Agricultural machine multi-machine collaborative operation path planning method

Abstract

The invention belongs to the field of multi-machine collaborative operation path planning, and particularly relates to an agricultural machine multi-machine collaborative operation path planning method, which comprises the following steps: constructing environment perception; grid modeling is carried out; agricultural machine capability modeling; the local tasks are optimized and fragmented; agricultural machine operation paths are planned hierarchically; constructing an intention broadcast and carrying out conflict detection; remote feedback is carried out by setting online triggering conditions, and online learning is carried out according to the feedback; according to the invention, the unmanned agricultural machine can sense the environment around the operation in real time while operating according to the statically planned route, and dynamically adjust the operation path according to the feedback sensed in real time, thereby overcoming the defects of the traditional path planning, and feeding back the information to the central control system in real time, thereby adjusting the operation plan in time, and improving the working efficiency. And the problem of delayed response of a traditional method is avoided.

Description

Technical Field

[0001] This invention belongs to the field of multi-machine collaborative operation path planning, and specifically relates to a method for multi-machine collaborative operation path planning of agricultural machinery. Background Technology

[0002] Chinese patent application number CN202311157865.4 discloses a method, system, and device for full-process path planning in multi-machine collaborative operations. This invention constructs a farm environment map by acquiring and utilizing farm road information and boundary point information; plans inter-plot transfer paths and full-coverage paths within plots based on the farm environment map; and, based on the inter-plot transfer paths, full-coverage paths within plots, and basic information of the agricultural machinery, employs an improved ant colony algorithm based on the full-process operation path, aiming to minimize path cost, to achieve the optimal task sequence result for each agricultural machine, thereby realizing the allocation of operation tasks for multiple agricultural machines; and obtains the full-process path for multi-machine collaborative operations based on the operation task allocation result and the path set.

[0003] While this invention addresses the issue of unmanned agricultural machinery (UAV) paths being unsuitable for fully automated operations, static route planning makes it difficult for UAVs to avoid obstacles encountered in farmland. For example, during rice harvesting, encountering wetlands can be problematic because static planning fails to account for fields obscured by rice. If the machine follows the optimal static route, it may become stuck in the wetlands, leading to reduced efficiency and machine damage. Therefore, a multi-machine collaborative operation path planning method is urgently needed. This method would allow UAVs to simultaneously perceive their surroundings in real-time while operating along statically planned routes, dynamically adjusting their paths based on feedback. This would overcome the shortcomings of traditional path planning and provide real-time feedback to the central control system, enabling timely adjustments to the work plan and avoiding the delayed response issues of traditional methods. Summary of the Invention

[0004] This invention provides a method for multi-machine collaborative operation path planning in agricultural machinery, aiming to solve the following problems: Static route planning for unmanned agricultural machinery makes it difficult for the unmanned agricultural machinery to avoid obstacles that occur in farmland operations, such as encountering wetlands during rice harvesting, because static planning cannot take into account the situation of farmland being covered by rice. If the machine operates according to the optimal route planned by static planning, the agricultural machinery will get stuck in the wetlands, which will lead to reduced operating efficiency and damage to the machine.

[0005] The technical solution adopted by this invention to solve the above problems is: a method for planning the path of multi-machine collaborative operation of agricultural machinery, comprising: acquiring operational environment data by deploying multi-source sensors to construct environmental perception; performing grid-based modeling based on the acquired environmental data to generate a grid map of occupancy and attributes with confidence;

[0006] Agricultural machinery capabilities are modeled by establishing capability vectors for each machine; the work area is divided into several connected local tasks based on a grid map, and task-equipment matching and allocation are performed; candidate tasks are locally optimized by constructing an evaluation definition function.

[0007] Perform hierarchical path planning for agricultural machinery operations on assigned tasks, generating executable trajectories that satisfy vehicle dynamics constraints; construct intent broadcasting and perform conflict detection;

[0008] By setting online trigger conditions, rapid processing and remote feedback can be achieved at the edge, and key parameters can be adaptively updated using online learning.

[0009] As a preferred implementation, the specific steps for deploying multi-source sensors to acquire operational environment data and construct environmental perception are as follows: Operational environment data is acquired by deploying multi-source sensors within the agricultural facility park, resulting in visual observation data covering the entire operational area; wireless sensor network nodes are deployed within the agricultural facility park to collect soil moisture content, temperature, and humidity environmental parameters, obtaining fixed-point measurement data of the ground environment; positioning modules are installed on vehicles within the agricultural facility park, and positioning base stations are deployed to obtain unified spatial coordinates for vehicles within the park; and GPU acceleration units are configured on edge computing nodes and a network connection is established with the vehicle controller of the unmanned agricultural machinery.

[0010] As a preferred embodiment, the specific steps for generating a grid map of occupancy and attributes with confidence based on the obtained environmental data are as follows:

[0011] By dividing the agricultural facility park into a set of two-dimensional grids with fixed dimensions, a unified spatial discretization basis for the agricultural facility park is obtained. By constructing a linear weighting formula to recalculate the comprehensive weight of each grid in the facility agricultural park, a priority sequence that can be directly used for task partitioning and scheduling is obtained. The linear weighting formula is as follows:

[0012] ,

[0013] The overall weight of grid C, The growth state score weighting coefficients for the grid. The score for the crop growth status on grid C. These are the weighting coefficients of the obstacle probability function. Let C be the obstacle probability function. These are the weighting coefficients for the soil condition scoring function. Let be the soil condition scoring function on grid C, where C represents the grid into which the agricultural facility park is divided.

[0014] As a preferred embodiment, the specific steps for modeling agricultural machinery capabilities by establishing capability vectors for each machine are as follows:

[0015] Capability vectors are constructed for unmanned agricultural machinery participating in operations in agricultural facility parks. These capability vectors can be obtained by combining parameters provided by the manufacturer, calibration tests conducted on-site in the agricultural facility parks, and online operational data. The unmanned agricultural machinery's processing rate per unit time and peak speed can be obtained through calibration tests conducted on-site in the agricultural facility parks, while the remaining fuel level can be reported in real time by the on-board controller.

[0016] As a preferred embodiment, the specific steps of dividing the work area into several connected local tasks based on the grid map and performing task-device matching and allocation are as follows:

[0017] Based on the comprehensive weight of the grid Driven by this approach, the peak grid comprehensive weight of the agricultural facility park is selected as the seed, and connected growth is used to merge adjacent grids to form candidate subtasks until a single vehicle reaches the peak of the workload it can handle. The diameter of the subtask is judged by setting a connectivity threshold. When the diameter exceeds the connectivity threshold, the spatial bisection method is used to split the subtask that exceeds the connectivity threshold into several connected subtasks.

[0018] As a preferred implementation, the specific steps for locally optimizing candidate tasks by constructing an evaluation definition function are as follows:

[0019] The scheduler generates a set of candidate tasks and broadcasts them. An evaluation definition function is then constructed to evaluate the candidate tasks. The formula for the evaluation definition function is as follows:

[0020] ,

[0021] Performing tasks for vehicles The overall valuation score at the time, A is the overall valuation score. The overall valuation score for the i-th agricultural machine is calculated as follows: Let B be the j-th candidate task, i be the index of the unmanned agricultural machinery, and j be the index of the candidate task. For time weighting coefficients, This refers to the actual time the vehicle spends working while it is in motion. This represents the estimated distance the vehicle will travel from its current location to the mission entrance, where `speed` is the vehicle's peak speed. `norm` is the time normalization constant, and `norm` indicates that a normalization operation is performed. This is a weighting factor for energy consumption. Performing tasks for vehicles E represents the estimated energy consumption during the task, where E is the energy consumption of the vehicle performing the task. Let be the energy consumption normalization constant. This is the weighting coefficient for task urgency. Performing tasks for vehicles Urgent indicates the urgency of the task at hand. This is the urgency normalization constant. This is a compatibility weighting coefficient. Let G be the hard constraint related to the compatibility between the i-th unmanned agricultural machine and the j-th candidate task, and let G be the hard constraint related to the compatibility between the vehicle and the candidate task.

[0022] As a preferred embodiment, the specific steps for performing hierarchical path planning for agricultural machinery operations on the assigned tasks and generating executable trajectories that satisfy vehicle dynamics constraints are as follows:

[0023] By constructing a task layer, a trajectory layer, and a local obstacle avoidance layer, the operation path of the unmanned agricultural machine is planned, enabling the unmanned agricultural machine to perceive the surrounding environment in real time and replan the route. The specific layers are as follows: Task layer: By performing directional analysis on the connected grid set allocated to the unmanned agricultural machine and generating equidistant scan bars based on the working width of the unmanned agricultural machine, a waypoint sequence covering the operation path of the unmanned agricultural machine is obtained.

[0024] Trajectory layer: By using curve fitting to the waypoint sequence output by the task layer and constraining the curvature of the trajectory, and combining the peak driving speed of the unmanned agricultural machine to generate a velocity profile and a time-series reference trajectory, a trajectory that satisfies vehicle kinematics constraints and enables the controller to accurately track the time-series execution trajectory is obtained.

[0025] Local obstacle avoidance layer: By deploying a rapid perception-obstacle avoidance module locally on the vehicle, it can perform short-term trajectory prediction and rapid planning local emergency response when central commands are unreachable by implementing trajectory fine-tuning or emergency stop control for short-term sudden obstacles.

[0026] As a preferred implementation, the specific steps for constructing intent broadcast and performing conflict detection are as follows:

[0027] By constructing the intent message in the current task identifier field, an information unit is obtained that can be used by nearby vehicles and the scheduling module for short-term behavior prediction. By setting the intent broadcast period to a range of 0.4–1.8s and the prediction window to a range of 6–18s, risks can be identified in advance and information can be kept updated under limited communication resources. By sending the intent message to nearby vehicles, edge gateways, and the cloud, real-time response to local scheduling can be taken into account. By constructing a vehicle distance calculation formula to calculate the valley Euclidean distance of the predicted trajectories of any two unmanned agricultural machines, a quantitative conflict risk index is obtained. The vehicle distance calculation formula is as follows:

[0028] ,

[0029] dmin is the Euclidean distance between the troughs of the predicted trajectories of the two unmanned agricultural machines, N is the index of the Nth unmanned agricultural machine currently in operation, and M is the index of the Mth unmanned agricultural machine currently in operation. To find the Euclidean distance between the troughs of the predicted trajectories of the two unmanned agricultural machines, For the Fan family symbol, Let N represent the predicted pose trajectory of the Nth unmanned agricultural machine. Let Q represent the predicted pose trajectory of the Mth unmanned agricultural machine, and let Q be the predicted pose trajectory of the unmanned agricultural machine.

[0030] By setting a safe distance threshold and a time threshold, a high-risk conflict is judged when the trough Euclidean distance between the predicted trajectories of the two unmanned agricultural machines is less than or equal to the safe distance threshold and the time period during which the trough Euclidean distance between the predicted trajectories of the two unmanned agricultural machines is less than the safe distance threshold; if the trough Euclidean distance between the predicted trajectories of the two unmanned agricultural machines is only less than the safe distance threshold, it is marked as a warning event, and only real-time observation of the warning vehicle is required; after each speed adjustment or trajectory modification, the intention message is immediately broadcast to update, and the conflict handling result and event log are reported to the scheduling module.

[0031] As a preferred embodiment, the specific steps for rapid processing and remote feedback at the edge by setting online trigger conditions are as follows:

[0032] By setting triggers to be tiered according to priority: high-priority triggers are for security events that are triggered immediately, medium-priority triggers are for operational anomalies that are quickly checked, and low-priority triggers are for events that require continuous monitoring; by setting tiered online trigger conditions, it is possible to respond to events of different severities in a layered manner; by adopting an edge-first, cloud-aggregated feedback architecture: real-time rapid processing is performed at the edge and structured operational data is reported to the cloud, which can meet the real-time response of local rapid processing of edge controllers and support the structured data closed loop of centralized learning.

[0033] As a preferred embodiment, the specific steps for adaptively updating key parameters using online learning are as follows:

[0034] By employing recursion with a forgetting factor for online estimation of linear parameters, parameter calibration of unmanned agricultural machinery can be performed quickly and stably. Online smoothing of the continuous output of unmanned agricultural machinery energy consumption and time predictions is achieved using exponential weighted moving averages, with small-step updates to a small regression network, resulting in a more stable and progressively converging prediction model under varying noise and operating conditions. Locally updated parameters are validated through playback simulations before taking effect. Threshold gating is used to determine whether to gradually roll out the updated parameters.

[0035] The beneficial effects of this invention are:

[0036] 1. By fusing real-time observations from multi-source vision and ground sensors with logarithmic probability and weighting to generate a grid map with confidence, a significant improvement in perception accuracy and timeliness is achieved. This reduces false positives and false negatives, provides reliable input for online replanning, and thus improves the feasibility of the system.

[0037] 2. By combining the connected growth segmentation of unmanned agricultural machinery with auction-local optimization allocation, load balancing and energy consumption optimization are achieved under heterogeneous fleet conditions, and redundant coverage is reduced, thereby shortening the total operation time and generating quantifiable technical effects to support creative judgment.

[0038] 3. By constructing capability vectors, periodic intent broadcasting, hierarchical conflict handling, and an edge-cloud online learning closed loop, real-time collision-free collaboration capabilities and adaptive parameter optimization that evolves with tasks are obtained. This improves system robustness, reduces response latency, and provides measurable evidence of long-term performance improvement to enhance the persuasiveness of authorization. Attached Figure Description

[0039] Figure 1 This is a flowchart of a method for planning the path of multi-machine collaborative operation in agricultural machinery.

[0040] Figure 2 This is a technical rendering of a method for planning the path of multiple agricultural machines working together. Detailed Implementation

[0041] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0042] Example 1 Figure 1 A flowchart of a multi-machine collaborative operation path planning method for agricultural machinery includes: constructing a perception base, specifically: constructing environmental perception; performing grid-based modeling; performing heterogeneous capability matching, specifically: agricultural machinery capability modeling; optimizing and segmenting local tasks; performing high-density safe collaboration, specifically: hierarchically planning agricultural machinery operation paths; constructing intent broadcasting and performing conflict detection; and performing robustness optimization, specifically: setting online trigger conditions, providing remote feedback, and learning online based on the feedback. The specific implementation steps are as follows:

[0043] Step 1: Construct the perception base, specifically: construct environmental perception; perform grid-based modeling.

[0044] Specifically, the steps for building environmental perception are as follows: First, deploying multi-source sensors within the agricultural facility park to obtain operational environment data, including several high-resolution cameras, multispectral cameras, and depth sensors, to obtain visual observation data covering the entire operational area, enabling real-time observation of crop status and obstacle images. Second, deploying wireless sensor network nodes within the agricultural facility park to collect environmental parameters such as soil moisture content, temperature, and humidity, thereby obtaining fixed-point measurement data of the ground environment and filling visual perception blind spots lacking soil and microclimate information. Third, installing positioning modules on vehicles within the agricultural facility park and deploying positioning base stations to obtain unified spatial coordinates for vehicles within the park, thus ensuring visual perception... Ground-based observation and ground-sensor data can be fused within the same world coordinate system. By configuring GPU acceleration units on edge computing nodes and establishing a network connection with the vehicle controller of the unmanned agricultural machinery, the perception latency between ground sensors and decision-making modules can be reduced, and the ability to aggregate data in real time in a cross-device, multi-source, heterogeneous environment can be improved. This meets real-time requirements and facilitates the real-time deployment of grid maps to the upper-level planning module. The upper-level planning module generates task segments and target waypoints based on network weights and device capabilities and issues the high-level operation plan to the decision-making unit of the lower-level trajectory control module. By applying intrinsic parameters and distortion correction parameters to all installed cameras, the accuracy of pixels in spatial coordinates can be improved. The camera performs extrinsic parameter calibration of the positioning module to obtain the spatial transformation relationship between various sensors, thereby ensuring alignment across sensor spaces. An edge timestamp alignment strategy is employed to synchronize network time, ensuring that multi-source observation data can be fused in parallel over time. This edge timestamp alignment strategy involves the edge controller recalibrating and interpolating data packets from different sensors and vehicles according to their arrival time and the local high-precision clock, generating synchronized timestamps on a unified timeline for multi-source data fusion and decision-making. Preprocessing the raw images captured by the camera yields denoised and geometrically corrected image frames, improving the high subsequent detection capability and segmentation accuracy of the initial event detection. The processing includes: preprocessing the acquired raw images such as distortion correction, color normalization, and multi-frame median filtering; the high subsequent detection capability of the initial detection event is to continuously track its evolution after the event occurs and promptly identify subsequent changes, thereby ensuring that path planning is based on the latest observation information; a target detection model is deployed on the vehicle side, and by inputting the preprocessed image into the target detection model, the categories, bounding boxes, and confidence scores of crop rows, crop plants, moving obstacles, and large obstacles are obtained, thereby providing target-level observations for grid occupancy judgment; the target detection model is a deep learning algorithm that identifies images in the camera in real time, locates targets in the images, and outputs target categories, bounding boxes, and confidence scores;By applying multi-frame tracking to the image detection results for real-time monitoring, the pose and velocity estimates of the tracked target are obtained. This enables short-term movement prediction and intent broadcasting of the detected target. The intent broadcasting involves periodically sending the short-term predicted trajectory, confidence level, and velocity of the tracked target to an edge gateway for collision prediction and collaborative decision-making by nearby devices, thereby preventing collisions when agricultural machinery deviates from its pre-planned route.

[0045] The specific steps for grid-based modeling are as follows: First, by dividing the agricultural facility park into a set of two-dimensional grids with fixed dimensions, a unified spatial discretization basis for the agricultural facility park is obtained, facilitating subsequent segmentation, planning, and scheduling of the gridded work areas. Second, when unmanned agricultural machinery detects obstacles, to prevent damage and plan a new route, the comprehensive weight of each grid in the facility agricultural park is recalculated using a linear weighted formula. This yields a priority sequence that can be directly used for task segmentation and scheduling, thereby guiding task allocation. The linear weighted formula is as follows:

[0046] ,

[0047] The overall weight of grid C, The growth state score weighting coefficients for the grid. The score for the crop growth status on grid C. These are the weighting coefficients of the obstacle probability function. Let C be the obstacle probability function. These are the weighting coefficients for the soil condition scoring function. Let be the soil condition scoring function on grid C, where C represents the grid into which the agricultural facility park is divided.

[0048] Step two involves heterogeneous capability matching, specifically: agricultural machinery capability modeling; optimization and segmentation of local tasks.

[0049] Specifically, the steps for assessing agricultural machinery capabilities are as follows: A capability vector is constructed for unmanned agricultural machinery participating in operations within the agricultural facility park, providing precise input for subsequent task feasibility assessment and evaluation functions. This capability vector includes: the unmanned agricultural machinery's processing rate per unit time (deal), continuous processing capacity (sustained), remaining fuel (rest), minimum turning radius (turn), peak speed (speed), energy consumption per unit time (consume), and task compatibility identifier (contain). The capability vector can be obtained by combining parameters provided by the manufacturer, calibration tests conducted on-site within the agricultural facility park, and online operational data. The processing rate per unit time (deal) and peak speed of the unmanned agricultural machinery can be obtained through calibration tests on-site within the agricultural facility park, while the remaining fuel (rest) can be reported in real-time by the onboard controller.

[0050] The specific steps for optimizing and partitioning local tasks are as follows: using the comprehensive weights of the grid... Driven by this mechanism, the peak grid comprehensive weight of the agricultural facility park is selected as the seed. Connectivity growth is used to merge adjacent grids into candidate subtasks until a single vehicle reaches its peak workload capacity. The seed refers to the grid cell selected as the starting point for region merging after sorting the grid comprehensive weights from high to low, triggering the subsequent neighbor expansion task generation process. Connectivity growth is a process of merging adjacent grids that satisfy connectivity, starting from the selected seed grid, according to their adjacency relationships, into connected subtasks. The grid adjacency relationship refers to two grids being connected and adjacent, used to determine grid connectivity and expansion paths. To ensure that each subtask is a connected set during the growth process, a connectivity threshold is set to determine the diameter of the subtask; if the diameter exceeds the connectivity threshold... The spatial bisection method is used to split subtasks exceeding the connectivity threshold into several connected subtasks. The diameter of each subtask is the maximum Euclidean distance between the center points of two grids within that task. The spatial bisection method involves dividing the region into two equal parts along the longest spatial axis of each subtask, and recursively performing diameter and capacity checks on the bisected sub-regions until all sub-regions are less than the connectivity threshold. This improves allocation feasibility and facilitates parallel execution. To ensure that unmanned agricultural machinery is matched to appropriate tasks based on equipment capabilities and execution costs, thereby improving allocation efficiency and accommodating heterogeneous capabilities, a candidate task set is generated and broadcast by the scheduler. An evaluation definition function is constructed to evaluate the candidate tasks. The formula for the evaluation definition function is:

[0051] ,

[0052] Performing tasks for vehicles The overall valuation score at the time, A is the overall valuation score. The overall valuation score for the i-th agricultural machine is calculated as follows: Let B be the j-th candidate task, i be the index of the unmanned agricultural machinery, and j be the index of the candidate task. For time weighting coefficients, This refers to the actual time the vehicle spends working while it is in motion. This represents the estimated distance the vehicle will travel from its current location to the mission entrance, where `speed` is the vehicle's peak speed. `norm` is the time normalization constant, and `norm` indicates that a normalization operation is performed. This is a weighting factor for energy consumption. Performing tasks for vehicles E represents the estimated energy consumption during the task, where E is the energy consumption of the vehicle performing the task. Let be the energy consumption normalization constant. This is the weighting coefficient for task urgency. Performing tasks for vehicles Urgent indicates the urgency of the task at hand. This is the urgency normalization constant. This is a compatibility weighting coefficient. G represents the hard constraints related to the compatibility between the i-th unmanned agricultural machine and the j-th candidate task, and G is the hard constraint related to the compatibility between the vehicle and the candidate task.

[0053] The scheduler sorts candidate tasks by their estimated value from low to high, initially assigns tasks according to the ranking, and verifies the initial assignment results item by item. It determines whether the vehicle's energy is sufficient to complete the task under hard constraints. If any hard constraint is violated, the system rolls back the assigned task, recalculates its value, and assigns it again. When the cost is appropriate, the cost matrix is ​​input into the Hungarian algorithm to find the total cost matching at the lowest possible solution. When the scale is significant, iterative optimization is performed using auctions combined with local exchange heuristics to improve vehicle load balance. The appropriate cost is calculated by the estimation function to ensure that the execution cost of executing candidate tasks does not exceed a set acceptable cost threshold. The execution cost includes: operation time, trips... Factors such as driving cost, energy consumption, and task urgency are considered. The cost matrix is ​​a two-dimensional matrix composed of all participating unmanned agricultural machines and candidate tasks. The Hungarian algorithm returns a suitable total pairing cost after inputting the cost matrix. The significant scale refers to the number of candidate tasks being sufficient to construct a complete cost matrix, thus enabling the Hungarian algorithm. To reduce communication between vehicles and the scheduler, vehicles are only allowed to match a subset of candidate tasks, and only candidate-estimated pairs with task costs below a preset threshold are uploaded. The aggregation matching is completed at the edge nodes, thereby reducing network bandwidth usage and communication latency, alleviating the processing load on edge nodes, and extending the communication energy consumption lifespan of onboard equipment. This improves the system's real-time performance, scalability, and robustness.

[0054] Step 3 involves high-density secure collaboration, specifically: planning agricultural machinery operation paths in layers; constructing intent broadcasting and performing conflict detection.

[0055] Specifically, the steps for hierarchical planning of agricultural machinery operation paths are as follows: To ensure the integrity of the operation and improve its feasibility, the operation path of the unmanned agricultural machinery is planned by constructing a task layer, a trajectory layer, and a local obstacle avoidance layer. This enables the unmanned agricultural machinery to perceive the surrounding environment in real time and replan its route. The specific layers are as follows: Task layer: By performing directional analysis on the connected grid set allocated to the unmanned agricultural machinery and generating equidistant sweep bars based on the working width of the unmanned agricultural machinery, a sequence of waypoints covering the unmanned agricultural machinery operation path is obtained. This provides a structured corridor for the subsequent refinement of the unmanned agricultural machinery operation trajectory while ensuring the integrity of the unmanned agricultural machinery operation path. The equidistant sweep bars are a set of parallel sweep lines generated in the working area with the working width of the unmanned agricultural machinery as the spacing, based on the working area as the main direction, thereby avoiding omissions or repeated coverage of the unmanned agricultural machinery's working area.

[0056] The path waypoint sequence refers to a work path consisting of a series of waypoints arranged in the execution order, which serves as a direct input for subsequent curve fitting, velocity profile generation, and motion control; the work path includes: each point contains coordinates, direction, and speed information of the unmanned agricultural machine.

[0057] Trajectory Layer: By using curve fitting to fit the waypoint sequence output by the task layer and constraining the curvature of the trajectory, and combining it with the peak driving speed of the unmanned agricultural machine to generate a velocity profile and a time-series reference trajectory, a trajectory that satisfies vehicle kinematics constraints and enables the controller to accurately track the time-series execution trajectory is obtained, thereby reducing mechanical wear and improving the smoothness of execution.

[0058] Local obstacle avoidance layer: By deploying a rapid perception-obstacle avoidance module locally on the vehicle, and by implementing trajectory fine-tuning or emergency stop control for short-term sudden obstacles, it can perform short-term trajectory prediction and rapid planning local emergency capabilities when central commands are unreachable, thereby improving system robustness and avoiding collisions when unmanned agricultural machinery changes its operating path.

[0059] The specific steps for constructing intent broadcasting are as follows: by constructing the intent message of the current task identifier field, information units that can be used by nearby vehicles and the scheduling module for short-term behavior prediction are obtained; the intent message of the current task identifier field includes: the identifier of the unmanned agricultural machine, timestamp, predicted trajectory points, etc.; the scheduling module calculates the task allocation and execution order by summarizing perception and status information and sends instructions to the control units of each unmanned agricultural machine to achieve resource optimization and operation coordination in multi-machine collaboration; by setting the intent broadcasting period to a range of 0.4–1.8s and the prediction window to a range of 6–18s, risks can be identified in advance and information can be kept updated under limited communication resources; by sending the intent message to nearby vehicles, edge gateways and the cloud, real-time response and local scheduling can be taken into account.

[0060] The specific steps for conflict detection are as follows: The trough Euclidean distance between the predicted trajectories of any two unmanned agricultural machines is calculated using a vehicle distance calculation formula. This yields a quantified conflict risk index, providing a reliable basis for risk classification and response decisions. The vehicle distance calculation formula is as follows:

[0061] ,

[0062] dmin is the Euclidean distance between the troughs of the predicted trajectories of the two unmanned agricultural machines, N is the index of the Nth unmanned agricultural machine currently in operation, and M is the index of the Mth unmanned agricultural machine currently in operation. To find the Euclidean distance between the troughs of the predicted trajectories of the two unmanned agricultural machines, For the Fan family symbol, Let N represent the predicted pose trajectory of the Nth unmanned agricultural machine. Let Q represent the predicted pose trajectory of the Mth unmanned agricultural machine, and let Q be the predicted pose trajectory of the unmanned agricultural machine.

[0063] By setting a safe distance threshold and a time threshold, a high-risk conflict is identified when the trough Euclidean distance between the predicted trajectories of two unmanned agricultural machines is less than or equal to the safe distance threshold, and the time period during which the trough Euclidean distance between the two unmanned agricultural machines is less than the safe distance threshold is less than the time threshold. If the trough Euclidean distance between the predicted trajectories of the two unmanned agricultural machines is only less than the safe distance threshold, it is marked as a warning event, requiring only real-time observation of the warning vehicle, thereby reducing misjudgments and improving safety. After each speed adjustment or trajectory modification, the intention message is immediately broadcast to update, and the conflict handling results and event logs are reported to the scheduling module, thereby maintaining the consistency of information across the network and providing operational records for offline optimization, providing a data source for subsequent parameter tuning and fault analysis.

[0064] Step four involves robust optimization, specifically setting online trigger conditions, providing remote feedback, and conducting online learning based on the feedback.

[0065] Specifically, the steps for setting online trigger conditions are as follows: Triggers are categorized by priority: high-priority triggers are safety events that trigger immediately, medium-priority triggers are operational anomalies requiring rapid inspection, and low-priority triggers are those requiring continuous monitoring. Examples of high-priority triggers include: a grid occupancy probability increasing from 0.3 to 0.6 within a short period; a single vehicle's remaining energy falling below a preset return-to-home threshold; a link loss time with the edge controller exceeding a threshold; and an unavoidable collision being predicted. Examples of medium-priority and low-priority triggers include: trajectory confidence levels falling below a preset trajectory confidence level, and long-term coverage efficiency degradation. By setting tiered online trigger conditions, different severity events can be responded to in a layered manner, thus prioritizing the safety of unmanned agricultural machinery in emergency situations and gradually optimizing in non-emergency situations.

[0066] The specific steps for remote feedback are as follows: An edge-first, cloud-aggregated feedback architecture is adopted. This architecture reports real-time vehicle events to a nearby edge controller, which then performs low-latency processing and structures the event data before uploading it to the cloud for centralized learning and long-term analysis. This allows for rapid processing at the edge and the uploading of structured operational data to the cloud. This approach satisfies both the real-time response requirements of the edge controller's local rapid processing and the structured data loop for centralized learning, thereby improving event processing efficiency and providing high-quality data for online learning.

[0067] The specific steps for online learning based on feedback are as follows: Online estimation of linear parameters using recursion with a forgetting factor enables rapid and stable parameter calibration of unmanned agricultural machinery. The linear parameters include: grid-weighted linear coefficients and normalized weights in the evaluation function. The recursion with a forgetting factor includes: least squares and weighted least squares. Online smoothing of continuous outputs such as energy consumption and time prediction of unmanned agricultural machinery is achieved using exponentially weighted moving averages, with small-step updates to a small regression network. This yields a more stable and gradually converging prediction model under noise and operating condition changes, reducing the interference of instantaneous fluctuations on decision-making and improving long-term prediction accuracy.

[0068] Before locally updated parameters take effect, the parameters are verified by replaying simulations. A threshold gating principle is used to determine whether to roll out gradually. For example, if the coverage decreases by less than 2% or the conflict rate increases by less than 5% after the parameter update, no rollback is required. If performance degradation occurs after the update, the system will automatically roll back to the most recent stable parameters and report to the audit log.

[0069] like Figure 2 The diagram illustrates the technical effect of a multi-machine collaborative operation path planning method for agricultural machinery. The black bars represent the technical effect of existing technologies, while the gray bars represent the technical effect of the present invention. Higher bar values ​​are better, thus demonstrating that the effect of the present invention is far superior to that of existing technologies.

[0070] Example 2, a method for planning the path of multi-machine cooperative operation in agricultural machinery based on Example 1, the specific scheme is as follows:

[0071] Step 1: Set real-time trigger conditions and perform real-time monitoring;

[0072] Specifically, the system monitors online triggering conditions in real time, such as a sudden increase in grid occupancy probability, vehicle energy below a threshold, and link loss. By setting tiered triggering conditions, it can distinguish between the detection capabilities of emergency and non-emergency events, thus prioritizing safety in emergency situations and supporting gradual optimization in non-emergency situations. The tiered triggering conditions are high-level, medium-level, and low-level. High-level triggering conditions are emergency events that affect safety within a short period of time, including a sudden increase in grid occupancy probability from low to high, and vehicle remaining energy below the return threshold. Medium-level triggering conditions are operational anomalies that require rapid inspection and local adjustment, including trajectory confidence levels below the trajectory confidence threshold and the difference between estimated and measured energy consumption exceeding the allowable percentage. Low-level triggering conditions are statistical signals reflecting long-term performance degradation, including a decrease in coverage efficiency relative to the baseline exceeding a set percentage and energy consumption per unit area consistently exceeding the target range.

[0073] Step two: The edge device processes the monitored events and then sends the processed data to the unmanned agricultural machinery.

[0074] Specifically, when a high-priority event is detected in step one, the edge controller immediately issues a local rapid response command to the affected vehicle and sends a feedback confirmation to the source vehicle. By implementing rapid response and sending feedback with edge priority, a low-latency emergency response capability is achieved, which can shorten the downtime of unmanned agricultural machinery and reduce the risk of accidents.

[0075] Step 3: Summarize the data and train candidate parameters in the cloud;

[0076] Specifically, the edge controller periodically aggregates structured data and uploads it to the cloud. After receiving the aggregated data, the cloud generates candidate parameters through training and sends them to the edge for shadow testing. Through batch training in the cloud and shadow testing at the edge, a stable set of candidate parameters is obtained, thereby improving the long-term allocation and planning accuracy and reducing model bias.

[0077] Step 4: Verify the candidate parameters and set up a rollback mechanism;

[0078] Specifically, when the cloud sends candidate parameters to the edge for execution, it performs replay simulation verification on the candidate parameters. If the performance indicators meet the threshold, the parameters are released; otherwise, the parameters are rolled back to the previous stable candidate parameter and an audit log is recorded. The performance indicators include coverage, energy consumption, and conflict rate. Through replay simulation and rollback mechanisms, erroneous updates that lead to system performance degradation are prevented.

[0079] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for path planning of multi-machine cooperative operation of agricultural machinery, characterized in that: environmental perception is obtained by deploying multi-source sensors to obtain operation environment data; a grid model is established according to the obtained environment data to generate an occupancy and attribute grid map with confidence; the ability of each agricultural machine is modeled by establishing a capability vector; the operation area is divided into several connected local tasks according to the grid map, and task-device matching and distribution are performed; local optimization of candidate tasks is performed by constructing an evaluation definition function; a hierarchical path planning agricultural machine operation path is performed on the distributed tasks to generate an executable trajectory that meets the vehicle dynamics constraints; an intention broadcast is constructed and conflict detection is performed; online triggering conditions are set to perform rapid disposal and remote feedback on the edge side, and key parameters are updated adaptively using online learning. 2.The method according to claim 1, characterized in that: The specific steps of deploying multi-source sensors to obtain operation environment data to construct environmental perception are: obtain visual observation data covering the entire operation area by deploying multi-source sensors in the agricultural facility park; obtain fixed-point measurement data of the ground environment by laying out wireless sensor network nodes in the agricultural facility park to collect soil moisture content, temperature and humidity environmental parameters; obtain the unified spatial coordinates of the vehicles in the park by installing positioning modules on the vehicles in the agricultural facility park and deploying positioning base stations; configure GPU acceleration units on the edge computing nodes and establish network connection with the vehicle controllers of unmanned agricultural machines. 3.The method according to claim 1, characterized in that: The specific steps of establishing a grid model according to the obtained environment data to generate an occupancy and attribute grid map with confidence are: divide the agricultural facility park into a set of two-dimensional grids with fixed size to obtain a unified spatial discretization base for the agricultural facility park; reconstruct the comprehensive weight of each grid in the agricultural facility park by constructing a linear weighted formula to obtain a priority sequence that can be directly used for task slicing and scheduling, and the linear weighted formula is: ， a weight for the grid C, a weight coefficient for the growth status score of the grid, a score for the crop growth status on the grid C, a weight coefficient for the obstacle probability function, an obstacle probability function for the grid C, a weight coefficient for the soil status score function, a soil status score function on the grid C, C represents the grid divided by the agricultural facility park.

4. The method according to claim 1, characterized in that: The specific steps of modeling the ability of each agricultural machine by establishing a capability vector are: construct a capability vector for unmanned agricultural machines participating in operation in the agricultural facility park, which can be obtained from manufacturer-provided parameters, on-site calibration tests in the agricultural facility park, and online operation data; wherein the operation processing rate per unit time deal and the peak speed of the unmanned agricultural machine speed can be obtained by on-site calibration tests in the agricultural facility park, and the fuel remaining amount rest can be reported in real time by the vehicle-mounted controller.

5. The method according to claim 1, characterized in that: The specific steps of dividing the operation area into several connected local tasks according to the grid map and performing task-device matching and distribution are: The integrated weight of the grid As the driving, by selecting the peak grid integrated weight of the agricultural facility park as the seed, the adjacent grids are merged to form candidate sub-tasks by using connected growth until the single vehicle reaches the peak of the workload that can be carried; by setting the connected threshold, the diameter of the sub-task is judged, and when the connected threshold is exceeded, the sub-task exceeding the connected threshold is split into several connected sub-tasks by using the space bisection method. 6.The path planning method for multi-machine coordinated operation of agricultural machines according to claim 1, characterized in that: The specific steps of constructing an evaluation definition function to locally optimize candidate tasks are: The specific steps of constructing an evaluation definition function to locally optimize candidate tasks are: generate a candidate task set through the scheduler and broadcast it, and construct an evaluation definition function to evaluate the candidate tasks, and the formula of the evaluation definition function is: ， the comprehensive evaluation score of the vehicle performing the task A is the comprehensive evaluation score, is the comprehensive evaluation score of the ith agricultural machine, is the jth candidate task, B is the candidate task, i is the index subscript of the unmanned agricultural machine, j is the index subscript of the candidate task, is the time weight coefficient, is the time actually used for work by the vehicle during movement, is the predicted travel distance of the vehicle from the current location to the task entrance, speed is the peak value of the vehicle travel speed, is the time normalization constant, norm indicates that the normalization operation is performed, is the weight coefficient of energy consumption, is the predicted energy consumption of the vehicle performing the task E is the energy consumption of the vehicle performing the task, is the energy consumption normalization constant, is the task urgency weight coefficient, is the task urgency of the vehicle performing the task Urgent indicates the urgency of the current task, is the urgency normalization constant, is the compatibility weight coefficient, is the hard constraint related to the compatibility of the ith unmanned agricultural machine and the jth candidate task, G is the hard constraint related to the compatibility of the vehicle and the candidate task.

7. The method according to claim 1, characterized in that: The specific steps of performing hierarchical path planning of agricultural machine operation path on the distributed tasks to generate an executable trajectory that meets the vehicle dynamics constraints are: The operation path of the unmanned agricultural machine is planned by constructing a task layer, a trajectory layer and a local obstacle avoidance layer, so that the unmanned agricultural machine can perceive the surrounding environment in real time and replan the route, wherein the specific levels are as follows: the task layer: through directional analysis on the connected grid set assigned to the unmanned agricultural machine and generating equidistant sweep strips according to the width of the unmanned agricultural machine work, a sequence of waypoints covering the operation path of the unmanned agricultural machine is obtained; the trajectory layer: through curve fitting method on the sequence of waypoints output by the task layer and combining the peak driving speed of the unmanned agricultural machine to generate a speed profile and a time-sequenced reference trajectory, an execution trajectory which can meet the vehicle dynamics constraints and can be accurately tracked by the controller is obtained; the local obstacle avoidance layer: through local deployment of a fast perception-obstacle avoidance module, through trajectory fine-tuning or emergency stop control on short-term sudden obstacles, the local emergency capability of short-term trajectory prediction and rapid planning can be realized when the central instruction is unreachable. 8.The method according to claim 1, characterized in that: The specific steps of constructing the intention broadcast and conflict detection are as follows: By constructing the intention message of the current task identifier field, the information unit for short-term behavior prediction of the adjacent vehicle and the scheduling module is obtained; by setting the range of the intention broadcast period to 0.4-1.8s and the range of the prediction time window to 6-18s, the risk can be identified in advance and the information can be updated under limited communication resources; by sending the intention message to the adjacent vehicle, the edge gateway and the cloud, the local scheduling can be considered in real time; by constructing the vehicle distance calculation formula to calculate the trough Euclidean distance of the predicted trajectories of any two unmanned agricultural machines, a quantitative conflict risk index is obtained, and the vehicle distance calculation formula is: ， dmin is the trough Euclidean distance of the predicted trajectories of two unmanned farm machines, N is the index of the Nth unmanned farm machine running, M is the index of the Mth unmanned farm machine running, is the trough Euclidean distance of the predicted trajectories of two unmanned farm machines, is the trough Euclidean distance of the predicted trajectories of two unmanned farm machines, represents the predicted pose trajectory of the Nth unmanned farm machine, represents the predicted pose trajectory of the Mth unmanned farm machine, and Q is the predicted pose trajectory of the unmanned farm machine; By setting the safety distance threshold and the time threshold, when the trough Euclidean distance of the predicted trajectories of the two unmanned agricultural machines is less than or equal to the safety distance threshold and the time when the trough Euclidean distance of the predicted trajectories of the two unmanned agricultural machines is less than the safety distance threshold is less than the time threshold, it is judged as a high-risk conflict; only when the trough Euclidean distance of the predicted trajectories of the two unmanned agricultural machines is less than the safety distance threshold, it is marked as a warning event, and only real-time observation of the warning vehicle is needed; after each speed adjustment or trajectory modification, the intention message is immediately broadcasted and updated, and the conflict processing result and the event log are reported to the scheduling module. 9.The method according to claim 1, characterized in that: The specific steps of setting online trigger conditions for rapid disposal on the edge side and remote feedback are as follows: By setting the trigger according to the priority: the high-priority trigger is the safety event triggered immediately, the medium-priority trigger is the running abnormality checked quickly, and the low-priority trigger needs to be monitored continuously; by setting the hierarchical online trigger condition, different severity events can be responded in layers; by adopting the feedback architecture of edge priority and cloud summary: real-time rapid disposal is carried out on the edge side and structured operation data is reported to the cloud, which can meet the real-time response of local rapid disposal of the edge controller and support the structured data closed loop of centralized learning.

10. The method according to claim 1, characterized in that: The specific steps of using online learning to adaptively update the key parameters are as follows: By using recursive estimation with forgetting factor for online estimation of linear parameters, the parameter calibration of unmanned agricultural machines can be quickly and stably performed; by using exponential weighted moving for online smoothing of the continuous output of energy consumption and time prediction of unmanned agricultural machines and updating small regression network with small step, a more stable and gradually convergent prediction model under noise and working condition change is obtained; the locally updated parameters are verified by playback simulation before taking effect; whether to gradually and rollingly release is determined by threshold gating principle.

Images