A dendrobium cultivation management method based on multi-agent collaborative optimization

Abstract

The application discloses a kind of based on multi-agent collaborative optimization's dendrobium cultivation management method, including the following steps: division greenhouse subarea and establish the coupling atlas of adjacent and fan action range of fusion space, configuration acquisition execution and light supplement inspection channel;Collect environment, equipment and leaf image data, generate subarea time series observation set and subarea visual evidence set;Fusion generates subarea risk potential field set;Collaborative optimization output dehumidifier and fan control instruction;Execution review forms the potential field change record after execution;Update coupling edge weight and correct confidence weighted mapping, form closed-loop iterative management.The application adopts multi-agent collaborative optimization and visual review closed-loop correction method, realizes dendrobium greenhouse subarea humidity control management, with the advantages of strong cooperativity, high adaptability and good humidity control stability.

Description

Technical Field

[0001] This invention relates to the field of Dendrobium cultivation technology, and in particular to a Dendrobium cultivation management method based on multi-agent collaborative optimization. Background Technology

[0002] Current Dendrobium greenhouse cultivation management techniques typically rely on monitoring environmental parameters such as temperature and relative humidity, combined with the use of dehumidifiers and fans for start-stop or threshold control to maintain a suitable growth environment within the greenhouse. Some solutions further introduce a zoned control approach, allowing independent adjustment of different areas to reduce energy consumption and improve the precision of environmental regulation. Meanwhile, agricultural image acquisition and analysis technologies have been gradually applied to crop growth status identification, disease monitoring, and leaf surface feature extraction, but most are still used as independent monitoring methods, failing to form a tightly coupled control loop with the greenhouse environmental regulation process.

[0003] However, existing technologies still have significant shortcomings in the Dendrobium cultivation scenario: On the one hand, there is a coupling and propagation effect between greenhouse zones due to spatial proximity and fan ventilation, making it difficult for traditional independent zone control to accurately characterize the mutual influence between zones, which can easily lead to effective local regulation but unstable overall humidity control; on the other hand, although the apparent evidence of leaf moisture can reflect the local microenvironment, it is greatly affected by factors such as supplemental lighting conditions, clarity, and shading. Existing solutions lack a quantitative mapping of the credibility of visual evidence and a post-execution review and correction mechanism, making it difficult to achieve continuous optimization and closed-loop iteration based on execution feedback.

[0004] Therefore, how to provide a Dendrobium cultivation and management method based on multi-agent collaborative optimization is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a Dendrobium cultivation and management method based on multi-agent collaborative optimization. This invention employs multi-agent collaborative optimization and visual verification closed-loop correction methods to achieve zoned humidity control management of Dendrobium greenhouses, which has the advantages of strong collaboration, high adaptability, and good humidity control stability.

[0006] A method for cultivating and managing Dendrobium officinale based on multi-agent cooperative optimization according to an embodiment of the present invention includes the following steps: The Dendrobium greenhouse was divided into multiple greenhouse zones. A greenhouse zone coupling map was established based on the spatial adjacency relationship and the operating range of the fans. Environmental acquisition channels, dehumidifier execution channels, fan execution channels, and standardized acquisition channels for supplementary lighting of mobile inspection devices were configured for each greenhouse zone. Based on the greenhouse zoning coupling map, temperature data, relative humidity data and equipment operation status data of each greenhouse zone are collected. The mobile inspection device collects leaf images under the standardized acquisition channel of supplemental lighting, calculates the visual humidity index and visual confidence, and forms a zoning time-series observation set and a zoning visual evidence set. Wetness risk statistics are generated based on the partitioned time series observation set, and then weighted and mapped with the partitioned visual evidence set according to visual confidence to form a partitioned risk potential field set. Based on the risk potential field set of the zones and the coupling map of the greenhouse zones, multi-agent collaborative optimization is performed to obtain a set of collaborative control commands, including dehumidifier control commands and fan control commands; The set of collaborative control commands is sent to the dehumidifier execution channel and the fan execution channel and execution feedback is collected. The mobile inspection device is driven to verify, collect and update the set of visual evidence in the partition, and form a record of the potential field change after execution. The coupling edge weights of the greenhouse zonal coupling map are updated based on the record of potential field changes after execution, and the confidence weighted mapping is corrected based on the updated zonal visual evidence set, forming a closed-loop iteration for Dendrobium cultivation management.

[0007] Optionally, the generation of the greenhouse zoning coupling map specifically includes: Obtain the planar boundaries, cultivation rack layout, dehumidifier installation location, fan installation location and air outlet direction of the Dendrobium greenhouse, establish a spatial reference coordinate system for the greenhouse and form a set of greenhouse structural elements; Based on the set of greenhouse structural elements, the Dendrobium greenhouse is divided into zones to generate a set of greenhouse zone objects; Based on the set of greenhouse zone objects, spatial adjacency relationships and spatial adjacency scores are calculated to form a set of spatial adjacency relationships; Based on the fan's installation location, air outlet direction, and rated air volume, the fan's effective range is determined and projected onto the greenhouse space reference coordinate system to construct a fan effective range mapping table. Based on the wind turbine operating range mapping table, generate a set of wind turbine coupling edges and generate wind turbine coupling scores to form a set of wind turbine operating range relationships; A greenhouse zoning coupling map is generated by integrating the set of spatial adjacency relationships and the set of fan action range relationships; Based on the set of greenhouse zone objects, configure environmental acquisition channels, dehumidifier execution channels, fan execution channels, and standardized acquisition channels for supplementary lighting of mobile inspection devices for each greenhouse zone, and generate a zone channel configuration table and a standardized supplementary lighting configuration table.

[0008] Optionally, the generation of the partitioned temporal observation set and the partitioned visual evidence set specifically includes: Based on the greenhouse zoning coupling map and zoning channel configuration table, a zoning acquisition plan is generated and expanded to form a set of zoning acquisition tasks; Collect time-series observation sets for each partition by task set, including partition environment sequences and partition device status sequences; The mobile inspection device is driven to collect leaf image sequences according to the task set of the zone, and the inspection positioning mark and the collection timestamp sequence are associated to form a zoned leaf image set. The leaf image set is preprocessed, and the effective area of ​​the leaf is extracted based on color segmentation and morphological closing operation. The effective area mask of the leaf is generated and bound to the corresponding leaf image sequence to form a leaf area sample set. Based on the sample set of leaf region, the specular highlight ratio feature, specular highlight connected domain density feature and texture attenuation feature are calculated. Interval normalization and weighted summation are performed, and the results are truncated to between zero and one to obtain the visual wetness index sequence. Based on the sample set of leaf area, the sharpness score, overexposure ratio score, occlusion ratio score and supplementary light consistency score are calculated. Interval normalization and weighted summation are performed, and the results are truncated to between zero and one to obtain the visual confidence sequence. The index information of the visual humidity index sequence, visual confidence sequence, and leaf image set of each zone is consistently bound with the inspection and positioning marker according to the collection timestamp sequence to form a set of visual evidence for each zone.

[0009] Optionally, the generation of the partitioned risk potential field set specifically includes: Based on the partitioned time-series observation set, a partitioned water vapor pressure difference sequence and a partitioned humidity control execution sequence are generated and bound together to form a partitioned humidity control feature set; Based on the zonal humidity control feature set, the proportion of high humidity duration, low water vapor pressure difference duration and humidity control failure are calculated within a preset statistical window to form a zonal window statistical feature set. Based on the statistical feature set of the partition window, the proportion of high humidity duration, the proportion of low water vapor pressure difference duration and the proportion of humidity control failure are calculated. The results are weighted and summed according to the preset non-negative statistical weights and truncated to between zero and one to obtain the humidity risk statistics. Interval normalization is then performed to form a humidity risk statistics sequence. Based on the partitioned visual evidence set, the visual wetness index sequence and visual confidence sequence are extracted and matched with the wetness risk statistics sequence over a time window to generate a visual wetness evidence value sequence. A confidence-weighted mapping sequence is generated based on the visual confidence sequence. The mapping weight sequence is then subtracted from the mapping weight sequence to obtain a complementary statistical weight sequence, which together constitute the set of confidence-weighted mapping parameters. The sequence of humidity risk statistics and the sequence of visual humidity evidence values ​​are weighted and mapped according to the set of confidence-weighted mapping parameters to obtain the sequence of risk potential field values. These values ​​are then aggregated according to the partition identifier of the greenhouse partition object set to form a partition risk potential field set.

[0010] Optionally, the generation of the cooperative control instruction set specifically includes: A set of collaborative optimization tasks is generated based on the set of risk potential fields in different zones and the coupled map of greenhouse zones; Calculate the set of intervention effect coefficients for each greenhouse zone according to the set of collaborative optimization tasks, perform sign consistency verification and amplitude truncation, and form a set of usable intervention effect coefficients. Based on the set of available intervention effect coefficients and the set of coupling weights, a risk potential field prediction sequence is constructed. Adjacent greenhouse zones are uniformly coupled according to the set of coupling weights to form a zone prediction potential field set corresponding to the candidate control scheme. Based on the partitioned predicted potential field set, a collaborative optimization objective structure is constructed, and a collaborative optimization objective function is formed by setting preset non-negative objective weights for each item. Construct a collaborative optimization constraint structure and a collaborative optimization feasible region based on the equipment capability information of the dehumidifier execution channel and the fan execution channel; Within the feasible region of collaborative optimization, iteratively solve the control strength of candidate dehumidifiers and candidate fans, and output the optimal control strength of candidate dehumidifiers and optimal control strength of candidate fans that satisfy the collaborative optimization constraint structure. The optimal candidate dehumidifier control strength and the optimal candidate fan control strength are converted into dehumidifier control commands and fan control commands, and then aggregated and output according to the partition identifier of the greenhouse partition object set to form a set of collaborative control commands.

[0011] Optionally, the collaborative optimization constraint structure includes upper and lower limits of partition control strength constraints, partition ramping constraints, global power constraints, and partition quota constraints.

[0012] Optionally, the set of partitioned intervention effect coefficients is obtained by comparing the changes in risk potential field value and equipment operating status between the current statistical window and the previous statistical window.

[0013] Optionally, the generation of the potential field change record after execution specifically includes: Based on the collaborative control instruction set, the greenhouse zone object set is divided into zones according to the zone identifier, generating a zone execution instruction package, which is then sent to the dehumidifier execution channel and fan execution channel of the corresponding greenhouse zone, forming a zone sending record; Based on the execution time of the partitioned distribution record, execution confirmation collection and consistency verification are triggered, and execution feedback packets and execution consistency flags are generated. Execution deviation records are calculated based on the execution consistency flag and partition distribution records, and then aligned and bound to the partition time series observation set according to the feedback timestamp to form a partition execution process record; Based on the partition execution process record, the mobile inspection device is driven to perform verification and data collection, generating a set of verification and data collection tasks; The mobile inspection device performs verification and collection according to the set of verification and collection tasks, forming a set of blade images for the verification zone. Preprocessing is performed on the leaf image set of the review zone to generate a mask of the effective leaf area, resulting in a sample set of the review leaf area. The visual wetness index sequence and visual confidence sequence are calculated and bound to the index information and inspection positioning mark of the leaf image set of the review zone according to the review collection time. The visual evidence set of the zone is updated to form a subset of visual evidence for review. Based on the subset of visual evidence for review and the set of time-series observations in the partition, the risk potential field value sequence for review is calculated. It is then matched with the risk potential field value sequence before execution according to a preset statistical window to calculate the potential field change record after execution. The records of potential field changes after execution are correlated with the set of coupling edges in the greenhouse partition coupling graph to generate the edge weight update input set.

[0014] Optionally, the verification acquisition is achieved by the mobile inspection device sequentially reaching the inspection positioning mark and calling the supplementary lighting standardization configuration table to complete the supplementary lighting standardization acquisition channel configuration. The leaf image sequence is acquired under the supplementary lighting standardization acquisition channel and associated with the partition mark and the verification acquisition time.

[0015] Optionally, the generation of the closed-loop iteration of Dendrobium cultivation management specifically includes: Based on the weighted update input set, a border-level time series sample sequence is constructed by the set of coupled edges of the greenhouse partition coupling graph. The temporal correlation and delay correlation of the coupled edges are calculated based on the edge-level temporal sample sequences to form an edge-level correlation feature set; Based on the set of related features at the edge level and the temporal sample sequence at the edge level, the coupled edge weights are updated to form the updated set of coupled edge weights. Based on the updated set of visual evidence from different zones, the visual humidity index sequence, visual confidence sequence, and a subset of verified visual evidence are extracted according to the zone identifier of the greenhouse zone object set. The visual consistency residual and confidence segment statistics are calculated to form a set of mapped weight correction samples. The confidence-weighted mapping parameter set is corrected based on the mapping weight correction sample set to form the corrected confidence-weighted mapping parameter set; A closed-loop iterative configuration is generated based on the updated set of coupled edge weights and the corrected set of confidence weighted mapping parameters, and evidence-driven inspection triggering judgment is executed. The updated set of coupled edge weights is written back to the greenhouse partition coupling map, the corrected set of confidence weighted mapping parameters is written back to the configuration field of the set of confidence weighted mapping parameters, and the closed-loop iteration configuration, the priority inspection task marker for review and the statistical window identifier for this round are associated and stored to form a closed-loop iteration for Dendrobium cultivation management.

[0016] The beneficial effects of this invention are: This invention constructs a greenhouse zoning coupling map, incorporating spatial adjacency relationships and the operating range of fans into the zoning influence model. Based on this, the zoning risk potential field is used as the basis for coordinated control. This allows the control of dehumidifiers and fans to no longer be limited to independent adjustment of a single zone, but can take into account the propagation effect of adjacent zones and the overall consistency of humidity control. This improves the overall stability and zoning coordination of humidity control in Dendrobium greenhouses, reduces energy waste caused by local over-adjustment or repeated adjustment, and enhances the matching degree between coordinated control commands and actual changes in the cultivation environment.

[0017] This invention further incorporates leaf image evidence obtained by a mobile inspection device under a standardized acquisition channel with supplemental lighting into the humidity control decision-making process. A set of visual evidence for each zone is formed using a visual humidity index and visual confidence level, and then weighted and mapped with humidity risk statistics. Simultaneously, after execution, a subset of visual evidence and a record of post-execution potential field changes are generated through review acquisition. This allows for closed-loop correction of the coupling edge weights and confidence-weighted mapping parameter set of the greenhouse zone coupling map, enabling the degree of visual evidence participation to dynamically adjust with execution results and evidence consistency. This improves the accuracy of risk potential field characterization, enhances adaptability to complex working conditions and visual quality fluctuations, and achieves continuous optimization and traceable closed-loop iteration in the Dendrobium cultivation management process. Attached Figure Description

[0018] 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 The flowchart shows a Dendrobium cultivation and management method based on multi-agent collaborative optimization proposed in this invention. Figure 2 This invention presents a method for generating greenhouse zoning coupled maps and a flowchart of zoning channel configuration for Dendrobium cultivation management based on multi-agent collaborative optimization. Figure 3 This is a closed-loop iterative flowchart of the coupled edge weight update and confidence-weighted mapping correction of a Dendrobium cultivation and management method based on multi-agent collaborative optimization proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figures 1-3 A method for cultivating and managing Dendrobium officinale based on multi-agent collaborative optimization includes the following steps: The Dendrobium greenhouse was divided into multiple greenhouse zones. A greenhouse zone coupling map was established based on the spatial adjacency relationship and the operating range of the fans. Environmental acquisition channels, dehumidifier execution channels, fan execution channels, and standardized acquisition channels for supplementary lighting of mobile inspection devices were configured for each greenhouse zone. Based on the greenhouse zoning coupling map, temperature data, relative humidity data and equipment operation status data of each greenhouse zone are collected. The mobile inspection device collects leaf images under the standardized acquisition channel of supplemental lighting, calculates the visual humidity index and visual confidence, and forms a zoning time-series observation set and a zoning visual evidence set. Wetness risk statistics are generated based on the partitioned time series observation set, and then weighted and mapped with the partitioned visual evidence set according to visual confidence to form a partitioned risk potential field set. Based on the risk potential field set of the zones and the coupling map of the greenhouse zones, multi-agent collaborative optimization is performed to obtain a set of collaborative control commands, including dehumidifier control commands and fan control commands; The set of collaborative control commands is sent to the dehumidifier execution channel and the fan execution channel and execution feedback is collected. The mobile inspection device is driven to verify, collect and update the set of visual evidence in the partition, and form a record of the potential field change after execution. The coupling edge weights of the greenhouse zonal coupling map are updated based on the record of potential field changes after execution, and the confidence weighted mapping is corrected based on the updated zonal visual evidence set, forming a closed-loop iteration for Dendrobium cultivation management.

[0021] In this embodiment, the generation of the greenhouse zoning coupling map specifically includes: Obtain the planar boundaries, cultivation rack layout, dehumidifier installation location, fan installation location and air outlet direction of the Dendrobium greenhouse, establish a spatial reference coordinate system for the greenhouse and form a set of greenhouse structural elements; Based on the set of greenhouse structural elements, the Dendrobium greenhouse is divided into zones to generate a set of greenhouse zone objects; The partitioning process divides the greenhouse into multiple partitions based on the boundaries of the cultivation rack passage and the equipment service boundaries, and generates a partition identifier, a partition boundary polygon, and a partition center point for each greenhouse partition. Based on the set of greenhouse zone objects, spatial adjacency relationships and spatial adjacency scores are calculated to form a set of spatial adjacency relationships; The spatial adjacency relationship set is generated by parallel rules of partition boundary contact determination and partition center point distance determination. The spatial adjacency mark of any two greenhouse partitions is set to be valid when the partition boundary is in contact and the distance between the partition center points does not exceed the distance threshold calculated by the partition scale and the channel width; otherwise, it is set to be invalid. The spatial adjacency score is obtained by weighted fusion of the distance between the partition center points normalized relative to a distance threshold and the contact length of the partition boundary normalized relative to the sum of the perimeters of the two partitions. Based on the fan's installation location, air outlet direction, and rated air volume, the fan's effective range is determined and projected onto the greenhouse space reference coordinate system to construct a fan effective range mapping table. The projection projectors the area of ​​operation of the fan onto the greenhouse space reference coordinate system and intersects with the partition boundary polygon of the greenhouse partition object set to obtain the partition coverage set and coverage area ratio corresponding to each fan, thus forming a fan area mapping table. Based on the wind turbine operating range mapping table, generate a set of wind turbine coupling edges and generate wind turbine coupling scores to form a set of wind turbine operating range relationships; When the coverage area ratio of the same fan to two greenhouse zones is not less than the coverage threshold, a fan coupling edge is established between the two greenhouse zones. The fan coupling score is obtained by calculating the smaller value of the coverage area ratio of the two greenhouse zones for all fans and taking the larger value. A greenhouse zoning coupling map is generated by integrating the set of spatial adjacency relationships and the set of fan action range relationships; The greenhouse zoning coupling graph includes a set of greenhouse zoning nodes, a set of coupling edges, and a set of coupling edge weights. For any pair of greenhouse zoning, if spatial adjacency is marked as valid or the fan coupling score is not less than the coupling threshold, a coupling edge is generated between them and the coupling edge weight is calculated. The coupling edge weight is obtained by weighting the spatial adjacency score and the fan coupling score with a non-negative fusion coefficient and then normalizing it. Based on the greenhouse zone object set, configure environmental acquisition channels, dehumidifier execution channels, fan execution channels and mobile inspection device supplementary lighting standardized acquisition channels for each greenhouse zone, and generate a zone channel configuration table and a supplementary lighting standardized configuration table. The standardized acquisition channel for supplementary lighting includes supplementary lighting brightness levels, supplementary lighting color temperature levels, camera exposure lock strategy, and inspection positioning markers.

[0022] In this embodiment, the generation of the partitioned temporal observation set and the partitioned visual evidence set specifically includes: Based on the greenhouse zoning coupling map and zoning channel configuration table, a zoning acquisition plan is generated and expanded to form a set of zoning acquisition tasks; The zonal data collection plan includes the data collection timestamp sequence for each greenhouse zone, the frequency of temperature data collection, the frequency of relative humidity data collection, the frequency of equipment operating status data collection, and the zonal inspection sequence of the mobile inspection device. The expansion is based on the zonal data collection plan, which expands the data collection and inspection actions of each greenhouse zone in the data collection timestamp sequence into task items with zone identifiers and execution times. Collect time-series observation sets for each partition by task set, including partition environment sequences and partition device status sequences; The zoned environment sequence includes reading temperature and relative humidity data of each greenhouse zone from the environmental acquisition channel, and performing timestamp alignment, range verification, abrupt change removal and missing segment interpolation. The zoned equipment status sequence includes reading equipment operation status data from the dehumidifier execution channel and the fan execution channel, and performing consistency verification between command status and feedback status. The mobile inspection device is driven to collect leaf image sequences according to the task set of the zone, and the inspection positioning mark and the collection timestamp sequence are associated to form a zoned leaf image set. The mobile inspection device sequentially arrives at the inspection positioning markers of each greenhouse zone, calls the supplementary light brightness level, supplementary light color temperature level and camera exposure lock strategy in the supplementary light standard configuration table to complete the supplementary light standard acquisition channel configuration, and acquires leaf image sequences of the greenhouse zone under the supplementary light standard acquisition channel. The leaf image set is preprocessed, and the effective area of ​​the leaf is extracted based on color segmentation and morphological closing operation. The effective area mask of the leaf is generated and bound to the corresponding leaf image sequence to form a leaf area sample set. The preprocessing includes distortion correction, brightness normalization, and noise suppression. Based on the sample set of leaf region, the specular highlight ratio feature, specular highlight connected domain density feature and texture attenuation feature are calculated. Interval normalization and weighted summation are performed, and the results are truncated to between zero and one to obtain the visual wetness index sequence. The specular highlight ratio feature is obtained by first limiting each frame of leaf image to the set of effective pixels covered by the effective area mask of the leaf in the sample set of the leaf region, and extracting the set of brightness values. Pixels with brightness values ​​greater than or equal to the preset high quantile brightness threshold are marked as highlight pixels. The ratio of the number of highlight pixels to the total number of effective pixels of the leaf is calculated. The specular highlight connected domain density feature is obtained by performing morphological opening operation on the binary labeled map of the highlight pixels to denoise, then marking the connected domains according to the eight-neighbor connected domain rule and performing area filtering to obtain the effective highlight connected domain set. The effective highlight connected domain set is obtained by the ratio of the number of connected domains in the effective highlight connected domain set to the effective area of ​​the leaf corresponding to the effective area mask of the leaf. The texture attenuation feature is obtained by converting the leaf image corresponding to the effective pixel set of the leaf into a grayscale image and then using a fixed kernel difference operator to calculate the horizontal gradient map and the vertical gradient map to generate a gradient amplitude map. The average gradient amplitude of the gradient amplitude map is statistically analyzed within the effective area of ​​the leaf mask to obtain the current texture intensity value. The drying reference texture intensity value is obtained by calculating the historical leaf image with a visual humidity index lower than the preset drying threshold under the same supplementary light standardized acquisition channel in the same greenhouse zone. The difference between the drying reference texture intensity value and the current texture intensity value is divided by the drying reference texture intensity value. Based on the sample set of leaf area, the sharpness score, overexposure ratio score, occlusion ratio score and supplementary light consistency score are calculated. Interval normalization and weighted summation are performed, and the results are truncated to between zero and one to obtain the visual confidence sequence. The sharpness score is obtained by first limiting each frame of leaf image in the leaf region sample set to the effective area of ​​the leaf mask and converting it into a grayscale image, performing Laplacian operator convolution and calculating the variance of the convolution response as the original sharpness value, and then using preset lower and upper sharpness limits to normalize the original sharpness value. The overexposure ratio score is obtained by counting the number of pixels whose brightness value is greater than or equal to the preset saturation brightness threshold within the mask coverage area of ​​the effective area of ​​the blade and dividing it by the total number of effective pixels of the blade to obtain the original overexposure ratio value, and then performing reverse normalization according to the preset overexposure ratio upper limit. The occlusion ratio score is obtained by comparing the number of effective pixels in the effective area mask of the leaf with the number of reference pixels in the preset leaf area during the generation of the effective area mask. The percentage of pixels with holes filled by morphological closing operation in the effective area mask of the leaf represents the degree of occlusion. The effective coverage and the degree of occlusion are fused according to the preset weight and then normalized to obtain the score. The supplementary lighting consistency score is based on the supplementary lighting standardization configuration table, which reads the supplementary lighting brightness level, supplementary lighting color temperature level, and camera exposure lock strategy. Within the effective area of ​​the leaf mask coverage, the mean values ​​and brightness values ​​of the red, green, and blue channels are calculated and compared with the pre-established channel mean value benchmarks under the same supplementary lighting brightness level and supplementary lighting color temperature level. The sum of the absolute values ​​of the channel mean differences is then reverse-normalized by a preset consistency upper limit. The index information of the visual humidity index sequence, visual confidence sequence, and leaf image set of each zone is consistently bound with the inspection and positioning marker according to the collection timestamp sequence to form a set of visual evidence for each zone.

[0023] In this embodiment, the generation of the partitioned risk potential field set specifically includes: Based on the partitioned time-series observation set, a partitioned water vapor pressure difference sequence and a partitioned humidity control execution sequence are generated and bound together to form a partitioned humidity control feature set; The partitioned water vapor pressure difference sequence extracts temperature data, relative humidity data, and equipment operating status data according to the partition identifier of the greenhouse partition object set. Within the preset statistical window, the saturated water vapor pressure and water vapor pressure difference are calculated to generate the partitioned water vapor pressure difference sequence. The partitioned humidity control execution sequence is generated by extracting the dehumidifier control status and fan control status from the equipment operating status data and aligning them according to the collection timestamp sequence. The saturated vapor pressure is calculated from the temperature data according to the exponential relationship between saturated vapor pressure and actual vapor pressure. The actual vapor pressure is obtained by scaling the saturated vapor pressure and relative humidity data by a percentage. The vapor pressure difference is obtained by subtracting the actual vapor pressure from the saturated vapor pressure. Based on the zonal humidity control feature set, the proportion of high humidity duration, low water vapor pressure difference duration and humidity control failure are calculated within a preset statistical window to form a zonal window statistical feature set. The high humidity duration ratio is the percentage of sampling points in the statistical window where the relative humidity data is not less than the preset high humidity threshold. The low water vapor pressure difference duration ratio is the percentage of sampling points in the statistical window where the water vapor pressure difference is not greater than the preset low water vapor pressure difference threshold. The humidity control failure ratio is the percentage of sampling points in the statistical window where the relative humidity data does not decrease according to the preset decrease range, and the dehumidifier control status is on and the fan control status is on. Based on the statistical feature set of the partition window, the proportion of high humidity duration, the proportion of low water vapor pressure difference duration and the proportion of humidity control failure are calculated. The results are weighted and summed according to the preset non-negative statistical weights and truncated to between zero and one to obtain the humidity risk statistics. Interval normalization is then performed to form a humidity risk statistics sequence. The high humidity duration ratio is obtained by dividing the number of sampling points with relative humidity data not less than the preset high humidity threshold within the preset statistical window by the total number of sampling points within the preset statistical window. The low water vapor pressure difference duration ratio is obtained by counting the number of sampling points within a preset statistical window where the water vapor pressure difference is not greater than the preset low water vapor pressure difference threshold, and then dividing the number of sampling points by the total number of sampling points within the preset statistical window. The humidity control failure ratio is obtained by filtering sampling points in a preset statistical window where the dehumidifier control status is on and the fan control status is on, forming a humidity control on sampling point set. For each sampling point, the current relative humidity data and the minimum relative humidity data within a preset decrease judgment period after the sampling point are extracted as the subsequent minimum relative humidity data. When the current relative humidity data minus the subsequent minimum relative humidity data is less than the preset decrease range, the sampling point is determined as a humidity control failure sampling point. The humidity control failure ratio is obtained by dividing the number of humidity control failure sampling points by the total number of sampling points in the humidity control on sampling point set. Based on the partitioned visual evidence set, the visual wetness index sequence and visual confidence sequence are extracted and matched with the wetness risk statistics sequence over a time window to generate a visual wetness evidence value sequence. The visual wetness evidence value sequence is obtained by time aggregation of the visual wetness index sequence within a preset statistical window and point-by-point multiplication with the visual confidence sequence within the same statistical window, taking the window mean, and truncating the result to between zero and one. A confidence-weighted mapping sequence is generated based on the visual confidence sequence. The mapping weight sequence is then subtracted from the mapping weight sequence to obtain a complementary statistical weight sequence, which together constitute the set of confidence-weighted mapping parameters. The mapping weight sequence is obtained by segmenting the visual confidence sequence into a linear sequence according to a preset lower confidence limit and a preset upper confidence limit, and the mapping result is truncated to between zero and the preset maximum visual weight. The sequence of wet risk statistics and the sequence of visual wet evidence values ​​are weighted and mapped according to the set of confidence weighted mapping parameters to obtain the sequence of risk potential field values. These values ​​are then grouped according to the partition identifier of the greenhouse partition object set to form a partition risk potential field set. The risk potential field value sequence is obtained by weighting and summing the wet risk statistic and the visual wet evidence value according to the statistical weight and the visual weight, and then truncating the result to between zero and one.

[0024] In this embodiment, the generation of the cooperative control instruction set specifically includes: A set of collaborative optimization tasks is generated based on the set of risk potential fields in different zones and the coupled map of greenhouse zones; The collaborative optimization task set includes the risk potential field value of each greenhouse zone in the current statistical window, the risk potential field value in the previous statistical window, the dehumidifier control status and fan control status in the previous statistical window, and the set of coupling edge weights in the greenhouse zone coupling graph, and sets the control cycle length and the upper limit number of iterations. Calculate the set of intervention effect coefficients for each greenhouse zone according to the set of collaborative optimization tasks, perform sign consistency verification and amplitude truncation, and form a set of usable intervention effect coefficients. Based on the set of available intervention effect coefficients and the set of coupling weights, a risk potential field prediction sequence is constructed. Adjacent greenhouse zones are uniformly coupled according to the set of coupling weights to form a zone prediction potential field set corresponding to the candidate control scheme. The risk potential field prediction sequence maps the candidate dehumidifier control intensity and candidate fan control intensity to the predicted potential field value for each greenhouse zone. The predicted potential field value for each greenhouse zone is obtained by subtracting the product of the dehumidifier intervention effect coefficient and the candidate dehumidifier control intensity from the current risk potential field value of the greenhouse zone, subtracting the product of the fan intervention effect coefficient and the candidate fan control intensity, and adding the difference between the risk potential field value of the adjacent greenhouse zone and the risk potential field value of the current greenhouse zone by weighted summation according to the coupled edge weights. Based on the partitioned predicted potential field set, a collaborative optimization objective structure is constructed, and a collaborative optimization objective function is formed by setting preset non-negative objective weights for each item. The collaborative optimization objective structure includes a potential field reduction term, a potential field consistency term, an action smoothing term, and a power cost term. The potential field reduction term is obtained by summing the predicted potential field values ​​of all greenhouse zones. The potential field consistency term is obtained by summing the squares of the differences between the predicted potential field values ​​of the greenhouse zones at both ends of any coupling edge in the set of coupling edges of the greenhouse zone coupling map, weighted by the coupling edge weights. The action smoothing term is obtained by summing the squares of the differences between the control strength of the candidate dehumidifiers and the control strength of the dehumidifiers in the previous statistical window and the squares of the differences between the control strength of the candidate fans and the control strength of the fans in the previous statistical window for each greenhouse zone. The power cost term is obtained by summing the control strength of the candidate dehumidifiers and the control strength of the candidate fans in each greenhouse zone, weighted by the power cost coefficient per unit control strength. Based on the equipment capability information of the dehumidifier execution channel and the fan execution channel, a collaborative optimization constraint structure and a collaborative optimization feasible region are constructed. Within the feasible region of collaborative optimization, iteratively solve the control strength of candidate dehumidifiers and candidate fans, and output the optimal control strength of candidate dehumidifiers and optimal control strength of candidate fans that satisfy the collaborative optimization constraint structure. The iterative solution adopts an update rule that combines neighborhood consistency update based on coupled edge weight set with local descent update based on collaborative optimization objective function. After each iteration, the feasible region projection of the candidate dehumidifier control strength and the candidate fan control strength satisfies the collaborative optimization constraint structure until the upper limit of the number of iterations is reached or the descent of the collaborative optimization objective function does not exceed the preset convergence threshold. The optimal candidate dehumidifier control strength and the optimal candidate fan control strength are converted into dehumidifier control commands and fan control commands, and then aggregated and output according to the partition identifier of the greenhouse partition object set to form a set of collaborative control commands; The conversion includes mapping the optimal candidate dehumidifier control strength to the dehumidifier start / stop status and power level, and mapping the optimal candidate fan control strength to the fan start / stop status and speed level.

[0025] In this embodiment, the collaborative optimization constraint structure includes the upper and lower limits of partition control strength constraints, partition ramping constraints, global power constraints, and partition quota constraints. Among them, the upper and lower limits of the zone control intensity constraint limit the candidate dehumidifier control intensity of each greenhouse zone to be between the preset minimum control intensity of the dehumidifier and the preset maximum control intensity of the dehumidifier, and the candidate fan control intensity to be between the preset minimum control intensity of the fan and the preset maximum control intensity of the fan. The zone ramp constraint limits the absolute value of the difference between the control strength of the candidate dehumidifier in each greenhouse zone and the control strength of the dehumidifier in the previous statistical window to no more than the preset dehumidifier ramp limit, and the absolute value of the difference between the control strength of the candidate fan and the control strength of the fan in the previous statistical window to no more than the preset fan ramp limit. The global power constraint limits the total power of the candidate dehumidifier control strength and candidate fan control strength in all greenhouse zones to no more than the preset global power limit after conversion by the power conversion factor. The zonal quota constraint limits the cumulative control intensity of candidate dehumidifiers and the cumulative control intensity of candidate fans in any greenhouse zone within a preset quota period to not exceed the upper limit of the corresponding zonal quota.

[0026] In this embodiment, the set of partitioned intervention effect coefficients is obtained by comparing the changes in the risk potential field value of the current statistical window with the changes in the equipment operating status; The dehumidifier intervention effect coefficient is calculated by dividing the difference between the risk potential field value of the previous statistical window and the current statistical window by the sum of the change in the dehumidifier control state and the preset zero bias, and then truncated to between zero and the preset upper limit of the amplitude. The wind turbine intervention effect coefficient is calculated by dividing the difference between the risk potential field value of the previous statistical window and the current statistical window by the sum of the wind turbine control state change and the preset zero bias, and then truncated to between zero and the preset upper limit of the amplitude.

[0027] In this embodiment, the generation of the potential field change record specifically includes: Based on the collaborative control instruction set, the greenhouse zone object set is divided into zones according to the zone identifier, generating a zone execution instruction package, which is then sent to the dehumidifier execution channel and fan execution channel of the corresponding greenhouse zone, forming a zone sending record; The partition execution instruction package includes partition identifier, execution time, duration, dehumidifier start / stop status, dehumidifier power level, fan start / stop status, and fan speed level; Based on the execution time of the partitioned distribution record, execution confirmation collection and consistency verification are triggered, and execution feedback packets and execution consistency flags are generated. The execution confirmation data collection reads the command status and feedback status from the dehumidifier execution channel and the fan execution channel. The execution feedback package includes the partition identifier, feedback timestamp, dehumidifier command status, dehumidifier feedback status, fan command status, and fan feedback status. Execution deviation records are calculated based on the execution consistency flag and partition distribution records, and then aligned and bound to the partition time series observation set according to the feedback timestamp to form a partition execution process record; The execution deviation record includes the partition identifier, deviation start time, deviation duration, dehumidifier power level deviation, and fan speed level deviation. The dehumidifier power level deviation is determined by the difference between the power level corresponding to the dehumidifier feedback state and the dehumidifier power level in the partition execution command package. The fan speed level deviation is determined by the difference between the fan speed level corresponding to the fan feedback state and the fan speed level in the partition execution command package. Based on the partition execution process record, the mobile inspection device is driven to perform verification and data collection, generating a set of verification and data collection tasks; The set of review collection tasks includes a partition identifier, a review collection time, and an inspection location identifier. The review collection time is determined by the execution time and duration of the record issued by the partition and is located within a preset review period after the duration ends. The mobile inspection device performs verification and data collection according to the set of verification and data collection tasks, forming a set of blade images for the verification zone. Preprocessing is performed on the leaf image set of the review zone to generate a mask of the effective leaf area, resulting in a sample set of the review leaf area. The visual wetness index sequence and visual confidence sequence are calculated and bound to the index information and inspection positioning mark of the leaf image set of the review zone according to the review collection time. The visual evidence set of the zone is updated to form a subset of visual evidence for review. Based on the subset of visual evidence for review and the set of time-series observations in the partition, the risk potential field value sequence for review is calculated. It is then matched with the risk potential field value sequence before execution according to a preset statistical window to calculate the potential field change record after execution. The post-execution potential field change record includes the partition identifier, statistical window identifier, pre-execution risk potential field value, review risk potential field value, potential field change amount, partition execution process record and review collection task set; The records of potential field changes after execution are correlated with the set of coupling edges of the greenhouse partition coupling graph to generate the edge weight update input set. The edge weight update input set includes the potential field change of each greenhouse zone, the execution deviation record of the corresponding zone, and the set of identifiers of the adjacent greenhouse zones connected to it.

[0028] In this embodiment, the verification acquisition is carried out by the mobile inspection device sequentially reaching the inspection positioning mark and calling the supplementary lighting standardization configuration table to complete the supplementary lighting standardization acquisition channel configuration. The leaf image sequence is acquired under the supplementary lighting standardization acquisition channel and associated with the partition mark and the verification acquisition time.

[0029] In this embodiment, the generation of the closed-loop iteration of Dendrobium cultivation management specifically includes: Based on the weighted update input set, a border-level time series sample sequence is constructed from the set of coupled edges of the greenhouse partition coupling graph. The edge-level time series sample sequence is formed by extracting the potential field change of the greenhouse partitions at both ends under the same statistical window identifier, performing deviation recording, performing consistency marking, statistical window identifier, and adjacent greenhouse partition identifier set for each coupled edge, and sorting them according to the statistical window identifier; The temporal correlation and delay correlation of the coupled edges are calculated based on the edge-level temporal sample sequences to form an edge-level correlation feature set; The temporal correlation is obtained by fusing the potential field changes of the greenhouse partitions at both ends of the coupling edge within a preset correlation window, performing window-based consistency statistics and amplitude proximity statistics. The delay correlation is obtained by performing delay-by-delay pairing comparisons between the potential field changes of one greenhouse partition and the potential field changes of the other greenhouse partition in subsequent statistical windows within a preset delay window, and selecting the maximum correlation result. Based on the set of related features at the edge level and the temporal sample sequence at the edge level, the coupled edge weights are updated to form the updated set of coupled edge weights. The coupling edge weight update generates a weight correction value by weighting and fusing the temporal correlation, delay correlation and execution deviation suppression factor. This value is then smoothed and synthesized with the coupling edge weight before the update and truncated to a preset edge weight range. The execution deviation suppression factor is determined based on the window mean of the power level deviation and speed level deviation in the execution deviation records of the greenhouse zones at both ends of the coupling edge. Based on the updated set of visual evidence from different zones, the visual humidity index sequence, visual confidence sequence, and a subset of verified visual evidence are extracted according to the zone identifier of the greenhouse zone object set. The visual consistency residual and confidence segment statistics are calculated to form a set of mapped weight correction samples. The visual consistency residual is determined by the difference between the visual wet evidence value within the statistical window corresponding to the verified visual evidence subset and the wet risk statistic within the same statistical window. The confidence segmentation statistic is obtained by dividing the visual confidence sequence into multiple confidence intervals according to a preset confidence segmentation threshold and calculating the sample size, residual mean, and residual dispersion of each confidence interval. The confidence-weighted mapping parameter set is corrected based on the mapping weight correction sample set to form the corrected confidence-weighted mapping parameter set; The correction includes segmented adjustments to the preset lower confidence limit, preset upper confidence limit, and preset maximum visual weight, and residual constraint correction to the mapping weight sequence corresponding to each confidence interval. Specifically, when the visual consistency residual continuously exceeds the preset residual threshold in the corresponding confidence interval, the upper bound of the visual weight in that confidence interval is reduced; when the visual consistency residual continuously falls below the preset residual threshold in the corresponding confidence interval and the visual confidence of the verified visual evidence subset is stable, the upper bound of the visual weight in that confidence interval is increased. A closed-loop iterative configuration is generated based on the updated set of coupled edge weights and the corrected set of confidence weighted mapping parameters, and evidence-driven inspection triggering judgment is executed. The evidence-driven inspection trigger determination is based on the joint determination of the potential field uncertainty mark, visual consistency residual, visual confidence segmentation statistics and execution deviation record in the partition risk potential field set. When the potential field uncertainty mark reaches the preset trigger level, or the visual consistency residual exceeds the limit continuously, or the execution deviation record shows the preset deviation event continuously, the review priority inspection task mark of the next control cycle is generated and written into the partition inspection sequence of the partition collection plan. The updated set of coupled edge weights is written back to the greenhouse partition coupling map, the corrected set of confidence weighted mapping parameters is written back to the configuration field of the set of confidence weighted mapping parameters, and the closed-loop iteration configuration, the priority inspection task marker for review and the statistical window identifier for this round are associated and stored to form a closed-loop iteration for Dendrobium cultivation management.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-span greenhouse cultivation scenario for Dendrobium officinale in a mountainous area of ​​southern China. This greenhouse has long suffered from significant humidity crosstalk between different zones, uneven airflow from fans, and persistent dampness on some leaf surfaces that environmental sensor data fails to reflect in a timely manner. This is particularly problematic during nighttime humidity control and early morning rehumidification, where repeated starting and stopping of dehumidifiers and fans in localized areas often results in insufficient humidity control stability. The site was equipped with zoned cultivation racks, dehumidifiers, fans, and a mobile inspection device that moves along the aisles, meeting the implementation requirements for environmental data acquisition, execution feedback data acquisition, and standardized leaf image acquisition under supplemental lighting as described in this invention.

[0031] In practical applications, the greenhouse is first divided into zones based on the structural boundaries, cultivation rack aisles, and fan arrangement. A greenhouse zone coupling map integrating spatial adjacency and fan operating range is then established. Simultaneously, each zone is equipped with environmental acquisition channels, dehumidifier execution channels, fan execution channels, and standardized acquisition channels for supplemental lighting from mobile inspection devices. During operation, temperature, relative humidity, and equipment operating status data are continuously collected from each zone. The mobile inspection device acquires leaf images under fixed supplemental lighting brightness and color temperature levels, and with a camera exposure lock strategy, generating a zone-specific temporal observation set and a zone-specific visual evidence set. Subsequently, the humidity risk statistics and visual humidity evidence are weighted and mapped according to visual confidence levels to form a zone-specific risk potential field set. This set is then combined with the greenhouse zone coupling map to perform multi-agent collaborative optimization, outputting dehumidifier control commands and fan control commands. After the instruction is issued, execution feedback is collected synchronously, and the mobile inspection device is driven to collect and verify the data within the preset review period. The visual evidence set of the partition is updated to form a record of the potential field change after execution. Based on this, the coupled edge weights are updated and the confidence weighted mapping is corrected, and the next round of closed-loop iteration begins.

[0032] During continuous operation, the changing trends of the risk potential field in different zones, equipment execution feedback records, verification and collection records, and coupled edge weight update trajectories were recorded on-site over multiple time periods. The performance of the humidity control process before and after closed-loop iteration was compared. The recorded results show that the present invention can identify local humidity risk clusters earlier, reduce misjudgments in humidity control caused by air propagation between zones, and reduce control deviations caused by fluctuations in the quality of visual evidence. Under the same cultivation site and similar meteorological conditions, the humidity control process exhibits better zonal coordination, smoother operation, and continuous stability. Moreover, verification and inspection resources can be more concentrated on areas with higher uncertainty, thus verifying the practical application value and closed-loop optimization effect of the present invention in Dendrobium greenhouse cultivation management.

[0033] Table 1 Comparison of Operational Performance Before and After Optimization of Zoned Humidity Control Closed-Loop System in Dendrobium Greenhouse

[0034] As shown in Table 1, the proposed solution exhibits better overall performance in terms of risk suppression, zonal coordination, and control stability. The peak and mean values ​​of the zonal risk potential fields both decreased significantly, indicating that the solution not only suppresses sudden increases in local humidity risk but also reduces the overall humidity risk level during continuous operation. The mean differences in risk potential fields between zonal zones decreased synchronously, indicating that the humidity control states between greenhouse zones are more coordinated, and the cascading disturbances caused by fan air propagation and spatial proximity between zones are more effectively addressed. This result is consistent with the technical approach described in the claims, which constructs a greenhouse zone coupling map based on spatial adjacency and fan range, and introduces a potential field consistency term in the collaborative optimization. This demonstrates that after the coupled edge weights participate in the control solution, the phenomenon of local improvement and overall instability caused by isolated regulation of a single zone can be reduced.

[0035] From the perspective of execution process indicators, the number of humidity control action switching times was significantly reduced, and the cumulative runtime of both the dehumidifier and the fan decreased. Simultaneously, the magnitude of the risk potential field decline after execution verification increased. This indicates that the present invention does not achieve risk reduction by increasing equipment operating intensity, but rather achieves smoother and more effective humidity control action allocation through the synergistic optimization of the target and constraint structures. The decrease in action switching times corresponds to the action smoothing term and zone ramping constraint in the claims, which can reduce control jitter caused by frequent start-stop cycles. The decrease in cumulative runtime is related to the power cost term, global power constraint, and zone quota constraint, allowing the equipment output to be more concentrated on the zones that truly require intervention. At the same time, the number of recovery cycles after execution deviation triggering was shortened, indicating that the closed-loop iteration formed by execution feedback, verification data collection, and post-execution potential field change recording can more quickly correct the impact of control deviations on subsequent control cycles, improving operational recovery capabilities.

[0036] From the perspectives of visual evidence and closed-loop correction effects, the proportion of high-risk zones hit by the review inspection increased, and the mean of the mapping residual decreased after the participation of visual evidence, further demonstrating the effectiveness of the evidence-driven inspection triggering mechanism and the confidence-weighted mapping correction mechanism. Specifically, this invention does not directly use leaf image results for control, but first calculates the visual humidity index and visual confidence, then performs a confidence-weighted mapping with the humidity risk statistics, and corrects the mapping weights by reviewing a subset of visual evidence after execution. Therefore, it can suppress the interference of low-quality visual evidence on control decisions under conditions such as fluctuations in supplemental lighting, shading, and overexposure. The decrease in the mean of the mapping residual indicates higher consistency between visual and statistical evidence, thereby improving the representation accuracy of the risk potential field set of zones. This is also one of the important reasons for the decrease in the mean number of zones with persistent high humidity exceeding the threshold and the proportion of humidity control failure. Overall, the data in Table 1 supports the beneficial effects of this invention in the Dendrobium greenhouse zone humidity control scenario, namely, more accurate risk identification, more stable collaborative control, and more effective review of resource utilization.

[0037] The above are merely preferred embodiments 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. A method for cultivating and managing Dendrobium officinale based on multi-agent collaborative optimization, characterized in that, The steps include the following: The Dendrobium greenhouse was divided into multiple greenhouse zones. A greenhouse zone coupling map was established based on the spatial adjacency relationship and the operating range of the fans. Environmental acquisition channels, dehumidifier execution channels, fan execution channels, and standardized acquisition channels for supplementary lighting of mobile inspection devices were configured for each greenhouse zone. Based on the greenhouse zoning coupling map, temperature data, relative humidity data and equipment operation status data of each greenhouse zone are collected. The mobile inspection device collects leaf images under the standardized acquisition channel of supplemental lighting, calculates the visual humidity index and visual confidence, and forms a zoning time-series observation set and a zoning visual evidence set. Wetness risk statistics are generated based on the partitioned time series observation set, and then weighted and mapped with the partitioned visual evidence set according to visual confidence to form a partitioned risk potential field set. Based on the risk potential field set of the zones and the coupling map of the greenhouse zones, multi-agent collaborative optimization is performed to obtain a set of collaborative control commands, including dehumidifier control commands and fan control commands; The set of collaborative control commands is sent to the dehumidifier execution channel and the fan execution channel and execution feedback is collected. The mobile inspection device is driven to verify, collect and update the set of visual evidence in the partition, and form a record of the potential field change after execution. The coupling edge weights of the greenhouse zonal coupling map are updated based on the record of potential field changes after execution, and the confidence weighted mapping is corrected based on the updated zonal visual evidence set, forming a closed-loop iteration for Dendrobium cultivation management.

2. The Dendrobium cultivation and management method based on multi-agent collaborative optimization according to claim 1, characterized in that, The generation of the greenhouse zoning coupling map specifically includes: Obtain the planar boundaries, cultivation rack layout, dehumidifier installation location, fan installation location and air outlet direction of the Dendrobium greenhouse, establish a spatial reference coordinate system for the greenhouse and form a set of greenhouse structural elements; Based on the set of greenhouse structural elements, the Dendrobium greenhouse is divided into zones to generate a set of greenhouse zone objects; Based on the set of greenhouse zone objects, spatial adjacency relationships and spatial adjacency scores are calculated to form a set of spatial adjacency relationships; Based on the fan's installation location, air outlet direction, and rated air volume, the fan's effective range is determined and projected onto the greenhouse space reference coordinate system to construct a fan effective range mapping table. Based on the wind turbine operating range mapping table, generate a set of wind turbine coupling edges and generate wind turbine coupling scores to form a set of wind turbine operating range relationships; A greenhouse zoning coupling map is generated by integrating the set of spatial adjacency relationships and the set of fan action range relationships; Based on the set of greenhouse zone objects, configure environmental acquisition channels, dehumidifier execution channels, fan execution channels, and standardized acquisition channels for supplementary lighting of mobile inspection devices for each greenhouse zone, and generate a zone channel configuration table and a standardized supplementary lighting configuration table.

3. The Dendrobium cultivation and management method based on multi-agent collaborative optimization according to claim 1, characterized in that, The generation of the partitioned temporal observation set and the partitioned visual evidence set specifically includes: Based on the greenhouse zoning coupling map and zoning channel configuration table, a zoning acquisition plan is generated and expanded to form a set of zoning acquisition tasks; Collect time-series observation sets for each partition by task set, including partition environment sequences and partition device status sequences; The mobile inspection device is driven to collect leaf image sequences according to the task set of the zone, and the inspection positioning mark and the collection timestamp sequence are associated to form a zoned leaf image set. The leaf image set is preprocessed, and the effective area of ​​the leaf is extracted based on color segmentation and morphological closing operation. The effective area mask of the leaf is generated and bound to the corresponding leaf image sequence to form a leaf area sample set. Based on the sample set of leaf region, the specular highlight ratio feature, specular highlight connected domain density feature and texture attenuation feature are calculated. Interval normalization and weighted summation are performed, and the results are truncated to between zero and one to obtain the visual wetness index sequence. Based on the sample set of leaf area, the sharpness score, overexposure ratio score, occlusion ratio score and supplementary light consistency score are calculated. Interval normalization and weighted summation are performed, and the results are truncated to between zero and one to obtain the visual confidence sequence. The index information of the visual humidity index sequence, visual confidence sequence, and leaf image set of each zone is consistently bound with the inspection and positioning marker according to the collection timestamp sequence to form a set of visual evidence for each zone.

4. The Dendrobium cultivation and management method based on multi-agent collaborative optimization according to claim 1, characterized in that, The generation of the partitioned risk potential field set specifically includes: Based on the partitioned time-series observation set, a partitioned water vapor pressure difference sequence and a partitioned humidity control execution sequence are generated and bound together to form a partitioned humidity control feature set; Based on the zonal humidity control feature set, the proportion of high humidity duration, low water vapor pressure difference duration and humidity control failure are calculated within a preset statistical window to form a zonal window statistical feature set. Based on the statistical feature set of the partition window, the proportion of high humidity duration, the proportion of low water vapor pressure difference duration and the proportion of humidity control failure are calculated. The results are weighted and summed according to the preset non-negative statistical weights and truncated to between zero and one to obtain the humidity risk statistics. Interval normalization is then performed to form a humidity risk statistics sequence. Based on the partitioned visual evidence set, the visual wetness index sequence and visual confidence sequence are extracted and matched with the wetness risk statistics sequence over a time window to generate a visual wetness evidence value sequence. A confidence-weighted mapping sequence is generated based on the visual confidence sequence. The mapping weight sequence is then subtracted from the mapping weight sequence to obtain a complementary statistical weight sequence, which together constitute the set of confidence-weighted mapping parameters. The sequence of humidity risk statistics and the sequence of visual humidity evidence values ​​are weighted and mapped according to the set of confidence-weighted mapping parameters to obtain the sequence of risk potential field values. These values ​​are then aggregated according to the partition identifier of the greenhouse partition object set to form a partition risk potential field set.

5. The Dendrobium cultivation and management method based on multi-agent collaborative optimization according to claim 1, characterized in that, The generation of the coordinated control instruction set specifically includes: A set of collaborative optimization tasks is generated based on the set of risk potential fields in different zones and the coupled map of greenhouse zones; Calculate the set of intervention effect coefficients for each greenhouse zone according to the set of collaborative optimization tasks, perform sign consistency verification and amplitude truncation, and form a set of usable intervention effect coefficients. Based on the set of available intervention effect coefficients and the set of coupling weights, a risk potential field prediction sequence is constructed. Adjacent greenhouse zones are uniformly coupled according to the set of coupling weights to form a zone prediction potential field set corresponding to the candidate control scheme. Based on the partitioned predicted potential field set, a collaborative optimization objective structure is constructed, and a collaborative optimization objective function is formed by setting preset non-negative objective weights for each item. Construct a collaborative optimization constraint structure and a collaborative optimization feasible region based on the equipment capability information of the dehumidifier execution channel and the fan execution channel; Within the feasible region of collaborative optimization, iteratively solve the control strength of candidate dehumidifiers and candidate fans, and output the optimal control strength of candidate dehumidifiers and optimal control strength of candidate fans that satisfy the collaborative optimization constraint structure. The optimal candidate dehumidifier control strength and the optimal candidate fan control strength are converted into dehumidifier control commands and fan control commands, and then aggregated and output according to the partition identifier of the greenhouse partition object set to form a set of collaborative control commands.

6. The Dendrobium cultivation and management method based on multi-agent collaborative optimization according to claim 5, characterized in that, The collaborative optimization constraint structure includes upper and lower limits of partition control strength constraints, partition ramping constraints, global power constraints, and partition quota constraints.

7. The Dendrobium cultivation and management method based on multi-agent collaborative optimization according to claim 5, characterized in that, The set of partitioned intervention effect coefficients is obtained by comparing the changes in risk potential field value and equipment operating status between the current statistical window and the previous statistical window.

8. The Dendrobium cultivation and management method based on multi-agent collaborative optimization according to claim 1, characterized in that, The generation of the potential field change record after execution specifically includes: Based on the collaborative control instruction set, the greenhouse zone object set is divided into zones according to the zone identifier, generating a zone execution instruction package, which is then sent to the dehumidifier execution channel and fan execution channel of the corresponding greenhouse zone, forming a zone sending record; Based on the execution time of the partitioned distribution record, execution confirmation collection and consistency verification are triggered, and execution feedback packets and execution consistency flags are generated. Execution deviation records are calculated based on the execution consistency flag and partition distribution records, and then aligned and bound to the partition time series observation set according to the feedback timestamp to form a partition execution process record; Based on the partition execution process record, the mobile inspection device is driven to perform verification and data collection, generating a set of verification and data collection tasks; The mobile inspection device performs verification and collection according to the set of verification and collection tasks, forming a set of blade images for the verification zone. Preprocessing is performed on the leaf image set of the review zone to generate a mask of the effective leaf area, resulting in a sample set of the review leaf area. The visual wetness index sequence and visual confidence sequence are calculated and bound to the index information and inspection positioning mark of the leaf image set of the review zone according to the review collection time. The visual evidence set of the zone is updated to form a subset of visual evidence for review. Based on the subset of visual evidence for review and the set of time-series observations in the partition, the risk potential field value sequence for review is calculated. It is then matched with the risk potential field value sequence before execution according to a preset statistical window to calculate the potential field change record after execution. The records of potential field changes after execution are correlated with the set of coupling edges in the greenhouse partition coupling graph to generate the edge weight update input set.

9. The Dendrobium cultivation and management method based on multi-agent collaborative optimization according to claim 8, characterized in that, The verification acquisition is achieved by the mobile inspection device sequentially reaching the inspection positioning mark and calling the supplementary lighting standardization configuration table to complete the supplementary lighting standardization acquisition channel configuration. The leaf image sequence is acquired under the supplementary lighting standardization acquisition channel and associated with the partition mark and the verification acquisition time.

10. The Dendrobium cultivation and management method based on multi-agent collaborative optimization according to claim 1, characterized in that, The generation of the closed-loop iteration of Dendrobium cultivation management specifically includes: Based on the weighted update input set, a border-level time series sample sequence is constructed by the set of coupled edges of the greenhouse partition coupling graph. The temporal correlation and delay correlation of the coupled edges are calculated based on the edge-level temporal sample sequences to form an edge-level correlation feature set; Based on the set of related features at the edge level and the temporal sample sequence at the edge level, the coupled edge weights are updated to form the updated set of coupled edge weights. Based on the updated set of visual evidence from different zones, the visual humidity index sequence, visual confidence sequence, and a subset of verified visual evidence are extracted according to the zone identifier of the greenhouse zone object set. The visual consistency residual and confidence segment statistics are calculated to form a set of mapped weight correction samples. The confidence-weighted mapping parameter set is corrected based on the mapping weight correction sample set to form the corrected confidence-weighted mapping parameter set; A closed-loop iterative configuration is generated based on the updated set of coupled edge weights and the corrected set of confidence weighted mapping parameters, and evidence-driven inspection triggering judgment is executed. The updated set of coupled edge weights is written back to the greenhouse partition coupling map, the corrected set of confidence weighted mapping parameters is written back to the configuration field of the set of confidence weighted mapping parameters, and the closed-loop iteration configuration, the priority inspection task marker for review and the statistical window identifier for this round are associated and stored to form a closed-loop iteration for Dendrobium cultivation management.

Images