Multi-agent greenhouse global closed-loop control system and control method

Abstract

The application provides a multi-agent greenhouse global closed-loop control system and a control method. A global perception matrix module collects multi-dimensional data in the greenhouse in real time; an edge computing node module runs a multi-agent central system, which utilizes the hardware encoding and decoding capabilities and AI inference capabilities of a full-function GPU computing unit based on a MUSA architecture, and generates control instructions by real-time calculation of the collected data through each model; a cloud-side intelligent calculation cluster module constructs and runs a growth model training task and a global strategy optimization algorithm across greenhouses, and periodically updates the trained model to the edge computing node module; and a body execution layer module receives the control instructions and feeds back the task execution results to the edge computing node module to form a data closed-loop link. The application realizes the cooperation of fast thinking and slow thinking, replaces manpower with computing power, and realizes the whole-process closed-loop control from environmental perception, disease diagnosis, task planning to physical execution.

Description

Technical Field

[0001] This invention relates to the fields of smart agriculture, industrial automation control and artificial intelligence application technology. Specifically, it relates to a multi-agent greenhouse full-domain closed-loop control system and control method that utilizes full-function GPU computing power, edge-cloud collaborative architecture and multi-agent technology. Background Technology

[0002] With the rapid development of modern agricultural technology, controlled environment agriculture (CEA) has become an important means of ensuring the supply of high-quality agricultural products. High-end intelligent greenhouses achieve automated control of light, temperature, water, air, and fertilizer through sophisticated environmental control systems, greatly improving crop yield and quality.

[0003] However, despite increasingly sophisticated hardware, the level of intelligent agricultural production still faces significant challenges. While hardware such as greenhouse structures, irrigation machines, and environmental control computers have largely reached international advanced levels, there are still significant shortcomings in terms of human efficiency. The main reasons for this core technological pain point of insufficient human efficiency include the following aspects: Computing power deficit and underutilization of high-performance equipment: While expensive environmental control computers record massive amounts of environmental data (temperature, humidity, EC values, pH values, etc.), this data is often only used for simple rule-based control, such as turning on fans when the temperature exceeds 28°C. The system lacks the ability to mine the deeper value of the data and cannot handle complex nonlinear biological growth equations. Existing industrial PCs (IPCs) or PLC controllers have insufficient computing power to support deep learning-based computer vision (CV) models or complex fluid dynamics (CFD) simulations. This results in greenhouses with tens of millions of dollars in investment still relying on personal experience and Excel spreadsheets for decision-making.

[0004] Data silos and broken causal chains: Current agricultural Internet of Things (IoT) systems typically focus only on environmental data, neglecting the crop's own growth status data. There is a lack of causal relationship models between environmental data (input) and crop growth performance (output). For example, the system knows yesterday's temperature was 30 degrees Celsius, but doesn't know the specific impact of this high temperature on the rosebuds in area 3. This disconnect between data and the object makes the production process a black box, with quality control relying solely on final harvest grading, unable to intervene in the process during the over 1000-hour growth cycle.

[0005] The lack of automated execution: Current greenhouse automation mainly focuses on uniform actions across the entire greenhouse (such as uniform window opening and irrigation), lacking the ability to perform precise operations on individual plants or localized areas. When pests or diseases are found on a particular crop, manual inspection, marking, and spraying are still required, which is the bottleneck limiting the improvement of human efficiency.

[0006] To address the aforementioned problems, there is an urgent need in this field for a new generation of greenhouse control technology that can integrate comprehensive sensing data, possess high-performance edge computing capabilities, utilize artificial intelligence models to replace human experience, and command physical equipment to perform closed-loop execution. Currently, no descriptions or reports of technologies similar to this invention have been found, nor have similar domestic or international materials been collected. Summary of the Invention

[0007] To address the aforementioned shortcomings in the prior art, this invention provides a multi-agent greenhouse full-domain closed-loop control system and control method.

[0008] According to a first aspect of the present invention, a multi-agent greenhouse full-domain closed-loop control system is provided, comprising: a full-domain perception matrix module, an edge computing node module, a cloud-based intelligent computing cluster module, and an embodied execution layer module; wherein: The global perception matrix module is deployed inside the greenhouse to collect environmental field data, crop phenotypic image data and equipment status data in real time, and send them to the edge computing node module via the local area network. The edge computing node module, deployed locally in the greenhouse, is equipped with a full-featured GPU computing unit based on the MUSA architecture to run a multi-agent central system. This multi-agent central system comprises multiple independent and cooperative agents. Each agent utilizes the hardware encoding / decoding capabilities and AI inference capabilities of the full-featured GPU computing unit to perform real-time calculations on the collected data using computer vision models, 3D digital twin models, and crop growth mechanism models. The 3D digital twin model is directly constructed and iteratively updated by the edge computing node module based on local real-time multi-dimensional data to generate control commands. The cloud-based intelligent computing cluster module is connected to the edge computing node module via a wide area network. It is used to store historical production data of the greenhouse, build and run training tasks for computer vision models and crop growth mechanism models, as well as global strategy optimization algorithms across greenhouses. The trained and iterated models are periodically distributed and updated to the edge computing node module. The embodied execution layer module is used to receive control commands issued by the edge computing node module, perform specific task operations within the greenhouse, and feed back the execution results to the edge computing node module for control command correction, thus forming a data closed-loop link.

[0009] Preferably, the full-featured GPU computing unit integrates a graphics rendering engine, an AI inference engine, a video processing engine, and a physics simulation engine; wherein: The video processing engine is used to decode multiple RTSP high-definition video streams from the global perception matrix module in parallel, and to complete the decoding directly in the video memory to obtain video frames. The AI ​​inference engine is used to perform real-time target detection, instance segmentation, and disease feature extraction on the decoded video frames. The real-time target detection results are used to locate the pixel positions and categories of individual crops, lesions, pests, and operating equipment in the image, providing the spatial agent of the multi-agent central system with 3D coordinate back-projection. The instance segmentation results are used to obtain the precise pixel boundaries of each crop and each leaf, providing the health agent of the multi-agent central system with single-plant nutrient diagnosis and the operating agent with the ability to generate precise local spraying areas. The disease feature extraction results are used to output the disease category, severity level, and spread trend, providing the health agent of the multi-agent central system with the ability to generate health diagnosis events and trigger SOP matching for the operating agent. The graphics rendering engine is used to render a three-dimensional greenhouse scene based on 3D Gaussian sputtering technology in real time, providing a visual interaction for remote management personnel. The physical simulation engine is used to perform collision detection and dynamic simulation of the robotic arm path or movement trajectory of the embodied execution layer module in virtual space.

[0010] Preferably, the multi-agent central system adopts an event bus-based collaborative architecture, including the following four types of agents: The health intelligence agent is used to analyze crop phenotypic image data using computer vision models and combine the instance segmentation results obtained by the AI ​​inference capability of the full-function GPU computing unit to obtain the precise pixel boundaries of each crop and each leaf, perform single-plant nutrient diagnosis, and diagnose the crop's pest and disease status, nutritional status and growth stage in real time. At the same time, based on the disease feature extraction results obtained by the AI ​​inference capability of the full-function GPU computing unit, it outputs the disease category, severity level and spread trend, and generates health diagnosis events to be published to the event bus. The spatial intelligent agent is used to locate the pixel positions and categories of individual crops, lesions, pests, and operating equipment in the image by using the real-time target detection results obtained by the AI ​​inference capability of the full-function GPU computing unit, perform three-dimensional coordinate back projection, maintain a unified spatiotemporal coordinate system in the greenhouse, generate task path control instructions based on the three-dimensional digital twin model, and provide path planning, navigation obstacle avoidance, and object spatial positioning services for the embodied execution layer module. The operation intelligence agent is used to subscribe to the health diagnosis event, trigger SOP matching according to the preset standard operating procedure and the disease feature extraction result, and obtain the local precise spraying area according to the instance segmentation result. Based on the three-dimensional digital twin model, it generates task work order control instructions containing three-dimensional coordinates and action definitions, and distributes them to the embodied execution layer module through the event bus. The predictive agent combines current environmental data, health diagnostic events released by the health agent, and crop growth mechanism models to predict crop yield, quality grade, and market cycle within a preset time period. During the prediction process, two parallel simulation branches are constructed, one including abnormal states and one not including abnormal states, to quantitatively evaluate the impact of the current abnormal state on the final yield, serving as the basis for the economic threshold decision of the operational agent.

[0011] Preferably, the task work order control instructions generated by the task intelligence agent include: a unique task identifier ID, the three-dimensional spatial coordinates of the target object, the task type code, the execution priority, and the completion verification standard; The autonomous mobile robot in the embodied execution layer module, after receiving the task work order control command, uses its onboard sensors to scan passive visual markers in the greenhouse for secondary precise positioning, ensuring that the error between the physical operation position and the three-dimensional spatial coordinates in the task work order control command is less than a preset threshold.

[0012] Preferably, the above system further includes: a full-process closed-loop verification mechanism trigger module; wherein: After the physical operation is completed by the embodied execution layer module, the full-process closed-loop verification mechanism trigger module automatically triggers the re-inspection process instruction, controls the full-domain perception matrix module to automatically collect post-operation image data of the operation area, and sends it to the edge computing node module. The health agent or operation agent of the edge computing node module compares and analyzes the post-operation image data to verify whether the operation effect meets the completion verification standard; if the verification fails, a rework task work order control instruction is automatically generated and the execution priority is set and fed back to the embodied execution layer module. The edge computing node module simultaneously feeds back the entire operation process data to the cloud-based intelligent computing cluster module for training and strategy iteration of computer vision models and crop growth mechanism models; wherein, the strategy includes one or more of the following: disease identification and treatment strategy, environmental field regulation strategy, operation task scheduling strategy, energy consumption optimization strategy, and water and fertilizer formula strategy.

[0013] According to a second aspect of the present invention, a method for closed-loop control of the entire greenhouse area using the system described in any one of the above claims is provided, comprising: The global perception matrix module collects multi-dimensional data from inside the greenhouse, while the edge computing node module utilizes the real-time rendering capabilities of a full-featured GPU to fuse discrete video streams with sensor data and update the real-time 3D scene of the greenhouse. The edge computing node module identifies abnormal crop conditions based on computer vision models, maps the abnormal conditions to specific locations in the world coordinate system of the 3D scene based on a 3D digital twin model, and publishes and responds to abnormal events. The three-dimensional digital twin model combines environmental data and production process knowledge base to generate standardized work order control instructions and task path control instructions, which are then distributed through an event bus. The embodied execution layer module schedules the corresponding robot or equipment according to the task work order control instructions, plans the optimal path to reach the target location to perform the task operation according to the task path control instructions, and provides real-time feedback on the execution status.

[0014] Preferably, the above method further includes: The global perception matrix module automatically collects image data of the work area after the operation, verifies the effect through the edge computing node module, and then cleans the data of the entire operation process through the edge computing node module before feeding it back to the cloud intelligent computing cluster module for model training and strategy iteration of computer vision models and crop growth mechanism models. Preferably, the real-time 3D scene of the greenhouse is constructed based on 3D Gaussian sputtering technology, wherein each Gaussian point carries 3D coordinates, color information and semantic labels for remote management personnel to perform visual interaction.

[0015] Preferably, the data from the entire operation process is cleaned by the edge computing node module and then fed back to the cloud-based intelligent computing cluster module, including: The data backflow process follows the principle of privacy protection. The edge computing node module performs data cleaning processing such as desensitization, encryption or federated learning gradient extraction on sensitive data before transmitting it to the cloud intelligent computing cluster module. Non-sensitive data is uploaded in full for big data analysis.

[0016] Preferably, when generating task work order control instructions, the edge computing node module simultaneously invokes the simulation results of the crop growth mechanism model. These simulation results are generated by the predictive agent based on current environmental data, the crop growth mechanism model, and health diagnostic events issued by the healthy agent. By running two simulation branches in parallel—one containing abnormal states and one without—the difference between the two yield curves is used to quantitatively assess the impact of the current abnormal state on the final crop yield. If, in the simulation results, the impact of the current abnormal state on the final crop yield is lower than a preset threshold, or if the current environmental data is unsuitable for task operation, the operational agent of the edge computing node module will suspend or cancel the generation of the task work order control instructions, achieving intelligent decision-making based on economic thresholds.

[0017] By adopting the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art: Significantly improves human efficiency: The multi-agent greenhouse full-domain closed-loop control system and control method provided by this invention effectively improves the problem of insufficient human efficiency by replacing manual inspection and decision-making with multiple agents and replacing repetitive manual tasks with robots.

[0018] Process Transparency and Computability: The multi-agent greenhouse full-domain closed-loop control system and method provided by this invention transform agricultural production from experience and intuition into the mathematical formula y=f(x). This mathematical mapping relationship is specifically reflected in: First, a computer vision model maps crop phenotypic images x to pest and disease categories and severity levels y; second, a crop growth mechanism model maps environmental field data x to future yield, quality level, and market cycle y; third, a three-dimensional digital twin model maps multi-source sensor data x to the three-dimensional greenhouse scene state y; and fourth, a multi-agent decision-making system maps abnormal events x to structured task orders y. This makes the production process quantifiable, predictable, and controllable.

[0019] Maximizing Hardware Utilization: The multi-agent greenhouse full-domain closed-loop control system and method provided by this invention activates previously dormant expensive hardware facilities and solves the computing power deficit problem through software-defined intelligence. Specifically, this is achieved through the following technical components: First, the edge computing node module is configured with a full-function GPU computing unit based on the MUSA architecture, integrating the capabilities of a graphics rendering engine, AI inference engine, video processing engine, and physical simulation engine into one, enabling a single device to handle video decoding, AI inference, 3D rendering, and physical simulation tasks that originally required multiple heterogeneous chips. Second, the multi-agent central system uses an event bus collaborative architecture to concurrently run four types of agents: health, space, operation, and prediction, reusing computing resources on the same GPU computing power base to avoid computing power idleness. Third, the dual-brain mechanism of edge-cloud collaboration sinks high-frequency real-time tasks to the edge for execution and moves large-scale training tasks to the cloud cluster, realizing on-demand matching of computing power.

[0020] Data security and privacy protection: The multi-agent greenhouse full-domain closed-loop control system and control method provided by this invention, with the privacy protection principle of data not leaving the greenhouse under the edge-cloud collaborative architecture, effectively protects the core technological secrets of planting enterprises (such as specific water and fertilizer formulas, light strategies, etc.).

[0021] This invention provides a multi-agent greenhouse full-domain closed-loop control system and method. The edge computing node module incorporates a multi-agent hub based on a full-featured GPU, which constructs a real-time digital twin of the greenhouse by processing multiple RTSP video streams and environmental sensor data in real time. Through the collaboration of four types of agents—health, space, operation, and prediction—an event bus mechanism is used to transform fuzzy biological perception information into precise mathematical control commands. The cloud-based intelligent computing cluster module is responsible for large-scale model training and global strategy optimization, and distributes the iterated model to the edge, achieving dual-brain collaboration between fast and slow thinking. By replacing human labor with computing power, a full-process closed-loop control is achieved, encompassing environmental perception, disease diagnosis, task planning, and physical execution. Attached Figure Description

[0022] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a working architecture diagram of a multi-agent greenhouse full-domain closed-loop control system in a preferred embodiment of the present invention.

[0023] Figure 2 This is a flowchart illustrating the workflow of a multi-agent greenhouse global closed-loop control method in a preferred embodiment of the present invention. Detailed Implementation

[0024] The embodiments of the present invention are described in detail below: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

[0025] Existing high-end intelligent greenhouses typically suffer from technical problems such as computing power deficit and underutilization of high-end configurations, data silos and broken causal chains, as well as a lack of automated execution, resulting in technical pain points of insufficient human efficiency. As a result, there is a lack in this field of a new generation of greenhouse control technology that can integrate full-domain perception data, possess high-performance edge computing capabilities, use artificial intelligence models to replace human experience, and command physical equipment to perform closed-loop execution.

[0026] To address the aforementioned issues, one embodiment of the present invention provides a multi-agent greenhouse full-domain closed-loop control system. This system is based on full-function GPU computing power and an edge-cloud collaborative architecture. By introducing the concept of computational agriculture, it transforms the fuzzy perception of the physical world into a precise mathematical model, realizing a paradigm shift from relying on weather conditions to data-driven approaches. This solves the technical problems of existing greenhouse environmental control systems being unable to understand crop growth status, relying on human experience for decision-making, and having insufficient computing power to support real-time reasoning of complex models.

[0027] Specifically, such as Figure 1 As shown, the multi-agent greenhouse full-domain closed-loop control system provided in this embodiment may include: a full-domain perception matrix module, an edge computing node module, a cloud-based intelligent computing cluster module, and an embodied execution layer module; wherein: The global perception matrix module, deployed inside the greenhouse, includes a fixed environmental sensor group, a mobile multispectral acquisition device, and a robotic telemetry device. It is used to collect environmental field data, crop phenotypic image data, and equipment status data in the greenhouse in real time, and send them to the edge computing node module via the local area network. The edge computing node module, deployed locally in the greenhouse, is equipped with a full-featured GPU computing unit based on the MUSA architecture to run a multi-agent central system. The multi-agent central system includes multiple independent and cooperative agents. Each agent utilizes the hardware encoding and decoding capabilities and AI inference capabilities of the full-featured GPU computing unit to perform real-time calculations on the collected data through computer vision models, 3D digital twin models, and crop growth mechanism models. Among them, the 3D digital twin model is directly constructed and iteratively updated on the edge computing node module based on local real-time multi-dimensional data to generate control commands. The cloud-based intelligent computing cluster module is connected to the edge computing node module via a wide area network. It is used to store historical production data of the greenhouse, build and run training tasks for computer vision models and crop growth mechanism models, as well as global strategy optimization algorithms across greenhouses. It also regularly distributes and updates the trained and iterated models to the edge computing node module. The embodied execution layer module, including autonomous mobile robots, automated environmental control equipment, and operation assistance equipment, is used to receive control commands issued by the edge computing node module, perform specific tasks within the greenhouse, and feed back the execution results to the edge computing node module for control command correction, forming a closed-loop data link.

[0028] The multi-agent greenhouse closed-loop control system provided in this embodiment adopts a dual-brain collaborative control strategy of fast thinking (System 1) and slow thinking (System 2). Among them, the edge computing node module executes the fast thinking strategy, which is responsible for millisecond-level real-time inference, 3D rendering and equipment control. Even when disconnected from the cloud intelligent computing cluster network module, it can maintain the safe operation of basic production operations based on a locally cached lightweight model. The cloud intelligent computing cluster module executes the slow thinking strategy, which is responsible for calculating the global optimal solution across cycles and greenhouses, performing full-element simulation and deduction of complex biophysical models and training large-scale deep learning models.

[0029] In some preferred embodiments, the aforementioned full-featured GPU computing unit integrates a graphics rendering engine (Graphics), an AI inference engine (AI), a video processing engine (Video), and a physics simulation engine (Physics); wherein: The video processing engine is used to decode multiple RTSP high-definition video streams from the global perception matrix module in parallel, and does not need to move the video data to the host memory. It can directly complete the decoding in the video memory to obtain video frames. The AI ​​inference engine is used for real-time target detection, instance segmentation, and disease feature extraction on decoded video frames. Real-time target detection results are used to locate the pixel positions and categories of individual crops, lesions, pests, and operating equipment in the image, providing the spatial agent of the multi-agent central system for 3D coordinate back-projection. Instance segmentation results are used to obtain the precise pixel boundaries of each crop and each leaf, providing the health agent of the multi-agent central system for single-plant nutrient diagnosis and the operating agent for generating precise local spraying areas. Disease feature extraction results are used to output disease categories, severity levels, and spread trends, providing the health agent of the multi-agent central system for generating health diagnosis events and triggering SOP matching for the operating agent. The graphics rendering engine is used to render greenhouse 3D scenes based on 3D Gaussian sputtering (3DGS) technology in real time, providing visual interaction for remote managers; The physics simulation engine is used to perform collision detection and dynamic simulation of the robotic arm path or movement trajectory of the embodied execution layer module in virtual space.

[0030] In some preferred embodiments, the above-mentioned multi-agent central system adopts an event bus-based cooperative architecture, including the following four types of agents: The Crop Health Agent is used to analyze crop phenotypic image data using computer vision models and combine the instance segmentation results obtained by the AI ​​inference capabilities of the full-featured GPU computing unit to obtain the precise pixel boundaries of each crop and each leaf, perform single-plant nutrient diagnosis, and diagnose the crop's pest and disease status, nutritional status, and growth stage in real time. At the same time, based on the disease feature extraction results obtained by the AI ​​inference capabilities of the full-featured GPU computing unit, it outputs the disease category, severity level, and spread trend, and generates health diagnosis events to be published to the event bus. The Spatial Agent is used to locate the pixel positions and categories of individual crops, lesions, pests, and operating equipment in images by using the AI ​​inference capabilities of a full-featured GPU computing unit to obtain real-time target detection results. It performs three-dimensional coordinate back projection, maintains a unified spatiotemporal coordinate system within the greenhouse, and generates task path control instructions based on a three-dimensional digital twin model. It provides path planning, navigation obstacle avoidance, and object spatial positioning services for the embodied execution layer module. The Operation Agent is used to subscribe to health diagnosis events, trigger SOP matching based on the preset standard operating procedures and the results of disease feature extraction, and obtain local precise spraying areas based on the instance segmentation results. Based on the three-dimensional digital twin model, it generates task work order control instructions containing three-dimensional coordinates and action definitions, and distributes them to the embodied execution layer module through the event bus. The Prediction Agent combines current environmental data, health diagnostic events released by the Health Agent, and crop growth mechanism models to predict crop yield, quality grade, and market cycle within a preset time period. During the prediction process, two parallel simulation branches are constructed, one including abnormal states and one not including abnormal states, to quantitatively evaluate the impact of current abnormal states on the final yield, serving as the basis for the economic threshold decision of the operational agent.

[0031] In some preferred embodiments, the process by which the aforementioned health intelligent agent generates health diagnosis events may further include: (1) Multispectral and RGB dual-modal fusion diagnosis: The early stress characteristics reflected by the invisible spectrum (near infrared, red edge band) are fused with the visible spot disease characteristics in the feature layer to achieve early warning of stresses such as water shortage and nitrogen deficiency that are invisible to the naked eye; (2) Lightweight inference at the edge: Based on the MUSA architecture operator optimization, the large model trained in the cloud is deployed to the edge after quantization pruning to achieve inference in the 30-millisecond level; (3) Event-driven structured output: The diagnostic results are encapsulated in a structured manner according to {type, severity level, confidence level, pixel position, timestamp} and published to the event bus, decoupled and collaborated with the spatial agent and the operational agent to form a traceable closed-loop link.

[0032] In some preferred embodiments, the process by which the space agent generates task path control instructions may further include: (1) Centimeter-level semantic map based on 3DGS: Each Gaussian point carries three-dimensional coordinates, color information and semantic labels (such as pillars, pipes, crops, and passages), so that path planning not only considers geometric obstacle avoidance, but also protects crops based on semantics (e.g., detouring instead of crossing dense flower bud areas). (2) Dynamic twin update mechanism: The three-dimensional digital twin model is continuously updated based on real-time video streams and sensor data, which can reflect dynamic obstacles such as temporary piles and mobile devices in the greenhouse, which is different from traditional static maps; (3) Pixel-world coordinate reverse projection: Combining the real-time pose of the acquisition device with the camera intrinsic parameters, the pixel-level disease location output by the health intelligent agent is converted into world coordinates in real time for the robot to execute directly.

[0033] In some preferred embodiments, the above-mentioned intelligent agent for generating task order control instructions containing three-dimensional coordinates and action definitions may further include: (1) Three-level linkage mechanism of event subscription - SOP matching - twin positioning: After subscribing to a health diagnosis event, the corresponding treatment plan is first matched in the SOP knowledge base (such as local spraying of 50ml sulfur suspension for Level 2 powdery mildew), and then the precise three-dimensional coordinates and reachable path of the target crop are obtained through the three-dimensional digital twin model. (2) Economic threshold decision fusion: Before generating a work order, the inference results of the predictive agent are called. If the impact of the current anomaly on the final output is lower than the economic threshold, the work order is suspended to avoid ineffective input. (3) Structured work order protocol: The work order includes fields such as task ID, three-dimensional coordinates, job type code, priority, and completion verification standard, and supports unified scheduling of heterogeneous equipment (drones, AMRs, spraying robots).

[0034] In some preferred embodiments, the aforementioned predictive agent, which extrapolates crop yield, quality grade, and market cycle within a preset time period, may further include: (1) Coupling of mechanism model and data-driven model: The crop growth mechanism equation based on photosynthesis and transpiration is integrated with the deep learning residual model trained on historical data, taking into account both interpretability and prediction accuracy. (2) Dual-branch comparative simulation: Two simulation trajectories, one for the healthy group and one for the abnormal group, are run in parallel. The difference is the output impact of the abnormal state, which provides a direct basis for economic threshold decision-making. (3) Full-factor simulation: Simultaneously outputs output, quality grade and market launch cycle to support multi-dimensional business decision-making.

[0035] In some preferred embodiments, the task order control instructions generated by the above-mentioned task agent include: a unique task identifier ID, the three-dimensional spatial coordinates (x, y, z) of the target object, the task type code, the execution priority, and the completion verification criteria. The autonomous mobile robot in the embodied execution layer module, after receiving the task order control command, uses its onboard sensors to scan the passive visual markers in the greenhouse for secondary precise positioning, ensuring that the error between the physical working position and the three-dimensional spatial coordinates in the task order control command is less than a preset threshold (e.g., 10mm).

[0036] In some preferred embodiments, the system further includes: a full-process closed-loop verification mechanism trigger module; wherein: After the physical operation is completed by the embodied execution layer module, the full-process closed-loop verification mechanism trigger module automatically triggers the re-inspection process instruction, controls the full-domain perception matrix module to automatically collect post-operation image data of the operation area, and sends it to the edge computing node module; The health agent or operation agent of the edge computing node module compares and analyzes the image data after the operation to verify whether the operation effect meets the completion verification standard; if the verification fails, it automatically generates a rework task work order control instruction and sets the execution priority to feed back to the embodied execution layer module. The edge computing node module simultaneously feeds back the data of the entire operation process to the cloud-based intelligent computing cluster module for training and strategy iteration of computer vision models and crop growth mechanism models; among them, strategies include one or more of the following: disease identification and treatment strategies, environmental field control strategies, operation task scheduling strategies, energy consumption optimization strategies, and water and fertilizer formula strategies.

[0037] In some preferred embodiments, the global strategy optimization algorithm across greenhouses in the aforementioned cloud-based intelligent computing cluster module may further include: (1) Federated learning algorithm: Under the premise of ensuring that the data of each greenhouse does not leave the garden, the model gradients of multiple greenhouses are aggregated in the cloud and iteratively generated a general growth model applicable to different climate zones and different crop varieties. (2) Cross-greenhouse transfer learning algorithm: Transfer the model parameters such as disease identification and environmental control strategies accumulated in mature greenhouses to newly built greenhouses to shorten the cold start period; (3) Multi-objective optimization algorithm: Taking output, quality, energy consumption and labor cost as optimization objectives, the Pareto optimal combination of environmental control and operation parameters is solved by using evolutionary algorithms such as NSGA-II; (4) Swarm intelligence scheduling algorithm: Based on the order demand, growth progress and logistics conditions of multiple greenhouses, dynamically generate cross-park production schedules and resource scheduling strategies.

[0038] The multi-agent greenhouse closed-loop control system provided in the above embodiments of the present invention features an edge computing node module with a built-in multi-agent hub based on a full-featured GPU. This hub processes multiple RTSP video streams and environmental sensor data in real time to construct a real-time digital twin of the greenhouse. Through the collaboration of four types of agents—health, space, operation, and prediction—an event bus mechanism is used to transform fuzzy biological perception information into precise mathematical control commands. The cloud-based intelligent computing cluster module is responsible for large-scale model training and global strategy optimization, and distributes the iterated model to the edge, achieving dual-brain collaboration between fast and slow thinking. The system provided in the above embodiments of the present invention replaces human labor with computing power, realizing a closed-loop control throughout the entire process from environmental perception, disease diagnosis, task planning to physical execution.

[0039] Based on the multi-agent greenhouse global closed-loop control system provided in the above embodiments of the present invention, an embodiment of the present invention also provides a greenhouse global closed-loop control method implemented using the above system.

[0040] Specifically, such as Figure 2 As shown, the greenhouse full-area closed-loop control method provided in this embodiment may include: S1, Perception and Modeling: The global perception matrix module collects multi-dimensional data in the greenhouse, and the edge computing node module uses the real-time rendering capability of a full-featured GPU to fuse discrete video streams and sensor data to update the real-time three-dimensional scene of the greenhouse. S2, Diagnosis and Localization: The health agent of the edge computing node module identifies abnormal crop conditions based on a computer vision model, and the spatial agent of the edge computing node module maps the abnormal conditions to specific locations in the world coordinate system of the three-dimensional scene based on a three-dimensional digital twin model, and publishes abnormal events; wherein, the abnormal condition is the abnormal condition carried by the health diagnosis event generated by the health agent; S3, Decision and Distribution: The operational agent of the edge computing node module responds to abnormal events, generates standardized work order control instructions and task path control instructions based on a three-dimensional digital twin model, combined with environmental data and production process knowledge base, and distributes them through the event bus. S4, Execution and Feedback: The embodied execution layer module schedules the corresponding robot or equipment according to the task work order control instructions, plans the optimal path to reach the target location to perform the task operation according to the task path control instructions, and provides real-time feedback on the execution status.

[0041] In some preferred embodiments, the above method may further include: S5, Verification and Optimization: The global perception matrix module automatically collects image data of the work area after the operation, verifies the effect through the edge computing node module, and cleans the data of the entire operation process through the edge computing node module before it flows back to the cloud intelligent computing cluster module for model training and strategy iteration of computer vision model and crop growth mechanism model.

[0042] In some preferred embodiments, in S1 above, the real-time three-dimensional scene of the greenhouse is constructed based on 3D Gaussian sputtering technology, wherein each Gaussian point carries three-dimensional coordinates, color information and semantic labels, which are used by remote management personnel for visual interaction, enabling direct distance measurement, area calculation and object attribute query in three-dimensional space.

[0043] In some preferred embodiments, in S5 above, the data from the entire operation process is cleaned by the edge computing node module and then flowed back to the cloud intelligent computing cluster module, including: The data return process follows the privacy protection principle of data not leaving the park. The edge computing node module performs data cleaning processing such as desensitization, encryption or federated learning gradient extraction on sensitive data (such as core biological asset characteristics and key process parameters) before transmitting it to the cloud intelligent computing cluster module. Non-sensitive data (such as environmental monitoring data and equipment operation logs) are uploaded in full for big data analysis.

[0044] In some preferred embodiments, in S3 above, when the edge computing node module generates task work order control instructions, it simultaneously calls the deduction results of the crop growth mechanism model. The deduction results are generated by the predictive agent based on the current environmental field data, the crop growth mechanism model, and health diagnosis events issued by the healthy agent. By running two simulation branches in parallel, one containing abnormal states and one not containing abnormal states, the difference between the two yield curves is used to quantitatively evaluate the impact of the current abnormal state on the final crop yield. If, in the deduction results, the impact of the current abnormal state on the final crop yield is lower than a preset threshold, or if the current environmental field data is not suitable for task operation (e.g., excessive sunlight makes spraying unsuitable), the operating agent of the edge computing node module will suspend or cancel the generation of task work order control instructions, thereby realizing intelligent decision-making based on economic thresholds.

[0045] It should be noted that the steps in the method provided by the present invention can be implemented using corresponding modules, devices, units, etc. in the system. Those skilled in the art can refer to the technical solution of the system to implement the steps and flow of the method. That is, the embodiments in the system can be understood as preferred examples of the method, and will not be elaborated here.

[0046] The specific application effects of the technical solutions provided in the above embodiments of the present invention will be further explained in detail below with reference to a specific application example.

[0047] In this specific application example, the multi-agent greenhouse full-domain closed-loop control system is based on full-function GPU computing power and edge-cloud collaborative architecture, and mainly consists of four layers: full-domain perception matrix module, edge computing node module, cloud intelligent computing cluster module, and embodied execution layer module.

[0048] The basic features of this system will be designed and constructed from the following three aspects.

[0049] I. Core Architecture Design Full-featured GPU base: Unlike traditional CPUs or dedicated AI chips, this system uses chips based on the MUSA architecture (such as Moore's Threads full-featured GPUs) for its edge computing node modules. This choice is based on the complexity of agricultural scenarios: it requires a Video Engine to decode hundreds or thousands of high-definition monitoring video streams, a Graphics Engine to render 3D digital twin interfaces in real time, an AI Engine to run high-concurrency plant phenotypic analysis models, and a Physics Engine to assist in robot path planning simulations. The all-in-one nature of the full-featured GPU eliminates the latency of data transfer between different chips, achieving ultimate real-time performance.

[0050] Edge-Cloud Collaborative Dual-Brain Mechanism: This system adopts a dual-brain architecture of Nexeed (edge) × KUAE (cloud).

[0051] Nexeed (on the edge): As a System 1 system, it is responsible for millisecond-level real-time responses. It processes real-time data streams from the field, executes pre-trained policy models, and directly controls fans, water valves, and robots. Importantly, it has the ability to operate even when the network connection is interrupted. In extreme cases where the connection to the cloud is lost, it can still ensure the normal operation of the greenhouse's basic life support system based on the local model, safeguarding the bottom line of agricultural production safety.

[0052] Cloud-based (KUAE): As a System 2 system, it possesses near-infinite computing power and massive storage. It aggregates historical data across greenhouses and seasons, is responsible for training large-scale general growth models, performing full-element virtual simulations, and calculating globally optimal strategies.

[0053] II. Multi-Agent Cooperation Mechanism This system innovatively introduces multi-agent technology, decomposing complex agricultural management tasks to four specialized digital experts: Crop Health Agent: Acting as a plant doctor. It uses multispectral analysis and computer vision models to monitor each crop in real time. It can not only identify visible lesions (such as powdery mildew and spider mites), but also detect invisible stress conditions (such as water shortage and nitrogen deficiency) in advance through spectral changes.

[0054] The Spatial Agent acts as a navigator. Based on 3DGS technology, it constructs a centimeter-level high-precision 3D map of the greenhouse. It knows the exact location of every pillar, every pipe, and every flower in the greenhouse. When the healthy agent detects disease, the Spatial Agent is responsible for converting its pixel coordinates into 3D coordinates in the physical world.

[0055] Operation Agent: Acts as a scheduler. It is the core of decision-making, receiving diagnostic reports from the health agent and combining them with agronomic SOPs (e.g., spraying pesticide A within 24 hours upon detecting powdery mildew) to generate specific task orders. It is responsible for scheduling various previously isolated devices (such as drones, AMR robots, and fertilizer applicators) to work collaboratively.

[0056] The prediction agent acts as an actuary. It continuously runs growth prediction models to answer the question of how much future output will increase if this operation is performed now. It provides economic assessments for the operational agent's decision-making, avoiding inefficient inputs.

[0057] III. Closed-Loop Control This system does more than just provide suggestions; it achieves a complete closed loop of perception, decision-making, execution, and verification through the Event Bus.

[0058] Observe: The global perception matrix collects data.

[0059] Orientation: The spatial agent confirms its location.

[0060] Decide: The task agent generates work orders.

[0061] Action (Act): The embodied execution layer (robot) performs physical operations.

[0062] Verification: This is the most crucial step. After the robot finishes its work, the system automatically calls upon sensing devices for a second check (e.g., taking another photo after spraying) to confirm whether the damage has been eliminated. Only after successful verification is the task considered complete; otherwise, rework is triggered.

[0063] Optimization (Learn): All closed-loop data is fed back to the cloud to train a stronger model.

[0064] Based on the above design and construction, the resulting system was applied to a modern rose cultivation greenhouse covering an area of ​​5 hectares.

[0065] Edge computing node module: Deploy one Nexeed AI Box, which has a built-in Moore's Threads MTT S-series full-featured GPU card. This device is connected to the greenhouse's core switch via 10 Gigabit fiber optic cable.

[0066] Global Perception Matrix Module: Environmental sensing: Deploy a set of high-precision sensor boxes for every 500 square meters to monitor temperature, humidity, CO2, light (PAR), and airflow speed.

[0067] Visual perception: Ten automated inspection drones equipped with RGB and multispectral cameras are deployed on the roof track of the greenhouse; 50 fixed-point high-definition monitoring cameras are deployed in key areas.

[0068] The embodied execution layer module is equipped with 5 wheeled AMR logistics robots and 2 tracked spraying robots, all of which are connected to the greenhouse's Wi-Fi 6 private network.

[0069] Based on the above system, an automated powdery mildew control process is executed, demonstrating how the system handles a specific agricultural anomaly.

[0070] Anomaly detection (System 1 real-time inference): During a routine patrol, an inspection drone captured images of white spots on the leaves of a rose bush in row 5 of section C10. The video stream was transmitted in real-time to the Nexeed AI Box. The AI ​​engine within the GPU ran a disease detection model and identified powdery mildew within 30 milliseconds with a 92% confidence level.

[0071] Event announcement: The health agent generates an event {Event: Disease_Detected, Type: Powdery_Mildew, Severity: Level_2} and publishes it to the Event Bus.

[0072] Spatial positioning: The spatial agent subscribes to the event, and calculates the precise world coordinates (X=120.5m, Y=45.2m, Z=0.8m) and PDO ID (C10-05-26) of the diseased plant by back-projecting the real-time pose of the drone and camera intrinsic parameters using the 3DGS world model, and publishes the {Event: Location_Resolved} event to the Event Bus.

[0073] Task generation and scheduling: The task agent subscribes to location events, queries the SOP knowledge base, and learns that Level 2 powdery mildew requires localized spraying of sulfur suspension. It checks current resources and finds that spraying robot #3 is idle and nearby. Therefore, it generates a task order: {Task_ID: T-20250107-001, Target: [120.5, 45.2, 0.8], Action: Spray_Agent_A,Volume: 50ml}, and publishes {Event: Task_Dispatched}.

[0074] Physical execution: Robot No. 3 received the instruction, planned a path to avoid obstacles, moved to the target point, adjusted the posture of the robotic arm, and performed precise spraying.

[0075] Closed-loop verification: After spraying, the robot sends a status message ({Status: Complete}) to the system. The operational agent does not immediately close the work order but instead generates a delayed re-inspection task. 24 hours later, a drone is dispatched to the coordinates again to take photos. The health agent analyzes the new photos, confirms the white spots have faded, and issues an event ({Event: Outcome_Success}).

[0076] Model Iteration (System 2 Slow Thinking): The entire process data (images, environmental parameters, pesticide dosage, and results) was packaged and uploaded to the cloud-based KUAE platform. The cloud platform then added this data to the training set to update the disease prediction model, resulting in more accurate predictions in the future.

[0077] As can be seen from the above specific applications, the technical solution provided by the above embodiments of the present invention truly realizes unmanned closed-loop management of agricultural production, greatly reduces the dependence on highly skilled workers, and improves the standardization level of production.

[0078] Any matters not covered in the above embodiments of the present invention are well-known in the art.

[0079] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A multi-agent greenhouse full-domain closed-loop control system, characterized in that, include: The module comprises a global perception matrix module, an edge computing node module, a cloud-based intelligent computing cluster module, and an embodied execution layer module; among which: The global perception matrix module is deployed inside the greenhouse to collect multi-dimensional data in the greenhouse in real time, including environmental field data, crop phenotypic image data, and device status data of the embodied execution layer module, and sends it to the edge computing node module through the local area network. The edge computing node module, deployed locally in the greenhouse, is equipped with a full-featured GPU computing unit based on the MUSA architecture to run a multi-agent central system. This multi-agent central system comprises multiple independent and cooperative agents. Each agent utilizes the hardware encoding / decoding capabilities and AI inference capabilities of the full-featured GPU computing unit to perform real-time computation on the collected multi-dimensional data using computer vision models, 3D digital twin models, and crop growth mechanism models. The 3D digital twin model is directly constructed and iteratively updated by the edge computing node module based on local real-time multi-dimensional data to generate control commands. The cloud-based intelligent computing cluster module is connected to the edge computing node module via a wide area network. It is used to store historical production data of the greenhouse, build and run training tasks for computer vision models and crop growth mechanism models, as well as global strategy optimization algorithms across greenhouses. The trained and iterated models are periodically distributed and updated to the edge computing node module. The embodied execution layer module is used to receive control commands issued by the edge computing node module, perform specific task operations within the greenhouse, and feed back the execution results to the edge computing node module for control command correction, thus forming a data closed-loop link.

2. The multi-agent greenhouse global closed-loop control system according to claim 1, characterized in that, The full-featured GPU computing unit integrates a graphics rendering engine, an AI inference engine, a video processing engine, and a physics simulation engine; among which: The video processing engine is used to decode multiple RTSP high-definition video streams from the global perception matrix module in parallel, and to complete the decoding directly in the video memory to obtain video frames. The AI ​​inference engine is used to perform real-time target detection, instance segmentation, and disease feature extraction on the decoded video frames. The real-time target detection results are used to locate the pixel positions and categories of individual crops, lesions, pests, and operating equipment in the image, which are then used by the multi-agent central system for 3D coordinate back-projection. The instance segmentation results are used to obtain the precise pixel boundaries of each crop and each leaf, which are then used by the multi-agent central system for single-plant nutrient diagnosis and generation of precise local spraying areas. The disease feature extraction results are used to output the disease category, severity level, and spread trend, which are then used by the multi-agent central system to generate health diagnosis events and trigger SOP matching. The graphics rendering engine is used to render a three-dimensional greenhouse scene based on 3D Gaussian sputtering technology in real time, providing a visual interaction for remote management personnel. The physical simulation engine is used to perform collision detection and dynamic simulation of the robotic arm path or movement trajectory of the embodied execution layer module in virtual space.

3. The multi-agent greenhouse global closed-loop control system according to claim 1, characterized in that, The multi-agent central system adopts an event bus-based collaborative architecture, including the following four types of agents: The health intelligence agent is used to analyze crop phenotypic image data using computer vision models and combine the instance segmentation results obtained by the AI ​​inference capability of the full-function GPU computing unit to obtain the precise pixel boundaries of each crop and each leaf, perform single-plant nutrient diagnosis, and diagnose the crop's pest and disease status, nutritional status and growth stage in real time. At the same time, based on the disease feature extraction results obtained by the AI ​​inference capability of the full-function GPU computing unit, it outputs the disease category, severity level and spread trend, and generates health diagnosis events to be published to the event bus. The spatial intelligent agent is used to locate the pixel positions and categories of individual crops, lesions, pests, and operating equipment in the image by using the real-time target detection results obtained by the AI ​​inference capability of the full-function GPU computing unit, perform three-dimensional coordinate back projection, maintain a unified spatiotemporal coordinate system in the greenhouse, generate task path control instructions based on the three-dimensional digital twin model, and provide path planning, navigation obstacle avoidance, and object spatial positioning services for the embodied execution layer module. The operation intelligence agent is used to subscribe to the health diagnosis event, trigger SOP matching according to the preset standard operating procedure and the disease feature extraction result, and obtain the local precise spraying area according to the instance segmentation result. Based on the three-dimensional digital twin model, it generates task work order control instructions containing three-dimensional coordinates and action definitions, and distributes them to the embodied execution layer module through the event bus. The predictive agent combines current environmental data, health diagnostic events released by the health agent, and crop growth mechanism models to predict crop yield, quality grade, and market cycle within a preset time period. During the prediction process, two parallel simulation branches are constructed, one including abnormal states and one not including abnormal states, to quantitatively evaluate the impact of the current abnormal state on the final yield, serving as the basis for the economic threshold decision of the operational agent.

4. The multi-agent greenhouse global closed-loop control system according to claim 3, characterized in that, The task work order control instructions generated by the task intelligence agent include: a unique task identifier ID, the three-dimensional spatial coordinates of the target object, the task type code, the execution priority, and the completion verification standard; The autonomous mobile robot in the embodied execution layer module, after receiving the task work order control command, uses its onboard sensors to scan passive visual markers in the greenhouse for secondary precise positioning, ensuring that the error between the physical operation position and the three-dimensional spatial coordinates in the task work order control command is less than a preset threshold.

5. The multi-agent greenhouse global closed-loop control system according to claim 3, characterized in that, Also includes: The end-to-end closed-loop verification mechanism trigger module; including: After the physical operation is completed by the embodied execution layer module, the full-process closed-loop verification mechanism trigger module automatically triggers the re-inspection process instruction, controls the full-domain perception matrix module to automatically collect post-operation image data of the operation area, and sends it to the edge computing node module. The health agent or operation agent of the edge computing node module compares and analyzes the post-operation image data to verify whether the operation effect meets the completion verification standard; if the verification fails, a rework task work order control instruction is automatically generated and the execution priority is set and fed back to the embodied execution layer module. The edge computing node module simultaneously feeds back the entire operation process data to the cloud-based intelligent computing cluster module for training and strategy iteration of computer vision models and crop growth mechanism models; wherein, the strategy includes one or more of the following: disease identification and treatment strategy, environmental field regulation strategy, operation task scheduling strategy, energy consumption optimization strategy, and water and fertilizer formula strategy.

6. A method for closed-loop control of the entire greenhouse area using the system described in any one of claims 1-5, characterized in that, include: The global perception matrix module collects multi-dimensional data from inside the greenhouse, while the edge computing node module utilizes the real-time rendering capabilities of a full-featured GPU to fuse discrete video streams with sensor data and update the real-time 3D scene of the greenhouse. The edge computing node module identifies abnormal crop conditions based on computer vision models, maps the abnormal conditions to specific locations in the world coordinate system of the 3D scene based on a 3D digital twin model, and publishes and responds to abnormal events. The three-dimensional digital twin model combines environmental data and production process knowledge base to generate standardized work order control instructions and task path control instructions, which are then distributed through an event bus. The embodied execution layer module schedules the corresponding robot or equipment according to the task work order control instructions, plans the optimal path to reach the target location to perform the task operation according to the task path control instructions, and provides real-time feedback on the execution status.

7. The greenhouse whole-area closed-loop control method according to claim 6, characterized in that, Also includes: The global perception matrix module automatically collects image data of the work area after the operation and then verifies the effect through the edge computing node module. The data of the entire operation process is cleaned by the edge computing node module and then fed back to the cloud intelligent computing cluster module for model training and strategy iteration of computer vision model and crop growth mechanism model.

8. The greenhouse whole-area closed-loop control method according to claim 6, characterized in that, The real-time 3D scene of the greenhouse is constructed based on 3D Gaussian sputtering technology, in which each Gaussian point carries 3D coordinates, color information and semantic labels, which are used for visual interaction by remote management personnel.

9. The greenhouse whole-area closed-loop control method according to claim 6, characterized in that, The entire operation process data is cleaned by the edge computing node module and then fed back to the cloud intelligent computing cluster module, including: The data backflow process follows the principle of privacy protection. The edge computing node module performs data cleaning processing such as desensitization, encryption or federated learning gradient extraction on sensitive data before transmitting it to the cloud intelligent computing cluster module. Non-sensitive data is uploaded in full for big data analysis.

10. The greenhouse whole-area closed-loop control method according to claim 6, characterized in that, When generating task work order control instructions, the edge computing node module simultaneously invokes the simulation results of the crop growth mechanism model. The simulation results are generated by the predictive agent based on the current environmental field data, the crop growth mechanism model, and health diagnosis events issued by the healthy agent. By running two simulation branches in parallel, one containing abnormal states and the other not containing abnormal states, the difference between the two yield curves is used to quantitatively evaluate the impact of the current abnormal state on the final crop yield. If, in the simulation results, the impact of the current abnormal state on the final crop yield is lower than a preset threshold, or if the current environmental field data is not suitable for task operation, the operating agent of the edge computing node module will suspend or cancel the generation of task work order control instructions, realizing intelligent decision-making based on economic thresholds.

Images