Multi-robot field cooperative scheduling method and system based on machine learning

Abstract

The present application relates to the technical field of machine learning, in particular to a multi-robot field cooperative scheduling method and system based on machine learning, comprising: acquiring remote sensing images and radar images of a field area, and dividing the field area into a plurality of unit grids; acquiring the ground feature complexity of each unit grid based on the remote sensing images and radar images of each unit grid; acquiring a plurality of work areas according to the ground feature complexity; acquiring the complexity and community label value of each work area; determining the necessity of setting an overlapping area of two adjacent work areas by using the complexity and community label value of the two adjacent work areas; acquiring the overlapping area of the two adjacent work areas based on the necessity of setting the overlapping area; acquiring the final work area of each robot according to the overlapping area; and planning a coverage work path for each robot based on the final work area of the robot. The present application can improve the work quality and reliability of field robots.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a multi-robot field collaborative scheduling method and system based on machine learning. Background Technology

[0002] In the face of large-scale and complex modern farmland, a single robot can hardly complete high-quality operations efficiently and reliably within the limited agricultural time window, so collaborative scheduling of field operations is particularly important.

[0003] Existing methods for collaborative scheduling in field operations typically prioritize avoiding overlapping work areas. Through path planning and task allocation algorithms, they strive to strictly separate the work areas of each robot to maximize efficiency and minimize resource consumption. However, in real-world field environments, complex factors such as undulating terrain, uneven crop distribution, and obstacle occlusion often make it difficult to achieve completely uniform and comprehensive coverage in a single operation. Existing methods treat "overlap" as a purely negative factor and attempt to eliminate it, failing to recognize that under complex conditions, moderate and controlled area overlap can serve as an effective quality assurance mechanism. Consequently, while existing collaborative scheduling methods avoid resource waste, they lose the opportunity to utilize overlapping operations for process verification and quality review. This may result in blind spots or quality defects in local operations not being detected and corrected in a timely manner, affecting the overall reliability and consistency of the operation. Summary of the Invention

[0004] This invention provides a multi-robot field collaborative scheduling method and system based on machine learning to solve existing problems.

[0005] The multi-robot field collaborative scheduling method and system based on machine learning of the present invention adopts the following technical solution:

[0006] One embodiment of the present invention provides a multi-robot field collaborative scheduling method based on machine learning, the method comprising the following steps:

[0007] Acquire remote sensing images and radar images of the field area, and divide the field area into several unit grids;

[0008] The ground feature complexity of each cell grid is obtained based on remote sensing images and radar imagery of each cell grid.

[0009] Based on the terrain complexity of each unit grid, all unit grids are merged to obtain multiple operational areas;

[0010] Obtain the complexity and community tag value for each job region;

[0011] By utilizing the complexity and community tag values ​​of two adjacent work areas, the necessity of setting up overlapping areas between two adjacent work areas can be determined.

[0012] Based on the necessity of setting the overlapping area of ​​two adjacent work areas, obtain the overlapping area of ​​two adjacent work areas;

[0013] The final working area of ​​each robot is obtained based on the overlapping area of ​​two adjacent working areas;

[0014] Based on the final working area of ​​each robot, a coverage path is planned for it.

[0015] Furthermore, the specific steps for obtaining the ground feature complexity of each unit grid based on the remote sensing image and radar imagery of each unit grid are as follows:

[0016] A target detection network is used to identify crops in the remote sensing image of each cell grid, obtain the number of crops in each cell grid, and divide the number of crops by the area of ​​each cell grid to obtain the crop density of each cell grid.

[0017] Elevation maps of each cell grid are obtained from radar imagery of each cell grid.

[0018] Obtain the 3D normal vector of each pixel in each cell grid based on the elevation map of each cell grid;

[0019] Calculate the mean of all 3D normal vectors as the 3D normal vector for each element mesh;

[0020] The land feature complexity of each cell is determined by the tuple consisting of the crop density and the three-dimensional normal vector of each cell.

[0021] Furthermore, the specific steps for merging all cell grids based on the terrain complexity of each cell grid to obtain multiple operational areas are as follows:

[0022] Using each cell grid as a node, the land feature complexity of each cell grid as the node value, and the lines connecting corresponding nodes of adjacent cell grids as edges, a minimum area area map structure is constructed in the field.

[0023] The Laplacian graph clustering algorithm was used to cluster the nodes in the field minimum area region map structure to obtain multiple initial node clusters.

[0024] Get the number of nodes in each initial node cluster, and multiply the number of nodes by the area of ​​each node to get the area of ​​each initial node cluster;

[0025] If the area of ​​each initial node cluster is less than or equal to the preset area threshold, the cell grids contained in each initial node cluster are merged to obtain the working area corresponding to each node cluster.

[0026] If the area of ​​an initial node cluster is greater than the preset area threshold, then the subgraph corresponding to that cluster is extracted from the field minimum area region map structure, and the Laplacian graph clustering algorithm is re-executed on the subgraph to divide it into two or more sub-clusters;

[0027] Repeat the area verification and segmentation steps above until the area of ​​all clusters is less than or equal to the preset area threshold. Then merge the cell grids contained in each cluster to obtain the working area corresponding to each cluster.

[0028] Furthermore, the specific steps for obtaining the complexity and community tag value of each job area are as follows:

[0029] The average feature complexity of the unit grid corresponding to each operation area is used to determine the complexity of each operation area.

[0030] Based on the complexity of each work area, the DBSCAN density clustering algorithm is used to cluster all work areas, obtain multiple work area categories, and assign the same symbol to the work areas in each category;

[0031] A work area graph structure is constructed by using each work area as a node, the symbol of each work area as the node value, and the lines connecting adjacent work areas as edges.

[0032] Input the job area graph structure into the SLPA community discovery algorithm to obtain the community label value for each job area.

[0033] Furthermore, the specific steps for determining the necessity of setting up overlapping areas between two adjacent work areas by utilizing the complexity and community tag values ​​of two adjacent work areas are as follows:

[0034] Based on the community tag values ​​of two adjacent work areas, obtain the community relationship values ​​of the two adjacent work areas;

[0035] Based on the complexity of two adjacent work areas, it is necessary to obtain the overlapping area of ​​two adjacent work areas.

[0036] The necessity of the overlapping area between two adjacent work areas is used as the first weight, and the difference between 1 and the first weight is used as the second weight.

[0037] Multiply the second weight by the community relationship value of the two adjacent work areas to obtain the first product;

[0038] The second product is obtained by multiplying the first weight by the necessity of the overlapping area between two adjacent work areas.

[0039] The sum of the first and second products is used to determine the necessity of setting the overlapping area of ​​two adjacent work areas.

[0040] Furthermore, the specific steps for determining the necessity of obtaining the overlapping area of ​​two adjacent work areas based on their complexity are as follows:

[0041] Calculate the average complexity of two adjacent work areas and the absolute value of the difference in complexity between the two adjacent work areas. Multiply the average complexity by the absolute value of the difference to calculate the necessity of the overlapping area between the two adjacent work areas.

[0042] Furthermore, the specific steps for obtaining the overlapping area of ​​two adjacent work areas based on the necessity of setting the overlapping area of ​​two adjacent work areas are as follows:

[0043] Obtain the common boundary length and overlapping area width of two adjacent work areas;

[0044] The overlapping area of ​​two adjacent work areas is determined by multiplying the necessity of the overlapping area setting, the length of the common boundary, and the width of the overlapping area.

[0045] Furthermore, the specific steps for obtaining the final working area of ​​each robot based on the overlapping area of ​​two adjacent working areas are as follows:

[0046] The overlapping area of ​​two adjacent work areas is used to obtain the merged overlapping area;

[0047] Assign a unique robot to each work area;

[0048] For any merged overlapping area, identify all related work areas, and select the robot with the smallest original work area from the robots responsible for these areas, and assign it as the quality inspection robot for that merged overlapping area.

[0049] Merging overlapping areas will be included as an additional task in the final working area of ​​the quality inspection robot.

[0050] Furthermore, the specific steps for obtaining the merged overlapping area based on the overlapping area of ​​two adjacent work areas are as follows:

[0051] The effective length of the overlapping area is determined based on the overlapping area of ​​two adjacent work areas;

[0052] The effective length is taken from the common boundary length of two adjacent work areas as the long side of the overlapping area, and the width of the overlapping area is extended vertically into the work area with higher complexity as the short side of the overlapping area. The rectangular area defined by the long side and the short side is determined as the overlapping area of ​​two adjacent work areas.

[0053] For each work area, all overlapping areas that are adjacent and belong to the same community are merged to obtain the merged overlapping area.

[0054] One embodiment of the present invention provides a multi-robot field collaborative scheduling system based on machine learning, the system comprising the following modules:

[0055] The acquisition module is used to acquire remote sensing images and radar images of the field area and divide the field area into several unit grids;

[0056] The analysis module is used to obtain the ground feature complexity of each cell grid based on the remote sensing images and radar images of each cell grid.

[0057] Based on the terrain complexity of each unit grid, all unit grids are merged to obtain multiple operational areas;

[0058] Obtain the complexity and community tag value for each job region;

[0059] By utilizing the complexity and community tag values ​​of two adjacent work areas, the necessity of setting up overlapping areas between two adjacent work areas can be determined.

[0060] Based on the necessity of setting the overlapping area of ​​two adjacent work areas, obtain the overlapping area of ​​two adjacent work areas;

[0061] The final working area of ​​each robot is obtained based on the overlapping area of ​​two adjacent working areas;

[0062] The planning module is used to plan a coverage path for each robot based on its final work area.

[0063] The beneficial effects of the technical solution of the present invention are as follows: The embodiments of the present invention propose a multi-robot field collaborative scheduling method and system based on machine learning. By combining the complexity of the field in the collaborative scheduling, some adjacent areas are covered in a planned and limited manner. This not only allows the repeated areas to be used as a reference benchmark for cross-verification, but also allows for timely detection of possible omissions or anomalies through comparative analysis, thereby improving the overall quality and reliability of the operation. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart illustrating the steps of the multi-robot field collaborative scheduling method based on machine learning in this invention.

[0066] Figure 2 This is a block diagram of the multi-robot field collaborative scheduling system based on machine learning, as described in this invention. Detailed Implementation

[0067] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the multi-robot field collaborative scheduling method and system based on machine learning proposed in this invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0069] The following description, in conjunction with the accompanying drawings, details the specific scheme of the multi-robot field collaborative scheduling method and system based on machine learning provided by this invention.

[0070] Please see Figure 1 The diagram illustrates a flowchart of a machine learning-based multi-robot field collaborative scheduling method according to an embodiment of the present invention. The method includes the following steps:

[0071] Step S001: Acquire remote sensing images and radar images of the field area, and divide the field area into several unit grids.

[0072] It should be noted that the UAV system, equipped with multispectral or hyperspectral imaging sensors and radar, performs low-altitude flight missions over target farmland areas according to a preset automated flight path. During the flight, it simultaneously acquires high-resolution remote sensing image data of the field and raw radar data. The entire acquisition process integrates a high-precision GNSS positioning module and a flight control system to ensure that every frame of image and radar data has accurate geographical location information.

[0073] Multiple remote sensing images were stitched and fused to generate an orthophoto map covering the entire field area, which served as the remote sensing image of the field area.

[0074] The acquired raw radar data is processed by interferometric synthetic aperture radar to generate a digital elevation model, in which the elevation value corresponding to each pixel constitutes the radar image.

[0075] To address the problems of existing methods, this embodiment considers actual field conditions. Due to factors such as undulating terrain, uneven crop distribution, and obstructions, a single robot may experience incomplete coverage and inconsistent quality in localized areas. Relying solely on a single operation often fails to identify and correct these potential defects. Therefore, in collaborative scheduling, based on the complexity of the field, planned and limited repeated coverage of some adjacent areas is performed. This not only utilizes the repeated areas as a reference benchmark for cross-verification but also allows for timely detection of potential omissions or anomalies through comparative analysis, thereby improving the overall quality and reliability of the operation.

[0076] Based on this, this embodiment first divides the field area into zones, then determines which zones need to overlap, and then plans the robot's travel route for each zone.

[0077] Using the geographic coordinate system shared by the georegistered remote sensing and radar images of the field area as the spatial reference, the entire field area is uniformly divided into the smallest area unit grid (e.g., corresponding to 5m*5m in reality).

[0078] Step S002: Obtain the ground feature complexity of each cell grid based on the remote sensing image and radar imagery of each cell grid.

[0079] Step S002 further includes steps S0021-S0025:

[0080] Step S0021: Use an object detection network to identify crops in the remote sensing image of each cell grid, obtain the number of crops in each cell grid, and divide the number of crops by the area of ​​each cell grid to obtain the crop density of each cell grid.

[0081] It should be noted that the target detection network is a YOLO network. For each cell grid, crop identification is performed using the YOLO network to obtain the number of crops n in that cell grid. Dividing this number by the area of ​​the cell grid yields the crop density m1 for that cell grid. YOLO networks and YOLO network target identification are well-known technologies and will not be described in detail here.

[0082] Step S0022: Obtain the elevation map of each cell grid based on the radar imagery of each cell grid.

[0083] It should be noted that for each grid cell, all elevation data points falling within the geographical boundary of that grid cell are extracted from the digital elevation model to form a height information matrix representing the topographic relief of that grid area, which is denoted as the elevation map of that grid cell.

[0084] Step S0023: Obtain the three-dimensional normal vector of each pixel in each cell grid based on the elevation map of each cell grid.

[0085] It should be noted that: the gradient histogram of the unit grid elevation map is calculated, and the three-dimensional normal vector corresponding to each pixel in the elevation map is obtained through INSAR technology. The first component of the three-dimensional normal vector represents the east-west tilt information, the second component represents the north-south tilt information, and the third component describes the verticality or horizontality of the ground surface.

[0086] Step S0024: Calculate the mean of all three-dimensional normal vectors as the three-dimensional normal vector for each element mesh.

[0087] It should be noted that each cell grid contains many pixels, and each pixel's 3D normal vector has three components. The arithmetic mean of the first (east-west), second (north-south), and third (vertical) components of the normal vectors of all pixels within that cell is calculated independently. These three means are then combined to form a new 3D vector, which serves as the 3D normal vector for each cell grid. This 3D normal vector represents the slope of that cell grid.

[0088] Step S0025: Determine the land feature complexity of each unit grid by the tuple formed by the crop density and the three-dimensional normal vector of each unit grid.

[0089] It should be noted that for each cell grid, the land feature complexity of the grid area is represented by a tuple formed by crop density and three-dimensional normal vector. For example, [m1, (s1, s2, s3)] represents the tuple, where s1, s2, and s3 represent the first component, the second component, and the third component, respectively.

[0090] Step S003: Merge all cell grids according to the terrain complexity of each cell grid to obtain multiple work areas.

[0091] Step S003 further includes steps S0031-S0036:

[0092] Step S0031: Using each unit grid as a node, the land feature complexity of each unit grid as the node value, and the connection between corresponding nodes of adjacent unit grids as edges, construct the minimum area region map structure in the field.

[0093] Step S0032: Use the Laplacian graph clustering algorithm to cluster the nodes in the field minimum area region graph structure to obtain multiple initial node clusters.

[0094] It's important to note that Laplacian graph clustering is a graph-based clustering method. It groups tightly connected nodes with similar features into clusters by calculating the eigenvectors of the Laplacian matrix of a graph, while simultaneously making the connections between different clusters as sparse as possible. The algorithm first constructs a Laplacian matrix based on the graph's connectivity (edges) and node attributes (feature complexity tuples). By solving for the eigenvectors of this matrix, the high-dimensional graph node data is mapped to a low-dimensional feature space. In this space, nodes that are initially tightly connected and have similar feature complexity values ​​will cluster together, and then be further divided into multiple initial node clusters using subsequent clustering methods (such as K-means). Laplacian graph clustering is a well-known technique and will not be elaborated upon further here.

[0095] The tuples of the cell grids corresponding to the same initial node cluster are similar, that is, the terrain complexity is similar. Multiple cell grids corresponding to the same initial node cluster are connected, which allows a robot to perform operations such as harvesting and sowing in the area corresponding to the initial node cluster.

[0096] Step S0033: Obtain the number of nodes in each initial node cluster, and multiply the number of nodes by the area of ​​each node to obtain the area of ​​each initial node cluster.

[0097] It should be noted that each initial node cluster contains several cell grids. To avoid the actual area of ​​a certain initial node cluster being too large, the area of ​​each cell grid is calculated by multiplying the area of ​​each cell grid in the cluster by the number of cell grids in the cluster.

[0098] Step S0034: If the area of ​​each initial node cluster is less than or equal to the preset area threshold, merge the cell grids contained in each initial node cluster to obtain the working area corresponding to each node cluster.

[0099] It should be noted that the preset area threshold Y refers to the working capacity of a single robot, and this value can be determined based on the actual battery life and working efficiency of the robot.

[0100] Step S0035: If there is an initial node cluster area greater than the preset area threshold, extract the subgraph corresponding to the cluster from the field minimum area region map structure, and re-execute the Laplacian graph clustering algorithm on the subgraph to divide it into two or more sub-clusters.

[0101] Step S0036: Repeat the above area verification and segmentation steps until the area of ​​all clusters is less than or equal to the preset area threshold. Then merge the cell grids contained in each cluster to obtain the working area corresponding to each cluster.

[0102] It should be noted that: grid regions belonging to the same cluster are merged into one region, and then one robot is responsible for agricultural operations such as harvesting in that region. Grid regions corresponding to the same cluster are recorded as work areas, and each cluster corresponds to one work area, resulting in multiple work areas. Each work area may be irregular in shape.

[0103] Step S004: Obtain the complexity and community tag value for each job region.

[0104] Step S004 further includes steps S0041-S0044:

[0105] Step S0041: Determine the complexity of each work area by taking the average value of the feature complexity of the unit grid corresponding to each work area.

[0106] It should be noted that for each task area, the mean of each element in the tuples of all grid areas within that area is calculated. The tuples formed by these mean values ​​represent the complexity of that task area. Higher complexity means greater task difficulty and a greater likelihood of omissions or poor task quality.

[0107] Step S0042: Based on the complexity of each work area, the DBSCAN density clustering algorithm is used to cluster all work areas, obtain multiple work area categories, and assign the same symbol to the work areas in each category.

[0108] It should be noted that: then it is necessary to identify which work areas in the field should be treated as a whole or a closely related group in terms of scheduling, so that the robots responsible for these work areas can check the work status of adjacent areas in overlapping areas.

[0109] DBSCAN is a clustering algorithm based on the spatial distribution density of data points. It can automatically discover clusters of arbitrary shapes and identify points in low-density areas as noise, without needing to pre-specify the number of clusters. First, the complexity of each work area is considered as a feature point in two-dimensional space. Based on a preset neighborhood radius and minimum sample size, the algorithm starts from any unvisited work area point and searches for all points within its density reach, forming a cluster. If the number of points in the neighborhood of a point is less than the minimum sample size, it is temporarily marked as a boundary point or noise. This process is iterated until all work area points have been visited. Ultimately, regions with similar complexity features and connected spatial densities are clustered into the same category (e.g., high complexity, medium complexity, low complexity), while isolated regions may be identified as noise and not classified. The DBSCAN density clustering algorithm is a well-known technique and will not be elaborated upon here.

[0110] Different work area nodes within the same work area category are represented by the same symbol, such as A, B, C, etc.

[0111] Step S0043: Construct a work area graph structure with each work area as a node, the symbol of each work area as the node value, and the lines connecting adjacent work areas as edges.

[0112] Step S0044: Input the job area graph structure into the SLPA community discovery algorithm to obtain the community label value of each job area.

[0113] It's important to note that SLPA is a community discovery algorithm based on dynamic label propagation. It identifies tightly connected community structures in a network by iteratively propagating labels among nodes in a graph, allowing nodes to belong to multiple communities simultaneously. First, a unique label is initialized for each node (i.e., the work area) in the work area graph structure. The algorithm then iterates through multiple rounds: SLPA discovers circles by having nodes "chat" with each other. Each node carries a notebook (memory list). In repeated communication, neighboring nodes share their "identity labels," which are then recorded in the notebook. After many rounds of communication, each node's notebook contains various labels and their frequency. Finally, each node retains only the labels that appear frequently enough; these labels represent the circle to which it belongs. Because a node can retain multiple labels, it can belong to multiple circles simultaneously; that is, each node's final label value may be one or more.

[0114] If a node eventually obtains multiple label values, the label value with the highest frequency among the nodes directly connected to that node is selected as the final label value for that node, i.e., the community label value for each work area.

[0115] Step S005: Determine the necessity of setting up overlapping areas between two adjacent work areas by using the complexity and community tag values ​​of the two adjacent work areas.

[0116] Step S005 further includes steps S0051-S0056:

[0117] Step S0051: Based on the community tag values ​​of two adjacent work areas, obtain the community relationship values ​​of the two adjacent work areas.

[0118] It should be noted that nodes that are adjacent in the task area graph structure and have the same label value are regarded as nodes of the same community. Areas with the same label mean that they have a higher necessity for collaborative work in terms of task complexity and spatial association.

[0119] Therefore, nodes in the same community are represented as a whole or a closely related group in the scheduling, providing a basis for community relationships for subsequent overlapping decisions.

[0120] At this point, each node has three attributes: its own node number (representing the uniqueness of the node), the complexity of the work region corresponding to the node, and the community tag value.

[0121] For each group of spatially adjacent work areas i and j, f(Li, Lj) is used as the label function. Li and Lj represent the community label values ​​of the two work areas. If Li = Lj, that is, the two work areas belong to the same community, then f(Li, Lj) = 1, that is, the community relationship value of the two adjacent work areas is 1, indicating that the intra-group verification of the two work areas should be strengthened.

[0122] Otherwise, f(Li, Lj) = r, that is, the community relationship value between two adjacent work areas is r. r is a low value, which can be set according to the actual situation in the field. This value is a small value in the range of 0-1. In this embodiment, it is preferably 0.2.

[0123] Step S0052: Based on the complexity of two adjacent work areas, determine the necessity of the overlapping area between the two adjacent work areas.

[0124] Specifically, it includes:

[0125] Calculate the average complexity of two adjacent work areas and the absolute value of the difference in complexity between the two adjacent work areas. Multiply the average complexity by the absolute value of the difference to calculate the necessity of the overlapping area between the two adjacent work areas.

[0126] It should be noted that for each pair of spatially adjacent regions i and j, g(hi, hj) is used as the complexity function, and hi and hj represent the complexity of the two working regions respectively.

[0127] The higher the complexity of these two task areas, the more necessary it is to set up an overlapping area for checking; the greater the difference in complexity between the two task areas, the more necessary it is to set up an overlapping area to verify the more complex area. This leads to the following necessity for the overlapping area between the two task areas based on complexity:

[0128]

[0129] in, The overlapping region of two adjacent work regions i and j is necessary, and hi and hj represent the complexity of the two work regions respectively.

[0130] Step S0053: The necessity of the overlapping area between two adjacent work areas is taken as the first weight, and the difference between 1 and the first weight is taken as the second weight.

[0131] Step S0054: Multiply the second weight by the community relationship value of the two adjacent work areas to obtain the first product.

[0132] Step S0055: Multiply the first weight by the necessity of the overlapping area of ​​two adjacent work areas to obtain the second product.

[0133] Step S0056: Determine the necessity of setting the overlapping area of ​​two adjacent work areas by the sum of the first product and the second product.

[0134] It should be noted that both community relationships and regional complexity influence whether overlapping regions need to be set. This embodiment uses a weighted summation method to obtain the final necessity for setting overlapping regions. The greater the regional complexity, the more necessary it is to set overlapping regions, thus using regional complexity itself as its weight. The necessity for setting overlapping regions between regions i and j is then obtained using the following formula:

[0135]

[0136] in, This indicates the necessity of setting up an overlapping area between two adjacent work areas i and j. Indicates the first weight. = ; =1- , Indicates the second weight; This represents the community relationship value between two adjacent work areas i and j.

[0137] actual and The value can also be determined based on the number of neighboring nodes of each node in the obtained work area map structure, and the overall complexity of the area (the more neighboring nodes each node has, the greater the overall complexity of the field area). (Take the larger value) to obtain.

[0138] Step S006: Based on the necessity of setting the overlapping area of ​​two adjacent work areas, obtain the overlapping area of ​​two adjacent work areas.

[0139] It should be noted that the necessity of setting overlapping areas is not used as an indicator to select whether to set overlapping areas, but rather as a basis for setting the size of the overlapping area. The greater the necessity of setting overlapping areas, the larger the overlapping area needs to be set so that robots in adjacent areas can perform quality inspection on the overlapping parts of the area.

[0140] Specifically, it includes:

[0141] Obtain the common boundary length and overlapping area width of two adjacent work areas.

[0142] The overlapping area of ​​two adjacent work areas is determined by multiplying the necessity of the overlapping area setting, the length of the common boundary, and the width of the overlapping area.

[0143] It should be noted that for each pair of spatially adjacent work areas i and j, the greater the necessity of setting an overlapping area, the larger the overlap area of ​​the two areas should be; the larger the length c of the common boundary, the larger the overlap area should be; combined with the pre-set width d of the overlapping area (set according to specific circumstances), the overlap area is obtained as follows:

[0144]

[0145] in, This represents the overlapping area of ​​two adjacent work areas i and j. For two adjacent work areas, this overlapping area is set in the work area with greater complexity, and the robot in the other work area performs quality inspection on this overlapping area.

[0146] Step S007: Obtain the final working area of ​​each robot based on the overlapping area of ​​two adjacent working areas.

[0147] Step S007 further includes steps S0071-S0074:

[0148] Step S0071: Obtain the merged overlapping area based on the overlapping area of ​​two adjacent work areas.

[0149] Specifically, it includes:

[0150] The effective length of the overlapping area is determined based on the overlapping area of ​​two adjacent work areas.

[0151] It should be noted that: As the effective length of the overlapping region.

[0152] The effective length is taken from the common boundary length of two adjacent work areas as the long side of the overlapping area, and the width of the overlapping area is extended vertically into the work area with higher complexity as the short side of the overlapping area. The rectangular area defined by the long side and the short side is determined as the overlapping area of ​​the two adjacent work areas.

[0153] It should be noted that: The length of the overlapping region is d, and the width of the overlapping region is d, thus obtaining the rectangular overlapping region; and this overlapping region is set from the more complex work region between two adjacent work regions.

[0154] For each work area, all overlapping areas that are adjacent and belong to the same community are merged to obtain the merged overlapping area.

[0155] It should be noted that for each work area, there are multiple adjacent areas, resulting in multiple overlapping areas. These may consist of small, discrete overlapping blocks, which is detrimental to efficient robot operation. Therefore, overlapping areas belonging to the same community are spatially connected and merged to form one or more larger overlapping areas, referred to as merged overlapping areas. This helps reduce the number of overlapping blocks and the robot's starting, stopping, and turning, creating smoother work boundaries.

[0156] Step S0072: Assign a unique robot to each work area.

[0157] Step S0073: For any merged overlapping area, determine all related work areas, and select the robot with the smallest original work area from the robots responsible for these areas, and assign it as the quality inspection robot for that merged overlapping area.

[0158] Step S0074: Merge the overlapping areas as an additional task and incorporate them into the final working area of ​​the quality inspection robot.

[0159] It should be noted that: for merging overlapping areas, all related work areas can be obtained. Then, each work area will be assigned an agricultural robot to carry out the work. The robot with the smallest work area area will be selected as the robot for merging overlapping areas. That is, after completing the work in its own work area, the robot needs to go to the merged overlapping area to take corresponding photos and complete the collection of quality inspection data for the merged overlapping area.

[0160] Step S008: Based on the final working area of ​​each robot, plan a coverage working path for it.

[0161] It should be noted that at this point, the final working area of ​​each robot can be obtained. If the final working area includes a merged overlapping area, the merged overlapping area is also taken as the final working area of ​​that robot.

[0162] For each robot's final work area, a genetic algorithm is used to plan the optimal coverage path for each individual robot. Different weights can be assigned to overlapping areas in the fitness function of the genetic algorithm. For example, overlapping areas can be used at a slower speed for quality checks, or designated as critical sections that must be precisely covered. This allows for efficient, low-repetition coverage paths within the area, ultimately yielding the optimal work path for each robot within its assigned area.

[0163] The robot operates along a planned path. When it passes through overlapping areas, it executes a preset high-quality operation mode (such as slowing down and improving operational precision). After completing the operation, it immediately takes photos of the area's operational results (such as photos of the soil after sowing and the ridges after harvesting).

[0164] The system then performs rapid AI analysis on the returned overlapping area photos to determine the quality of the work and reacts immediately:

[0165] If it is a localized problem (such as a small missing video): command the current robot to return to repair.

[0166] If it is a systemic problem (such as the overall sowing being too shallow): automatically adjust the operating parameters of all robots (such as increasing the sowing depth).

[0167] In case of a serious malfunction (such as machine damage): immediately call the alarm and notify personnel for handling.

[0168] The "job performance data" collected from the overlapping areas is then fed back to the central scheduling system. The system will learn from this experience and be able to allocate tasks more intelligently, set overlapping areas more rationally, and continuously optimize the overall job plan in the future.

[0169] Please see Figure 2 The diagram illustrates a block diagram of a machine learning-based multi-robot field collaborative scheduling system according to an embodiment of the present invention. The system includes the following modules:

[0170] The acquisition module 100 is used to acquire remote sensing images and radar images of the field area and divide the field area into several unit grids;

[0171] Analysis module 200 is used to obtain the ground feature complexity of each cell grid based on remote sensing images and radar images of each cell grid;

[0172] Based on the terrain complexity of each unit grid, all unit grids are merged to obtain multiple operational areas;

[0173] Obtain the complexity and community tag value for each job region;

[0174] By utilizing the complexity and community tag values ​​of two adjacent work areas, the necessity of setting up overlapping areas between two adjacent work areas can be determined.

[0175] Based on the necessity of setting the overlapping area of ​​two adjacent work areas, obtain the overlapping area of ​​two adjacent work areas;

[0176] The final working area of ​​each robot is obtained based on the overlapping area of ​​two adjacent working areas;

[0177] The planning module 300 is used to plan a coverage path for each robot based on its final work area.

[0178] In summary, in this embodiment of the invention, by combining the complexity of the site with planned and limited repeated coverage of some adjacent areas during collaborative scheduling, not only can the repeated areas be used as a reference benchmark for cross-verification, but also possible omissions or anomalies can be discovered in a timely manner through comparative analysis, thereby improving the overall quality and reliability of the operation.

[0179] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-robot field collaborative scheduling method based on machine learning, characterized in that, The method includes the following steps: Acquire remote sensing images and radar images of the field area, and divide the field area into several unit grids; The ground feature complexity of each cell grid is obtained based on remote sensing images and radar imagery of each cell grid. Based on the terrain complexity of each unit grid, all unit grids are merged to obtain multiple operational areas; Obtain the complexity and community tag value for each job region; By utilizing the complexity and community tag values ​​of two adjacent work areas, the necessity of setting up overlapping areas between two adjacent work areas can be determined. Based on the necessity of setting the overlapping area of ​​two adjacent work areas, obtain the overlapping area of ​​two adjacent work areas; The final working area of ​​each robot is obtained based on the overlapping area of ​​two adjacent working areas; Based on the final working area of ​​each robot, a coverage path is planned for it.

2. The multi-robot field collaborative scheduling method based on machine learning according to claim 1, characterized in that, The specific steps for obtaining the ground feature complexity of each unit grid based on remote sensing images and radar imagery are as follows: A target detection network is used to identify crops in the remote sensing image of each cell grid, obtain the number of crops in each cell grid, and divide the number of crops by the area of ​​each cell grid to obtain the crop density of each cell grid. Elevation maps of each cell grid are obtained from radar imagery of each cell grid. Obtain the 3D normal vector of each pixel in each cell grid based on the elevation map of each cell grid; Calculate the mean of all 3D normal vectors as the 3D normal vector for each element mesh; The land feature complexity of each cell is determined by the tuple consisting of the crop density and the three-dimensional normal vector of each cell.

3. The multi-robot field collaborative scheduling method based on machine learning according to claim 1, characterized in that, The specific steps involved in merging all cell grids based on the terrain complexity of each cell grid to obtain multiple operational areas are as follows: Using each cell grid as a node, the land feature complexity of each cell grid as the node value, and the lines connecting corresponding nodes of adjacent cell grids as edges, a minimum area area map structure is constructed in the field. The Laplacian graph clustering algorithm was used to cluster the nodes in the field minimum area region map structure to obtain multiple initial node clusters. Get the number of nodes in each initial node cluster, and multiply the number of nodes by the area of ​​each node to get the area of ​​each initial node cluster; If the area of ​​each initial node cluster is less than or equal to the preset area threshold, the cell grids contained in each initial node cluster are merged to obtain the working area corresponding to each node cluster. If the area of ​​an initial node cluster is greater than the preset area threshold, then the subgraph corresponding to that cluster is extracted from the field minimum area region map structure, and the Laplacian graph clustering algorithm is re-executed on the subgraph to divide it into two or more sub-clusters; Repeat the area verification and segmentation steps above until the area of ​​all clusters is less than or equal to the preset area threshold. Then merge the cell grids contained in each cluster to obtain the working area corresponding to each cluster.

4. The multi-robot field collaborative scheduling method based on machine learning according to claim 1, characterized in that, The specific steps involved in obtaining the complexity and community tag value for each job region are as follows: The average feature complexity of the unit grid corresponding to each operation area is used to determine the complexity of each operation area. Based on the complexity of each work area, the DBSCAN density clustering algorithm is used to cluster all work areas, obtain multiple work area categories, and assign the same symbol to the work areas in each category; A work area graph structure is constructed by using each work area as a node, the symbol of each work area as the node value, and the lines connecting adjacent work areas as edges. Input the job area graph structure into the SLPA community discovery algorithm to obtain the community label value for each job area.

5. The multi-robot field collaborative scheduling method based on machine learning according to claim 1, characterized in that, The specific steps involved in determining the necessity of setting up overlapping areas between two adjacent work areas by utilizing the complexity and community tag values ​​of two adjacent work areas are as follows: Based on the community tag values ​​of two adjacent work areas, obtain the community relationship values ​​of the two adjacent work areas; Based on the complexity of two adjacent work areas, it is necessary to obtain the overlapping area of ​​two adjacent work areas. The necessity of the overlapping area between two adjacent work areas is used as the first weight, and the difference between 1 and the first weight is used as the second weight. Multiply the second weight by the community relationship value of the two adjacent work areas to obtain the first product; The second product is obtained by multiplying the first weight by the necessity of the overlapping area between two adjacent work areas. The sum of the first and second products is used to determine the necessity of setting the overlapping area of ​​two adjacent work areas.

6. The multi-robot field collaborative scheduling method based on machine learning according to claim 5, characterized in that, The specific steps involved in determining the necessity of obtaining the overlapping area between two adjacent work areas based on their complexity are as follows: Calculate the average complexity of two adjacent work areas and the absolute value of the difference in complexity between the two adjacent work areas. Multiply the average complexity by the absolute value of the difference to calculate the necessity of the overlapping area between the two adjacent work areas.

7. The multi-robot field collaborative scheduling method based on machine learning according to claim 1, characterized in that, The specific steps for determining the necessity of setting the overlapping area of ​​two adjacent work areas and obtaining the overlapping area of ​​two adjacent work areas are as follows: Obtain the common boundary length and overlapping area width of two adjacent work areas; The overlapping area of ​​two adjacent work areas is determined by multiplying the necessity of the overlapping area setting, the length of the common boundary, and the width of the overlapping area.

8. The multi-robot field collaborative scheduling method based on machine learning according to claim 1, characterized in that, The specific steps for obtaining the final working area of ​​each robot based on the overlapping area of ​​two adjacent working areas are as follows: The overlapping area of ​​two adjacent work areas is used to obtain the merged overlapping area; Assign a unique robot to each work area; For any merged overlapping area, identify all related work areas, and select the robot with the smallest original work area from the robots responsible for these areas, and assign it as the quality inspection robot for that merged overlapping area. Merging overlapping areas will be included as an additional task in the final working area of ​​the quality inspection robot.

9. The multi-robot field collaborative scheduling method based on machine learning according to claim 8, characterized in that, The specific steps for obtaining the merged overlapping area based on the overlapping area of ​​two adjacent work areas are as follows: The effective length of the overlapping area is determined based on the overlapping area of ​​two adjacent work areas; The effective length is taken from the common boundary length of two adjacent work areas as the long side of the overlapping area, and the width of the overlapping area is extended vertically into the work area with higher complexity as the short side of the overlapping area. The rectangular area defined by the long side and the short side is determined as the overlapping area of ​​two adjacent work areas. For each work area, all overlapping areas that are adjacent and belong to the same community are merged to obtain the merged overlapping area.

10. A multi-robot field collaborative scheduling system based on machine learning, characterized in that, The system includes the following modules: The acquisition module is used to acquire remote sensing images and radar images of the field area and divide the field area into several unit grids; The analysis module is used to obtain the ground feature complexity of each cell grid based on the remote sensing images and radar images of each cell grid. Based on the terrain complexity of each unit grid, all unit grids are merged to obtain multiple operational areas; Obtain the complexity and community tag value for each job region; By utilizing the complexity and community tag values ​​of two adjacent work areas, the necessity of setting up overlapping areas between two adjacent work areas can be determined. Based on the necessity of setting the overlapping area of ​​two adjacent work areas, obtain the overlapping area of ​​two adjacent work areas; The final working area of ​​each robot is obtained based on the overlapping area of ​​two adjacent working areas; The planning module is used to plan a coverage path for each robot based on its final work area.

Images