Agricultural robot control method based on multi-agent cooperation

Abstract

The invention discloses an agricultural robot control method based on multi-agent cooperation, and the method comprises the following steps: S1, collecting environment data of a farmland region, and carrying out the preprocessing of the environment data; s2, constructing a job task graph, and forming structured task graph data; s3, inputting the operation task graph into the graph neural network model, and performing task initialization on all agricultural robots; s4, sharing the local operation sub-graphs among the agricultural robots, and fusing all the local operation sub-graphs based on a graph neural network model; s5, executing a multi-agent cooperative task division operation, and generating a task allocation matrix; s6, after the agricultural robot receives the operation instruction, executing a path planning and obstacle avoidance algorithm; and S7, summarizing operation states of the agricultural robots in real time, and dynamically updating weight parameters of the graph neural network model according to operation feedback. Efficient cooperation and intelligent scheduling of the agricultural robot are realized, and the operation efficiency and the coverage rate are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural automation control technology, and in particular to a kind of agricultural robot control method based on multi-agent cooperation. BACKGROUND

[0002] Under the background of rapid development of agricultural production automation, agricultural robots have been widely used in sowing, fertilizing, spraying and harvesting, etc., to replace manual work to improve work efficiency and quality. However, the existing agricultural robot control method mostly adopts single machine independent operation mode, lacks multi-machine cooperation mechanism, which leads to low work efficiency, high path repetition rate, uneven task allocation in large area farmland or complex terrain conditions, and seriously affects the overall performance of work.

[0003] The existing control method generally relies on static task planning and fixed path execution, which is difficult to cope with real-time environmental changes and obstacle interference; at the same time, in multi-robot cooperative work, there is a lack of effective information sharing and task coordination strategy, which is easy to cause conflicts between robots, overlapping coverage area or blind area omission, etc. In addition, the current task allocation mechanism is mostly based on preset rules or static area division, lacking dynamic perception and utilization of environmental elements such as terrain features, soil state and crop distribution, which limits the adaptability of the system in variable environment.

[0004] Therefore, how to provide an agricultural robot control method based on multi-agent cooperation is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide an agricultural robot control method based on multi-agent cooperation, which fully utilizes the technologies of graph neural network, task graph modeling and multi-agent cooperative decision making, and describes in detail the whole process of cooperative work of agricultural robots in complex farmland environment through information sharing, task division, path planning and work state feedback, which has the advantages of high work efficiency, reasonable resource allocation and strong environmental adaptability.

[0006] According to an agricultural robot control method based on multi-agent cooperation according to an embodiment of the present application, the method comprises the following steps:

[0007] S1, collecting environmental data of farmland area and pre-processing;

[0008] S2, based on the pre-processed environmental data, constructing a work task graph, the nodes of the work task graph representing work units, and the edges representing the spatial adjacency relationship between work units, forming structured task graph data;

[0009] S3, input the job task graph into the graph neural network model, extract the graph embedding vector of each job unit, and initialize the task of all agricultural robots, and construct the corresponding local job subgraph;

[0010] S4, share the local job subgraph between each agricultural robot through point-to-point communication mode, and fuse all local job subgraphs based on the graph neural network model to generate a global collaborative job graph;

[0011] S5, according to the global collaborative job graph, perform multi-agent collaborative task division operation to generate task allocation matrix, and according to the task allocation matrix, issue job instructions to corresponding agricultural robots;

[0012] S6, after the agricultural robot receives the job instruction, combines the current position information and obstacle position, executes the path planning and obstacle avoidance algorithm, generates the job path and synchronously uploads the job state;

[0013] S7, real-time summary of the job state of each agricultural robot, and dynamically updating the weight parameters of the graph neural network model according to the job feedback until completing all job tasks of the target farmland.

[0014] Optionally, the environmental data includes terrain structure, soil moisture, crop distribution and obstacle position.

[0015] Optionally, the preprocessing includes coordinate standardization, data interpolation completion, noise filtering and rasterization mapping.

[0016] Optionally, the S2 specifically includes:

[0017] S21, according to the terrain structure information after preprocessing, perform uniform coordinate conversion operation and planarization processing, and remove the non-workable farmland area;

[0018] S22, divide the farmland area into a plurality of standard grid units by using fixed edge length division method, each standard grid unit corresponds to a candidate job unit, and the position information and boundary coordinates of each grid unit are retained;

[0019] S23, spatially align the preprocessed soil moisture data and crop distribution image data, and use point-by-point raster mapping method to map all soil moisture information and crop distribution results to the corresponding standard grid unit;

[0020] S24, for the standard grid unit with missing data, perform multi-directional weighted interpolation completion processing according to the surrounding soil moisture and crop distribution, and delete the irregular noise block identified in the raster mapping process;

[0021] S25. Performing a spatial segmentation operation based on the obstacle location data, marking obstacle units in the standard grid units, setting non-operational flags, and constructing a complete operation candidate layer to form a farmland layer structure consisting of operation units and obstacle units.

[0022] S26. Traverse all the operation units and obtain edge information of the operation task graph based on the adjacent relationship of the operation units in the farmland layer structure. If any two operation units have common edges in contact in the horizontal, vertical, or diagonal direction, establish a connecting edge in the operation task graph;

[0023] S27. All work units constitute a node set of the work task graph, and the spatial adjacency relationship between the nodes is used as an edge set to form structured task graph data for collaborative control.

[0024] Optionally, the S3 specifically includes:

[0025] S31. Input the task graph into the graph neural network model, where the input content includes the attribute information of the task units corresponding to the nodes and the spatial adjacency relationship described by the edges;

[0026] S32. Perform a structural initialization operation in the graph neural network model to convert each work unit node into a graph embedding vector with attributes, and establish an edge weight encoding to reflect the adjacency strength and work coupling relationship between work units;

[0027] S33. Aggregate the adjacent node information of each work unit in the graph neural network through a multi-round graph information transmission mechanism, and update the graph embedding vector of each work unit so that the graph embedding vector contains both the local topological structure and the work unit attribute information;

[0028] S34, performing matching analysis on the graph embedding vector and the current position and operation capability parameters of the agricultural robot, performing task initialization operations on all agricultural robots based on the matching results, and determining a set of operation areas that each agricultural robot can cover;

[0029] S35. Based on the set of operation areas, extract the subgraph area overlapping with the position of each agricultural robot in the operation task graph, construct the operation unit nodes and their associated edges contained in the subgraph into a local operation subgraph, and complete the operation attribute binding;

[0030] S36. Establish a one-to-one mapping relationship between each local operation subgraph and the corresponding agricultural robot, which serves as an input basis for operation task division and path planning.

[0031] Optionally, the multi-round graph information transmission mechanism refers to iteratively aggregating and updating the adjacent node features of each work unit in the graph neural network, so that the graph embedding vector of each work unit fuses local structural relationships and attribute information.

[0032] Optionally, the S4 specifically comprises:

[0033] S41, after the agricultural robots complete the local work subgraph construction, starting the point-to-point communication mechanism, each agricultural robot sends the node set and edge set of the local work subgraph held by it to other agricultural robots within the communication radius range through short-distance wireless communication;

[0034] S42, after receiving the local work subgraph data sent by the agricultural robots, sequentially performing the graph structure splicing operation, performing position verification and boundary fusion on the work unit nodes shared by adjacent agricultural robots, eliminating duplicate nodes and edge information, and constructing an extended joint work subgraph;

[0035] S43, after the bidirectional sharing and graph structure splicing of the local work subgraph of each agricultural robot are completed, sequentially uploading the summary data to the graph neural network model, and performing feature fusion operation on the joint work subgraph through the graph neural network model, combining the structural features and attribute features of the nodes in each subgraph, and uniformly encoding the node representation;

[0036] S44, based on the joint work subgraph shared and fused by all agricultural robots, constructing a global collaborative work graph covering the entire farmland area, the global collaborative work graph containing all work unit nodes, complete edge connection relationship and work unit attribute information;

[0037] S45, synchronously returning the global collaborative work graph to the local cache of each agricultural robot.

[0038] Optionally, the S5 specifically comprises:

[0039] S51, based on the spatial position, task type and work complexity of each work unit in the global collaborative work graph, extracting the task feature information of all nodes in the graph, and simultaneously collecting the current position, motion ability and task carrying capacity information of all agricultural robots;

[0040] S52, according to the spatial relationship between the work unit and the agricultural robot, the task adaptation degree and the adjacent structure in the graph, performing multi-agent collaborative task division operation, using regional aggregation method to divide the work unit into multiple non-intersecting subtask regions, each subtask region corresponding to an agricultural robot;

[0041] S53, constructing a task allocation matrix according to the divided subtask region, identifying the work unit number, work order and task content corresponding to each agricultural robot in the task allocation matrix, ensuring that the work allocation has coverage, balance and execution accessibility;

[0042] S54, the instruction content in the task allocation matrix is structured and packaged to generate a work instruction package for the agricultural robot, each work instruction package containing a target work unit number, a work type identifier, a priority level and a completion condition parameter;

[0043] S55, the work instruction package is respectively issued to the corresponding agricultural robot terminal through short-range communication, and the instruction issuing state is recorded in the work scheduling module to ensure that each agricultural robot obtains a work task matching its capability.

[0044] Optionally, the task allocation matrix refers to a mapping relationship table in which the work units are allocated to specific agricultural robots based on the spatial proximity between the agricultural robots and the work units, the work capacity matching degree and the task load balancing principle on the basis of multi-agent collaborative task division.

[0045] Optionally, the S6 specifically comprises:

[0046] S61, after the agricultural robot receives the issued work instruction, the spatial position, task type and execution order parameters of the target work unit are parsed therefrom, and the current position information is obtained in real time;

[0047] S62, the obstacle position data of the corresponding work area is loaded, the current position information and the spatial position of the target work unit are input into the path calculation module, and the work path planning process is initialized;

[0048] S63, based on the work path planning process, a grid connected graph of the farmland work space is constructed, a set of passable paths is constructed in combination with the obstacle distribution, an optimal connected path is extracted and a work path sequence is generated;

[0049] S64, during the generation of the work path sequence, a path node screening operation is performed according to the obstacle boundary information, the path nodes overlapping or too close to the obstacles in the path are removed, and the path direction is structurally adjusted;

[0050] S65, the agricultural robot performs the moving work process according to the optimized work path sequence, and records the current position, task completion state and execution time data after completing each work unit;

[0051] S66, the recorded work state data is uploaded to the work scheduling module through the communication interface to complete the real-time synchronization and update of the work state.

[0052] The beneficial effects of the present application are:

[0053] Firstly, the application realizes deep modeling of the spatial structure and operation attributes of farmland by constructing a task graph and introducing a graph neural network model, enabling the agricultural robot to extract graph embedding features of the operation unit on the basis of multi-dimensional perception, providing accurate data support for subsequent task division and path planning, thereby effectively improving the intelligence and pertinence of operation decision-making.

[0054] Secondly, a multi-agent collaborative mechanism is adopted to realize local operation subgraph sharing and fusion among agricultural robots through point-to-point communication, generate a unified global collaborative operation graph, perform dynamic task division and build a task allocation matrix, which can significantly improve the coverage integrity of operation tasks and the balance of collaborative operation of robots, and avoid path conflicts and resource waste.

[0055] Finally, the current position information and obstacle position are combined to perform path planning and obstacle avoidance operation, and the operation status is uploaded in real time to form a closed-loop control process, ensuring the running stability and operation continuity of the agricultural robot in a dynamic environment, realizing the transformation of agricultural operation from static planning to intelligent collaboration, and having significant application promotion value and industrial landing potential. BRIEF DESCRIPTION OF DRAWINGS

[0056] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0057] Fig. 1 A flowchart of an agricultural robot control method based on multi-agent collaboration proposed by the application;

[0058] Fig. 2 A task graph construction flowchart of an agricultural robot control method based on multi-agent collaboration proposed by the application;

[0059] Fig. 3 A local operation subgraph construction flowchart of an agricultural robot control method based on multi-agent collaboration proposed by the application. DETAILED DESCRIPTION

[0060] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0061] REFERENCE Figs. 1-3 An agricultural robot control method based on multi-agent collaboration, comprising the following steps:

[0062] S1, collecting environmental data of the farmland area and performing preprocessing;

[0063] S2, based on the pre-processed environment data, a job task graph is constructed, the nodes of the job task graph represent job units, and the edges represent the spatial adjacency relationship between the job units, forming structured task graph data;

[0064] S3, inputting the job task graph into a graph neural network model, extracting a graph embedding vector of each job unit, and initializing tasks for all agricultural robots to construct corresponding local job subgraphs;

[0065] S4, sharing the local job subgraph between each agricultural robot through a point-to-point communication mode, and merging all local job subgraphs based on the graph neural network model to generate a global collaborative job graph;

[0066] S5, according to the global collaborative job graph, performing a multi-agent collaborative task division operation to generate a task allocation matrix, and according to the task allocation matrix, issuing a job instruction to the corresponding agricultural robot;

[0067] S6, after the agricultural robot receives the job instruction, combining the current position information and the obstacle position to execute a path planning and obstacle avoidance algorithm, generating a job path and synchronously uploading a job state;

[0068] S7, real-time summary of the job state of each agricultural robot, and dynamically updating the weight parameters of the graph neural network model according to the job feedback until the entire job task of the target farmland is completed.

[0069] The present application realizes information fusion, task division and path generation among agricultural robots by constructing a job task graph and combining a graph neural network, and improves the overall efficiency of multi-robot collaborative operation and the intelligent level of farmland operation.

[0070] In the embodiment, the environment data includes terrain structure, soil moisture, crop distribution and obstacle position.

[0071] The present application collects and utilizes multi-source environment data such as terrain structure, soil moisture, crop distribution and obstacle position, and provides a comprehensive and accurate environment basis for subsequent task graph construction and job path planning, improving the accuracy of data-driven decision-making.

[0072] In the embodiment, the preprocessing includes coordinate standardization, data interpolation completion, noise filtering and rasterization mapping.

[0073] The present application improves the processing quality of farmland data by performing coordinate standardization, interpolation completion, noise filtering and rasterization mapping on the environment data, ensuring the coherence of task graph construction and the rationality of job area division.

[0074] In the embodiment, the S2 specifically includes:

[0075] S21, according to the pre-processed terrain structure information, performing uniform coordinate conversion operation and planarization processing, and eliminating the non-workable farmland area;

[0076] S22, the farmland area is divided into a plurality of standard grid units in a fixed side length division manner, each standard grid unit corresponds to a candidate work unit, and the position information and boundary coordinates of each grid unit are retained;

[0077] S23, the pre-processed soil moisture data and the crop distribution image data are spatially aligned, and all soil moisture information and crop distribution results are mapped into the corresponding standard grid unit by using point-by-point grid mapping;

[0078] S24, for the standard grid unit with missing data, the surrounding soil moisture and crop distribution are used for multi-directional weighted interpolation completion processing, and the irregular noise block identified in the grid mapping process is deleted;

[0079] S25, based on the obstacle position data, performing spatial subdivision operation, marking the obstacle unit in the standard grid unit, and setting the non-workable identifier, constructing a complete work candidate layer, and forming a farmland layer structure composed of work units and obstacle units;

[0080] S26, traversing all work units, obtaining edge information of the work task graph according to the adjacent relationship of the work units in the farmland layer structure, and establishing a connection edge in the work task graph if any two work units have a common edge contact in the horizontal, vertical or diagonal direction;

[0081] S27, all work units constitute a node set of the work task graph, and the spatial adjacency relationship between the nodes is taken as an edge set, forming a structured task graph data for collaborative control.

[0082] The application refines the environmental data processing flow, accurately divides the work units and obstacle areas, and constructs a clear layer structure, providing an accurate atlas basis for collaborative path calculation and task planning.

[0083] In the embodiment, the S3 specifically comprises:

[0084] S31, inputting the work task graph into the graph neural network model, and the input content includes the work unit attribute information corresponding to the node and the spatial adjacency relationship described by the edge;

[0085] S32, performing structure initialization operation in the graph neural network model, converting each work unit node into a graph embedding vector with attributes, and establishing edge weight coding to reflect the adjacency strength and work coupling relationship between work units;

[0086] S33, through the multi-round graph information transmission mechanism, the adjacency node information of each job unit is aggregated in the graph neural network, and the graph embedding vector of each job unit is updated, so that the graph embedding vector contains local topological structure and job unit attribute information at the same time.

[0087] S34, the graph embedding vector is matched and analyzed with the current position of the agricultural robot and the job capacity parameter, and task initialization operation is performed on all agricultural robots according to the matching result, and the job area set that can be covered by each agricultural robot is determined.

[0088] S35, based on the job area set, the subgraph area overlapping with the position of each agricultural robot is extracted in the job task graph, the job unit nodes and their associated edges contained in the subgraph are constructed into a local job subgraph, and job attribute binding is completed.

[0089] S36, each local job subgraph is mapped to the corresponding agricultural robot one by one, as the input basis of job task division and path planning.

[0090] The application extracts the graph embedding vector and constructs the local job subgraph, so that the agricultural robot has the ability of autonomous perception and regional initialization, and realizes the accurate matching between the job task and the robot capability.

[0091] In the embodiment, the multi-round graph information transmission mechanism means that the features of the adjacent nodes of each job unit are iteratively aggregated and updated in the graph neural network, so that the graph embedding vector of each job unit fuses local structure relationship and attribute information.

[0092] The application fuses the local structure and attribute features of the job unit through the multi-round graph information transmission mechanism, improves the modeling depth and expression ability of the graph neural network in the farmland job task.

[0093] In the embodiment, the S4 specifically includes:

[0094] S41, after the agricultural robot completes the construction of the local job subgraph, a point-to-point communication mechanism is started, and each agricultural robot sends the node set and edge set of the local job subgraph held by it to other agricultural robots within the communication radius range through short-distance wireless communication;

[0095] S42, after receiving the local job subgraph data sent by the agricultural robot, the graph structure splicing operation is sequentially performed, the job unit nodes shared by adjacent agricultural robots are position checked and boundary fused, the repeated nodes and edge information are eliminated, and the expanded joint job subgraph is constructed.

[0096] S43, after each agricultural robot completes the bidirectional sharing of the local operation subgraph and the graph structure splicing, sequentially upload the summary data to the graph neural network model, perform feature fusion operation on the joint operation subgraph through the graph neural network model, and uniformly encode the node representation by combining the structural features and attribute features of the nodes in each subgraph;

[0097] S44, based on the joint operation subgraph shared and fused by all agricultural robots, a global collaborative operation graph covering the entire farmland area is constructed, and the global collaborative operation graph includes all operation unit nodes, complete edge connection relationships and operation unit attribute information;

[0098] S45, the local collaborative operation graph is returned to the local cache of each agricultural robot.

[0099] The application fuses the local operation subgraph through point-to-point communication, constructs a global collaborative operation graph, effectively improves the data sharing efficiency and global task collaboration ability between robots, and enhances the overall consistency of the system.

[0100] In the embodiment, the S5 specifically includes:

[0101] S51, based on the spatial position, task type and operation complexity of each operation unit in the global collaborative operation graph, the task feature information of all nodes in the graph is extracted, and the current position, motion ability and task carrying capacity information of all agricultural robots are collected at the same time;

[0102] S52, according to the spatial relationship between the operation unit and the agricultural robot, the task adaptation degree and the adjacent structure in the graph, a multi-agent collaborative task division operation is performed, and the operation unit is divided into a plurality of non-intersecting subtask regions by using a region aggregation method, each subtask region corresponds to an agricultural robot;

[0103] S53, a task allocation matrix is constructed according to the divided subtask region, and the operation unit number, operation sequence and task content corresponding to each agricultural robot are marked in the task allocation matrix, so that the operation allocation has coverage, balance and execution accessibility;

[0104] S54, the instruction content in the task allocation matrix is structured and packaged to generate an operation instruction package for the agricultural robot, and each operation instruction package includes target operation unit number, operation type identification, priority level and completion condition parameters;

[0105] S55, the operation instruction package is respectively sent to the corresponding agricultural robot terminal through short-distance communication, and the instruction sending state is recorded in the operation scheduling module, so that each agricultural robot obtains an operation task matched with its ability.

[0106] The application realizes the fairness, integrity and executability of task allocation by performing multi-agent task division on a synergy map and generating a task allocation matrix, and avoids resource waste and path conflict problems.

[0107] In the embodiment, the task allocation matrix refers to a mapping relationship table of allocating work units to specific agricultural robots according to the spatial proximity, work capacity matching degree and task load balancing principle between the agricultural robots and the work units on the basis of multi-agent collaborative task division.

[0108] The task allocation matrix constructed by the application accurately reflects the matching relationship between the work units and the agricultural robots, realizes the transparent controllability and dynamic optimization of the task allocation process.

[0109] In the embodiment, the S6 specifically comprises:

[0110] S61, after receiving the issued work instruction, the agricultural robot parses the spatial position, task type and execution order parameters of the target work unit, and obtains the current position information in real time;

[0111] S62, load the obstacle position data of the corresponding work area, input the current position information and the spatial position of the target work unit into the path calculation module, and initialize the work path planning process;

[0112] S63, construct a grid connected graph of farmland work space based on the work path planning process, construct a set of passable paths combined with the obstacle distribution, extract the optimal connected path and generate a work path sequence;

[0113] S64, during the work path sequence generation process, perform a path node screening operation according to the obstacle boundary information, remove the path nodes in the path that overlap or are too close to the obstacles, and adjust the structure of the path direction;

[0114] S65, the agricultural robot performs the moving work process according to the optimized work path sequence, and records the current position, task completion state and execution time data after executing each work unit;

[0115] S66, upload the recorded work state data to the work scheduling module through the communication interface to complete the real-time synchronization and update of the work state.

[0116] Through the path planning and obstacle avoidance control process, the application realizes the autonomous navigation and state synchronization feedback of the robot in a complex environment, ensures the continuity, safety and real-time visualization of information of the work process.

[0117] Example 1:

[0118] In order to verify the feasibility of the application in implementation, the application is applied to the spring ploughing stage operation of a certain farmland area, and the task target is to perform efficient and collaborative fertilization and pest and disease patrol operation on more than 100 irregular plots. The traditional agricultural robot has problems such as scattered deployment, frequent path conflict, high task repetition execution rate and the like in this scene, and cannot realize efficient collaboration and intelligent task scheduling among multiple robots, resulting in incomplete operation coverage, increased time cost and energy waste.

[0119] In this embodiment, six agricultural robots with environment perception ability and mobile operation ability are deployed, and a terrain scanning device is used to collect data of the target farmland area, and obtain multi-source environment data such as ground elevation, soil moisture, crop distribution and obstacle position. The data is standardized in coordinates, noise is cleaned and raster division is performed by the method of the application, a structured operation task graph containing 784 effective operation units is constructed, and the graph is imported as input into a graph neural network model to extract the graph embedding vector of each node. After constructing a local operation subgraph at the robot end, subgraph sharing is completed through point-to-point communication, and feature fusion is performed in the graph neural network to generate a global collaborative operation graph covering the whole farmland.

[0120] Based on the collaborative graph, a task division strategy is executed, each robot is allocated to a corresponding operation unit area according to the operation intensity, task type and obstacle distribution, a task allocation matrix is generated and issued to the robot terminal. In the actual operation process, each robot performs path optimization and dynamic obstacle avoidance according to the current position and obstacle layout, and automatically synchronously uploads the task status after completing the operation. The system continuously collects operation feedback information, and dynamically adjusts the weight parameters of the graph model to optimize the subsequent allocation strategy.

[0121] The actual running results show that the total operation time of the six robots is 3.2 hours, the average operation coverage rate reaches 98.6%, the task repetition rate is controlled within 1.3%, and the number of obstacle avoidance and detour is significantly lower than that of the traditional random path control method; compared with the traditional single machine independent operation mode, the overall operation efficiency is improved by about 41.8%, the energy consumption is reduced by 22.5%, and the average score of operation quality is improved to 92 points, which fully verifies the collaborative control ability and task scheduling efficiency of the application in the complex terrain and multi-target task scene.

[0122] Table 1 Comparison table of measured data of multi-agent agricultural robot collaborative operation

[0123]

[0124] This embodiment shows that the method of the application can effectively realize dynamic task division, path optimization and operation collaborative control of multiple agricultural robots in irregular farmland, and has higher running efficiency, lower energy consumption and better operation quality in complex environment.

[0125] The above description is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can make equivalent substitutions or changes according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A control method for agricultural robots based on multi-agent collaboration, characterized in that: The steps include: S1. Collect environmental data of farmland areas and perform preprocessing; S2. Based on the preprocessed environmental data, construct an operation task graph, where the nodes of the operation task graph represent operation units and the edges represent the spatial adjacency relationship between the operation units, thereby forming structured task graph data; S3. Input the task graph into the graph neural network model, extract the graph embedding vector of each task unit, initialize the tasks for all agricultural robots, and construct the corresponding local task subgraph; S4. Share local operation subgraphs among agricultural robots through point-to-point communication. Then, fuse all local operation subgraphs based on the graph neural network model to generate a global collaborative operation graph. S5. Based on the global collaborative operation map, perform multi-agent collaborative task division operations, generate a task allocation matrix, and issue operation instructions to corresponding agricultural robots according to the task allocation matrix; S6. After receiving the operation instruction, the agricultural robot combines the current position information with the obstacle position, executes the path planning and obstacle avoidance algorithm, generates the operation path and simultaneously uploads the operation status; S7. Summarize the operating status of each agricultural robot in real time, and dynamically update the weight parameters of the graph neural network model based on the operating feedback until all operating tasks on the target farmland are completed.

2. The agricultural robot control method based on multi-agent collaboration according to claim 1 is characterized in that: The environmental data includes terrain structure, soil moisture, crop distribution and obstacle locations.

3. The agricultural robot control method based on multi-agent collaboration according to claim 1 is characterized in that: The preprocessing includes coordinate standardization, data interpolation and completion, noise filtering and rasterization mapping.

4. The agricultural robot control method based on multi-agent collaboration according to claim 1, characterized in that: The S2 specifically includes: S21, performing a unified coordinate conversion operation and planarization process based on the pre-processed terrain structure information, and eliminating inoperable farmland areas; S22, dividing the farmland area into a plurality of standard grid cells using a fixed side length division method, where each standard grid cell corresponds to a candidate operation unit, and retaining the position information and boundary coordinates of each grid cell; S23, spatially aligning the pre-processed soil moisture data with the crop distribution image data, and mapping all soil moisture information and crop distribution results to corresponding standard grid cells using a point-by-point grid mapping method; S24. For missing standard grid cells, multi-directional weighted interpolation is performed based on the surrounding soil moisture and crop distribution, and irregular noise blocks identified during the grid mapping process are deleted. S25. Performing a spatial segmentation operation based on the obstacle location data, marking obstacle units in the standard grid units, setting non-operational flags, and constructing a complete operation candidate layer to form a farmland layer structure consisting of operation units and obstacle units. S26. Traverse all the operation units and obtain edge information of the operation task graph based on the adjacent relationship of the operation units in the farmland layer structure. If any two operation units have common edges in contact in the horizontal, vertical, or diagonal direction, establish a connecting edge in the operation task graph; S27. All work units constitute a node set of the work task graph, and the spatial adjacency relationship between the nodes is used as an edge set to form structured task graph data for collaborative control.

5. The agricultural robot control method based on multi-agent collaboration according to claim 1 is characterized in that: The S3 specifically includes: S31. Input the task graph into the graph neural network model, where the input content includes the attribute information of the task units corresponding to the nodes and the spatial adjacency relationship described by the edges; S32. Perform a structural initialization operation in the graph neural network model to convert each work unit node into a graph embedding vector with attributes, and establish an edge weight encoding to reflect the adjacency strength and work coupling relationship between work units; S33. Aggregate the adjacent node information of each work unit in the graph neural network through a multi-round graph information transmission mechanism, and update the graph embedding vector of each work unit so that the graph embedding vector contains both the local topological structure and the work unit attribute information; S34, performing matching analysis on the graph embedding vector and the current position and operation capability parameters of the agricultural robot, performing task initialization operations on all agricultural robots based on the matching results, and determining a set of operation areas that each agricultural robot can cover; S35. Based on the set of operation areas, extract the subgraph area overlapping with the position of each agricultural robot in the operation task graph, construct the operation unit nodes and their associated edges contained in the subgraph into a local operation subgraph, and complete the operation attribute binding; S36. Establish a one-to-one mapping relationship between each local operation subgraph and the corresponding agricultural robot, which serves as an input basis for operation task division and path planning.

6. The agricultural robot control method based on multi-agent collaboration according to claim 5 is characterized in that: The multi-round graph information transmission mechanism refers to iteratively aggregating and updating the adjacent node features of each operation unit in the graph neural network, so that the graph embedding vector of each operation unit integrates local structural relationships and attribute information.

7. The agricultural robot control method based on multi-agent collaboration according to claim 1 is characterized in that: The S4 specifically includes: S41. After the agricultural robot completes the construction of the local operation subgraph, it starts the point-to-point communication mechanism. Each agricultural robot sends the node set and edge set of the local operation subgraph held by it to other agricultural robots within its communication radius through short-range wireless communication. S42: After receiving the local operation subgraph data sent by the agricultural robot, perform graph structure splicing operations in sequence, perform position verification and boundary fusion on the operation unit nodes shared by adjacent agricultural robots, eliminate duplicate nodes and edge information, and construct an expanded joint operation subgraph; S43. After each agricultural robot completes the bidirectional sharing and graph structure splicing of the local operation subgraph, it sequentially uploads the aggregated data to the graph neural network model. The graph neural network model performs a feature fusion operation on the joint operation subgraph, combines the structural features and attribute features of the nodes in each subgraph, and uniformly encodes the node representation; S44. Based on the joint operation subgraph shared and integrated by all agricultural robots, a global collaborative operation graph covering the entire farmland area is constructed, wherein the global collaborative operation graph includes all operation unit nodes, complete edge connection relationships, and operation unit attribute information; S45. Synchronously return the collaborative operation map to the local cache of each agricultural robot.

8. The agricultural robot control method based on multi-agent collaboration according to claim 1 is characterized in that: The S5 specifically includes: S51. Based on the spatial position, task type, and task complexity of each operation unit in the global collaborative operation graph, extract the task feature information of all nodes in the graph, and simultaneously collect the current position, motion capability, and task carrying capacity information of all agricultural robots; S52. Based on the spatial relationship between the work unit and the agricultural robot, the degree of task adaptation, and the adjacency structure in the graph, perform a multi-agent collaborative task division operation, and use a region aggregation method to divide the work unit into multiple non-overlapping subtask areas, each of which corresponds to an agricultural robot; S53. Construct a task allocation matrix based on the divided subtask areas, and identify the operation unit number, operation sequence, and task content corresponding to each agricultural robot in the task allocation matrix to ensure that the operation allocation has coverage, balance, and execution accessibility; S54: Structurally encapsulate the instruction content in the task allocation matrix to generate an operation instruction package for the agricultural robot, where each operation instruction package includes a target operation unit number, an operation type identifier, a priority level, and completion condition parameters; S55. Send the operation instruction package to the corresponding agricultural robot terminal through short-distance communication, and record the instruction sending status in the operation scheduling module to ensure that each agricultural robot obtains the operation task that matches its capability.

9. The agricultural robot control method based on multi-agent collaboration according to claim 8, characterized in that: The task allocation matrix refers to a mapping relationship table that allocates work units to specific agricultural robots based on the division of multi-agent collaborative tasks, according to the spatial proximity between the agricultural robot and the work unit, the matching degree of work capabilities and the task load balancing principle.

10. The agricultural robot control method based on multi-agent collaboration according to claim 1, characterized in that: The S6 specifically includes: S61. After receiving the issued operation instruction, the agricultural robot analyzes the spatial position, task type, and execution sequence parameters of the target operation unit and obtains the current position information in real time. S62: Load obstacle location data corresponding to the work area, input the current location information and the spatial location of the target work unit into the path calculation module, and initialize the work path planning process; S63, constructing a grid connectivity graph of the farmland operation space based on the operation path planning process, constructing a set of traversable paths based on the obstacle distribution, extracting the optimal connectivity path, and generating an operation path sequence; S64: During the operation path sequence generation process, a path node screening operation is performed based on the obstacle boundary information, path nodes that overlap with or are too close to obstacles are removed, and the path direction is structurally adjusted; S65. The agricultural robot performs the mobile operation process according to the optimized operation path sequence, and records the current position, task completion status, and execution duration data after completing each operation unit; S66. Upload the recorded job status data to the job scheduling module through the communication interface to complete the real-time synchronization and update of the job status.

Images