A Deep Learning-Based Scheduling Method for Agricultural Pesticide Spraying Drones

Abstract

This invention discloses a deep learning-based method for scheduling pesticide spraying drones, comprising the following steps: generating a set of spraying operation states; constructing a two-layer generative graph coupling a task commitment layer and a resource settlement layer; binding spraying tasks and committed drones through a debt flow GFlowNet to form a set of task commitment anchors; identifying unexecutable task commitment anchors and writing them into a resource debt ledger; generating a debt settlement flow and a debt closure record table according to debt type; determining the target spraying scheduling chain from the debt closure candidate spraying scheduling chain and generating a scheduling scheme. This invention utilizes the debt flow GFlowNet scheduling algorithm to generate closed spraying scheduling chains, possessing the advantages of clear resource settlement, high executability, and stable scheduling.

Description

Technical Field

[0001] This invention relates to the field of intelligent drone scheduling technology, and in particular to a deep learning-based method for scheduling pesticide spraying drones. Background Technology

[0002] With the large-scale development of agricultural plant protection operations, pesticide spraying drones have been widely used for the prevention and control of pests and diseases in farmland. Existing scheduling methods typically generate spraying routes based on farmland boundaries, operation areas, drone battery power, remaining pesticide solution, resupply point locations, and operation time windows. Alternatively, they may use genetic algorithms, heuristic search, reinforcement learning, and other methods to sort and allocate multiple spraying tasks to improve operational coverage and reduce flight distance. In multi-drone collaborative scenarios, the flight routes, operation sequence, and resupply timing are also adjusted based on drone location, task status, and weather conditions.

[0003] However, existing methods often treat insufficient pesticide solution, insufficient return-to-base power, drift risks caused by wind speed and direction, resupply queuing, conflicting operation time windows, and airspace conflicts as independent constraints or penalties. When encountering spraying tasks that cannot be directly executed, they typically employ methods such as elimination, demotion, reordering, or partial replanning, lacking a unified recording and closed-loop processing mechanism for the reasons for non-executability. The above methods fail to preserve the correspondence between task commitments and resource replenishment actions, leading to potential issues in the scheduling results such as incomplete resupply arrangements, unclear takeover procedures, lack of verification of avoidance actions, and insufficient executability of the final spraying scheduling chain.

[0004] Therefore, how to provide a deep learning-based method for scheduling pesticide spraying drones is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based scheduling method for pesticide spraying drones. This invention utilizes the debt flow GFlowNet scheduling algorithm to generate a closed spraying scheduling chain, which has the advantages of clear resource settlement, high executability, and stable scheduling.

[0006] A method for scheduling pesticide spraying drones based on deep learning according to an embodiment of the present invention includes the following steps: Generate a set of spraying operation statuses based on farmland boundaries, spraying tasks, drone status, resupply nodes, no-spraying zones, wind speed and direction, and operation time windows; Based on the operational relationships between spraying tasks, drones, and supply nodes, a task commitment layer and a resource settlement layer are constructed. The task commitment layer forms a candidate set of task commitment anchors, and the resource settlement layer forms a set of settlement actions. The task commitment layer and the resource settlement layer are coupled to obtain a two-layer generative graph. The two-layer generative graph is input into the debt flow GFlowNet scheduling algorithm. The commitment generation unit reads the task commitment anchor candidates one by one and binds the spraying task with the commitment drone or candidate drone set to form the task commitment anchor set. The debt identification unit reads the task commitment anchor set and the spraying operation status set, identifies task commitment anchors that cannot be directly executed, and writes them into the resource debt ledger. The debt settlement unit selects the corresponding settlement action from the settlement action set according to the debt type in the resource debt ledger, generates a debt settlement flow, and forms a debt closure record table after the debt settlement flow completes the replenishment, avoidance, waiting, machine replacement or time period adjustment of the corresponding resource debt; The task commitment anchor set, resource debt ledger, and debt closure record table are concatenated into a debt closure candidate spraying scheduling chain set. The target spraying scheduling chain is determined from the debt closure candidate spraying scheduling chain set where all resource debts have been closed, and a pesticide spraying drone scheduling plan is generated.

[0007] Optionally, the generation of the spraying operation status set specifically includes: Within the farmland boundary, the corresponding work area for the spraying task is determined. The work area is divided into strips according to the executable spray width in the drone status and bound to the corresponding work time window to form a set of spraying task units. The system reads the operating area, required amount of pesticide, operating time window, and adjacent operating strips of each spraying task unit according to the spraying task unit set, and forms the task debt identification status. Record the current position, remaining battery power, remaining liquid volume, executable spray width, and return point of each drone according to its status, and form a set of drone operation status records; A set of supply node states is formed based on the location of the supply node, its replenishment capacity, its battery swapping capacity, and its queuing time. Based on the positional relationship between the no-spraying area, wind speed and direction, and the spraying task unit set, a drift risk status set is formed, and an initial flight path occupancy status is formed based on the current position of the drone and the occupied flight path segment; The spraying task unit set, task debt identification status, UAV operation status set, supply node status set, drift risk status set, and initial flight path occupancy status are encapsulated into a spraying operation status set.

[0008] Optionally, the generation of the two-layer generated graph specifically includes: The spraying task units in the spraying task unit set are used as task nodes, and the drones in the drone operation status set are used as drone nodes. The task commitment layer is constructed according to the correspondence between task nodes and drone nodes. According to the task commitment layer, the task node, the drone node and the task commitment candidate edge are encapsulated into task commitment anchor candidates, and then summarized to form a task commitment anchor candidate set. A resource settlement layer is constructed using the supply node status set, drift risk status set, UAV operation status set, operation time window, and initial flight path occupancy status as the source of actions; Settlement action sets are formed according to the resource settlement conditions in the resource settlement layer. The settlement action sets include replenishment action, battery swapping and return action, drift avoidance action, engine swapping and takeover action, and time period migration action. Match the task commitment anchor candidates with the action records in the settlement action set one by one, generate commitment-settlement coupling edges according to the matching results, and form the correspondence between task commitment anchor candidates and callable settlement actions; The task commitment layer, resource settlement layer, task commitment anchor candidate set, settlement action set, and commitment-settlement coupling edge are encapsulated into a two-layer generative graph.

[0009] Optionally, the generation of the task commitment anchor set specifically includes: The commitment generation unit of the debt flow GFlowNet scheduling algorithm reads the task commitment anchor candidate set and commitment-settlement coupling edge in the two-layer generation graph to establish the commitment generation state; The commitment generation unit extracts the range of callable settlement actions for each task commitment anchor candidate based on the settlement entry field and commitment-settlement coupling edge in the commitment generation status; The forward commitment flow allocation unit reads the commitment generation status and the scope of callable settlement actions to generate a forward commitment flow record, including the commitment flow type and the reason for debt retention; The types of commitment flows include direct commitments, debt retention, and non-closable types; The commitment sorting unit generates the commitment sorting result according to the forward commitment flow record corresponding to the same spraying task unit; Based on the commitment ranking results, the committed drones and the candidate drone set are bound together to form a mission commitment anchor; The task commitment anchors corresponding to each spraying task unit are summarized to form a task commitment anchor set.

[0010] Optionally, the debt flow GFlowNet scheduling algorithm includes a commitment generation unit, a forward commitment flow allocation unit, a commitment sorting unit, a debt identification unit, and a debt settlement unit; The commitment generation unit reads the task commitment anchor candidates and their corresponding commitment-settlement coupling edges in the two-layer generation graph, forms a commitment relationship between each spraying task unit and its corresponding UAV, and writes the range of callable settlement actions of the task commitment anchor candidate into the commitment relationship. The forward commitment flow allocation unit reads the remaining liquid volume, required liquid volume, remaining power, return power requirement, candidate execution time period, and range of callable settlement actions in the commitment relationship, and writes the commitment flow type and debt retention reason into the forward commitment flow record. The commitment flow type includes direct commitment type, debt retention type, and non-closable type. The commitment sorting unit determines the commitment UAV and candidate UAV set in the order of priority of direct commitment type over debt retention type, priority of candidate execution time period, and priority of larger remaining power under the same spraying task unit. The debt identification unit converts the task commitment anchor of debt retention type into resource debt in the resource debt ledger. The debt settlement unit selects the settlement action along the commitment-settlement coupling edge corresponding to the resource debt and writes it into the debt closure record table after the resource debt is settled.

[0011] Optionally, the generation of the resource debt ledger specifically includes: The debt identification unit reads the task commitment anchor set, the spraying operation status set, and the commitment-settlement coupling edge in the two-layer generation graph to establish a debt identification base table; The debt identification unit performs resource debt identification according to the debt identification base table to obtain the resource debt identification results; Based on the resource debt identification results, the resource debt is written into the resource debt ledger; Configure debt-settlement mapping records according to the debt type and corresponding commitment-settlement coupling edge in the resource debt record, including a set of primary settlement actions and standby settlement actions; Resource debt records and debt-settlement mapping records are encapsulated into a resource debt ledger.

[0012] Optionally, the configuration of the debt-settlement mapping record determines the main settlement action and the set of backup settlement actions from the scope of callable settlement actions according to the settlement priority sequence corresponding to the debt type. The settlement action that is ranked first in the settlement priority sequence and exists in the scope of callable settlement actions is determined as the main settlement action. The remaining settlement actions that exist in the scope of callable settlement actions and belong to the same debt type settlement priority sequence are written into the set of backup settlement actions. Settlement actions that are not written into the scope of callable settlement actions do not enter the main settlement action and the set of backup settlement actions.

[0013] Optionally, the generation of the debt closure record table specifically includes: The debt settlement unit reads resource debt records and debt-settlement mapping records from the resource debt ledger to form a sequence of resource debts to be settled. Based on the sequence of resources and liabilities to be settled, execute the main settlement action along the corresponding commitment-settlement coupling edge to generate the main settlement flow record; Perform debt closure verification based on the master settlement flow record and generate master settlement verification results; When the main settlement verification result is that the main settlement is not closed, the debt settlement unit generates a standby settlement flow record according to the standby settlement action set; A debt settlement flow is generated based on the main settlement flow record, the main settlement verification result, and the standby settlement flow record; Write the records in the debt settlement flow that are in a closed state into the debt closure record table.

[0014] Optionally, the generation of the pesticide spraying drone scheduling scheme specifically includes: Establish a commitment chain skeleton based on the set of task commitment anchors, and use each task commitment anchor as a commitment node in the commitment chain skeleton; Based on the correspondence between commitment nodes and resource debt records, the resource debt ledger is written into the commitment chain skeleton to form a commitment debt corresponding chain; Based on the correspondence between resource debt records and debt closure record tables, debt closure verification is performed on the corresponding chain of committed debts to form a closed commitment chain segment; Construct a set of candidate spraying scheduling chains for closed-loop commitment chains based on the operational time sequence between segments of the closed commitment chain and the ownership relationship of the committed drones. Determine the target spraying schedule chain from the set of candidate spraying schedule chains with closed debt; A pesticide spraying drone scheduling scheme is generated based on the target spraying scheduling chain.

[0015] The beneficial effects of this invention are: This invention constructs a two-layer generative graph coupling a task commitment layer and a resource settlement layer. It transforms the scheduling relationship between spraying tasks, UAV status, and supply nodes into a candidate set of task commitment anchors, a set of settlement actions, and commitment-settlement coupling edges. This ensures that the range of callable settlement actions is retained when a spraying task enters the scheduling generation process. Therefore, pesticide spraying UAV scheduling no longer relies solely on the order of plot visits or flight path distance as the generation object. Instead, it first forms task commitment anchors, and then performs subsequent debt identification and settlement based on the correspondence between the task commitment anchors and settlement actions. This improves the data connectivity between spraying task allocation, supply arrangement, drift avoidance, and UAV takeover.

[0016] This invention uses a resource debt ledger to uniformly record reasons why tasks cannot be directly executed, such as pesticide supply, return-to-base power, drift risk, operation time windows, resupply queuing, and airspace conflicts. It also maps each type of resource debt to a main settlement action, a set of backup settlement actions, and a commitment-settlement coupling edge, transforming previously scattered constraint judgments into traceable and settleable intermediate states. When a task commitment anchor has issues such as insufficient pesticide supply, insufficient power, or execution time conflicts, the task is not directly deleted or its priority simply lowered. Instead, the GFlowNet debt flow scheduling algorithm preserves settleable task commitment relationships, thereby reducing the premature exclusion of high-priority spraying tasks due to temporary resource shortages.

[0017] This invention further executes the main settlement action and the backup settlement action through a debt settlement flow, and forms a debt closure record table after execution. Only the commitment chain segments where all resource debts are closed participate in the construction of the debt closure candidate spraying scheduling chain. This processing method can perform a closed-loop verification of the settlement results for pesticide replenishment, battery swapping and return, drift avoidance, drone takeover, and time period migration before generating the pesticide spraying drone scheduling plan. This reduces the probability of insufficient pesticide replenishment, insufficient return battery, drift in prohibited spraying areas, resupply queue timeouts, and multi-drone airspace conflicts in the scheduling plan, and improves the executability, stability, and traceability of the final spraying scheduling plan. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based pesticide spraying drone scheduling method proposed in this invention; Figure 2 This is a flowchart illustrating the two-layer generative graph construction process of a deep learning-based pesticide spraying drone scheduling method proposed in this invention. Figure 3 This is a closed-loop flowchart of resource debt for a pesticide spraying drone scheduling method based on deep learning proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A method for scheduling pesticide spraying drones based on deep learning includes the following steps: Generate a set of spraying operation statuses based on farmland boundaries, spraying tasks, drone status, resupply nodes, no-spraying zones, wind speed and direction, and operation time windows; Based on the operational relationships between spraying tasks, drones, and supply nodes, a task commitment layer and a resource settlement layer are constructed. The task commitment layer forms a candidate set of task commitment anchors, and the resource settlement layer forms a set of settlement actions. The task commitment layer and the resource settlement layer are coupled to obtain a two-layer generative graph. The two-layer generative graph is input into the debt flow GFlowNet scheduling algorithm. The commitment generation unit reads the task commitment anchor candidates one by one and binds the spraying task with the commitment drone or candidate drone set to form the task commitment anchor set. The debt identification unit reads the task commitment anchor set and the spraying operation status set, identifies task commitment anchors that cannot be directly executed, and writes them into the resource debt ledger. The debt settlement unit selects the corresponding settlement action from the settlement action set according to the debt type in the resource debt ledger, generates a debt settlement flow, and forms a debt closure record table after the debt settlement flow completes the replenishment, avoidance, waiting, machine replacement or time period adjustment of the corresponding resource debt; The task commitment anchor set, resource debt ledger, and debt closure record table are concatenated into a debt closure candidate spraying scheduling chain set. The target spraying scheduling chain is determined from the debt closure candidate spraying scheduling chain set where all resource debts have been closed, and a pesticide spraying drone scheduling plan is generated.

[0021] In this embodiment, the generation of the spraying operation status set specifically includes: Within the farmland boundary, the corresponding work area for the spraying task is determined. The work area is divided into strips according to the executable spray width in the drone status and bound to the corresponding work time window to form a set of spraying task units. The system reads the operating area, required amount of pesticide, operating time window, and adjacent operating strips of each spraying task unit according to the spraying task unit set, and forms the task debt identification status. Record the current position, remaining battery power, remaining liquid volume, executable spray width, and return point of each drone according to its status, and form a set of drone operation status records; A set of supply node states is formed based on the location of the supply node, its replenishment capacity, its battery swapping capacity, and its queuing time. Based on the positional relationship between the no-spraying area, wind speed and direction, and the spraying task unit set, a drift risk status set is formed, and an initial flight path occupancy status is formed based on the current position of the drone and the occupied flight path segment; The spraying task unit set, task debt identification status, UAV operation status set, supply node status set, drift risk status set, and initial flight path occupancy status are encapsulated into a spraying operation status set.

[0022] In this embodiment, the generation of the two-layer generated graph specifically includes: The spraying task units in the spraying task unit set are used as task nodes, and the drones in the drone operation status set are used as drone nodes. The task commitment layer is constructed according to the correspondence between task nodes and drone nodes. The task commitment layer includes task nodes, drone nodes, and task commitment candidate edges. Each task commitment candidate edge records the required amount of pesticide, operation time window, adjacent operation strips, current position of the corresponding drone, remaining battery power, remaining amount of pesticide, executable spray width, return point, and return battery power requirement for the corresponding spraying task unit. The return battery power requirement is the amount of electricity required for the corresponding drone to reach the spraying task unit from its current position, complete the spraying task unit, and return to the return point. The task commitment candidate edge records the candidate commitment relationship between the spraying task unit and the drone, and is not used as a determination result that the spraying task unit can be directly executed by the corresponding drone. According to the task commitment layer, the task node, the drone node and the task commitment candidate edge are encapsulated into task commitment anchor candidates, and then summarized to form a task commitment anchor candidate set. The task commitment anchor wait options include spraying task unit, corresponding drone, required amount of liquid, operation time window, candidate execution period, adjacent operation strip, return point, return power requirement and non-committed status. The candidate execution period is formed by dividing the operation time window. The task commitment anchor wait options do not change the drone's remaining power and remaining amount of liquid. A resource settlement layer is constructed using the supply node status set, drift risk status set, UAV operation status set, operation time window, and initial flight path occupancy status as the source of actions; The resource settlement layer includes replenishment nodes, battery swapping and return nodes, drift avoidance nodes, drone takeover nodes, and time-slot migration nodes. Replenishment nodes are formed by replenishment capabilities in the supply node status set. Battery swapping and return nodes are formed by battery swapping capabilities in the supply node status set and return points in the drone operation status set. Drift avoidance nodes are formed by spraying task units, drift directions, drift impact ranges, and corresponding no-spray areas in the drift risk status set. Drone takeover nodes are formed by drones that can correspond to the same spraying task unit in the drone operation status set. Time-slot migration nodes are formed by operation time windows and initial flight path occupancy status. Settlement action sets are formed according to the resource settlement conditions in the resource settlement layer. The settlement action sets include replenishment action, battery swapping and return action, drift avoidance action, engine swapping and takeover action, and time period migration action. When the remaining liquid amount of the drone is less than the required liquid amount of the spraying task unit, the corresponding replenishment node, corresponding drone, liquid amount before replenishment, liquid amount after replenishment, and replenishment service period are written into the replenishment action. When the remaining battery power of the drone is less than the return battery power requirement, the corresponding replenishment node, corresponding drone, return point, battery power before battery swap, battery power after battery swap, and battery swap service period are written into the battery swap return action. When the drift influence range of the spraying task unit in the current wind speed and wind direction overlaps with the no-spray area, the corresponding spraying task unit, corresponding no-spray area, drift direction, execution period after avoidance, and spraying parameters after avoidance are written into the drift avoidance action. When there are more than two corresponding drones for a spraying task unit, the original corresponding drone, replacement drone, corresponding spraying task unit, and execution period after takeover are written into the drone takeover action. When the candidate execution period of the spraying task unit overlaps with the initial route occupancy status in both time and space, the corresponding spraying task unit, candidate execution period, execution period after migration, and route occupancy conflict record are written into the time period migration action. Match the task commitment anchor candidates with the action records in the settlement action set one by one, generate commitment-settlement coupling edges according to the matching results, and form the correspondence between task commitment anchor candidates and callable settlement actions; Item-by-item matching includes: When the corresponding drone in the replenishment action is the same as the corresponding drone in the mission commitment anchor option, and the amount of medicine after replenishment in the replenishment action is not less than the required amount of medicine in the mission commitment anchor option, the mission commitment anchor option, the replenishment action and the medicine supply matching identifier are written into the same commitment-settlement coupling edge. When the corresponding drone and return point in the battery swapping and return action are consistent with the corresponding drone and return point in the mission commitment anchor waiting option, and the battery power after the battery swapping and return action is not less than the return power requirement in the mission commitment anchor waiting option, the mission commitment anchor waiting option, the battery swapping and return action and the return power matching identifier are written into the same commitment-settlement coupling edge. When the spraying task unit in the drift avoidance action is consistent with the spraying task unit in the task commitment anchor waiting option, and the execution period after avoidance in the drift avoidance action is within the operation time window in the task commitment anchor waiting option, the task commitment anchor waiting option, the drift avoidance action and the drift risk matching identifier are written into the same commitment-settlement coupling edge. When the spraying task unit in the aircraft takeover action is the same as the spraying task unit in the task commitment anchor option, and the replacement drone in the aircraft takeover action belongs to the corresponding drone in the task commitment anchor option corresponding to the same spraying task unit, and the replacement drone is different from the corresponding drone in the task commitment anchor option, the task commitment anchor option, the aircraft takeover action and the aircraft takeover matching identifier are written into the same commitment-settlement coupling edge. When the spraying task unit in the time period migration action is consistent with the spraying task unit in the task commitment anchor waiting option, and the execution time period after migration is within the operation time window in the task commitment anchor waiting option, write the task commitment anchor waiting option, the time period migration action and the time period migration matching identifier into the same commitment-settlement coupling edge. The task commitment layer, resource settlement layer, task commitment anchor candidate set, settlement action set, and commitment-settlement coupling edge are encapsulated into a two-layer generative graph.

[0023] In this embodiment, the generation of the task commitment anchor set specifically includes: The commitment generation unit of the debt flow GFlowNet scheduling algorithm reads the task commitment anchor candidate set and commitment-settlement coupling edge in the two-layer generation graph to establish the commitment generation state; The commitment generation status is written into the commitment generation field for each task commitment anchor option, including the spraying task unit, corresponding drone, required amount of pesticide, operation time window, candidate execution time period, return point, and return power requirement. According to the pesticide supply matching identifier, return power matching identifier, drift risk matching identifier, drone takeover matching identifier, and time period migration matching identifier in the commitment-settlement coupling edge, a settlement entry field is established for the task commitment anchor option. The commitment generation unit extracts the range of callable settlement actions for each task commitment anchor candidate based on the settlement entry field and commitment-settlement coupling edge in the commitment generation status; The extraction of the scope of callable settlement actions is achieved by reading the settlement actions written to the same commitment-settlement coupling edge as the task commitment anchor option for any task commitment anchor option. When the commitment-settlement coupling edge writes a liquid supply matching identifier, the corresponding replenishment action is written into the scope of callable settlement actions. When the commitment-settlement coupling edge writes a return power matching identifier, the corresponding battery swap return action is written into the scope of callable settlement actions. When the commitment-settlement coupling edge writes a drift risk matching identifier, the corresponding drift avoidance action is written into the scope of callable settlement actions. When the commitment-settlement coupling edge writes a generator takeover matching identifier, the corresponding generator takeover action is written into the scope of callable settlement actions. When the commitment-settlement coupling edge writes a time period migration matching identifier, the corresponding time period migration action is written into the scope of callable settlement actions. Settlement actions not written to the same commitment-settlement coupling edge as the task commitment anchor option are not written into the scope of callable settlement actions. The forward commitment flow allocation unit reads the commitment generation status and the scope of callable settlement actions to generate a forward commitment flow record, including the commitment flow type and the reason for debt retention; The types of commitment flows include direct commitments, debt retention, and non-closable types; Among them, if the remaining amount of medicine in the current task commitment anchor waiting option is not less than the required amount of medicine, and the remaining power is not less than the return power requirement, and the scope of callable settlement actions does not include drift avoidance actions and time period migration actions, then the commitment flow type of the task commitment anchor waiting option shall be written as direct commitment type. Where the remaining liquid volume in the current mission commitment anchor waiting option is less than the required liquid volume and the scope of callable settlement actions includes replenishment action or aircraft takeover action; the remaining power in the current mission commitment anchor waiting option is less than the return power requirement and the scope of callable settlement actions includes power swap return action or aircraft takeover action; or the scope of callable settlement actions in the current mission commitment anchor waiting option includes drift avoidance action or time period migration action, the commitment flow type of the mission commitment anchor waiting option shall be written as debt retention type, and one or more debt retention reasons shall be written according to the triggered debt retention conditions, including liquid supply debt retention reason, return power debt retention reason, drift risk debt retention reason or time period migration debt retention reason; When the current task commitment anchor candidate cannot be written to the direct commitment type and there is no corresponding callable settlement action, the commitment flow type of the task commitment anchor candidate is written as an unclosed type, and the task commitment anchor candidate is prevented from entering the commitment sorting result. The commitment sorting unit generates the commitment sorting result according to the forward commitment flow record corresponding to the same spraying task unit; The commitment ranking results will group task commitment anchor candidates with the same spraying task unit into the same commitment ranking group. Task commitment anchor candidates with the commitment flow type of direct commitment will be ranked before task commitment anchor candidates with the commitment flow type of debt retention. For task commitment anchor candidates with the same commitment flow type of debt retention, they will be arranged according to the order of the candidate execution time. When there are two or more task commitment anchor candidates with the same ranking position in the same commitment ranking group, the drone with the larger remaining battery power will be ranked before the drone with the smaller remaining battery power. Based on the commitment ranking results, the committed drones and the candidate drone set are bound together to form a mission commitment anchor; Specifically, in each commitment ranking result, the top-ranked task commitment anchor candidate is read, and the corresponding drone in the task commitment anchor candidate is written into the commitment drone field. The task commitment anchor candidate other than the top-ranked one and whose commitment flow type is not non-closable is read, and the corresponding drone in the candidate drone set is written into the candidate drone set. The spraying task unit, commitment drone, candidate drone set, required amount of liquid, operation time window, candidate execution period, return point, return power requirement, forward commitment flow record, commitment flow type, debt retention reason, scope of callable settlement action and unidentified debt status are written into the same task commitment anchor. Summarize the task commitment anchors corresponding to each spraying task unit to form a task commitment anchor set; Specifically, the task commitment anchors are checked item by item according to the spraying task unit. The committed drones, candidate drone sets, forward commitment flow records, commitment flow type, debt retention reasons and scope of callable settlement actions are retained in each task commitment anchor. The retained task commitment anchors are sorted according to the operation time window of the spraying task unit to form a task commitment anchor set.

[0024] In this embodiment, the debt flow GFlowNet scheduling algorithm includes a commitment generation unit, a forward commitment flow allocation unit, a commitment sorting unit, a debt identification unit, and a debt settlement unit; The commitment generation unit reads the task commitment anchor candidates and their corresponding commitment-settlement coupling edges in the two-layer generation graph, forms a commitment relationship between each spraying task unit and the corresponding UAV, and writes the range of callable settlement actions of the task commitment anchor candidate into the commitment relationship. The forward commitment flow allocation unit reads the remaining liquid volume, required liquid volume, remaining power, return power requirement, candidate execution time period and range of callable settlement actions in the commitment relationship, and writes the commitment flow type and debt retention reason in the forward commitment flow record. The commitment flow type includes direct commitment type, debt retention type and non-closable type. The commitment sorting unit determines the commitment UAV and candidate UAV set in the order of priority of direct commitment type over debt retention type, priority of candidate execution time period, and priority of larger remaining power under the same spraying task unit. The debt identification unit converts the task commitment anchor of debt retention type into resource debt in the resource debt ledger. The debt settlement unit selects the settlement action along the commitment-settlement coupling edge corresponding to the resource debt and writes it into the debt closure record table after the resource debt is settled. During the training phase of the GFlowNet scheduling algorithm for debt flow, each training generated trajectory is divided into a task commitment segment, a resource debt segment, and a debt settlement segment. The forward generation probability of the task commitment segment, the resource debt segment, and the debt settlement segment are multiplied together to obtain the forward trajectory probability. The backward backtracking probability of the task commitment segment, the resource debt segment, and the debt settlement segment are multiplied together to obtain the backward trajectory probability. The forward generation probability of the task commitment segment is obtained by multiplying the forward commitment flow records of the task commitment anchor candidates. The forward generation probability of the resource debt segment is obtained by multiplying the debt generation probabilities of the resource debt records. The forward generation probability of the debt settlement segment is obtained by multiplying the action selection probabilities of the main settlement action and the backup settlement action. For each training trajectory, a closure indicator value is generated. When each resource debt record in the resource debt ledger has a debt closure record with the same task commitment anchor, the same debt type, and the same commitment-settlement coupling edge in the debt closure record table, the closure indicator value is closed. When there are unsettled resource debt records, non-closable type task commitment anchor candidates, or unmatched debt closure records, the closure indicator value is not closed. The termination reward is jointly determined by the closure indicator value, the number of spraying task units covered, the number of times the replenishment action and battery swap return action are called, the number of times the power is insufficient after settlement, the number of times the time window is not met after settlement, the number of times the drift overlap is after settlement, and the number of times the airspace conflict is after settlement. Among them, the number of spraying task units covered is the benefit item, and the rest are the cost items. The trajectory balance loss is calculated using normalized flow parameters, forward trajectory probability, backward trajectory probability, and termination reward. Specifically, the logarithm of the normalized flow parameters is added to the logarithm of the forward trajectory probability, and then the logarithm of the backward trajectory probability and the logarithm of the termination reward are subtracted to obtain the trajectory balance deviation. The square of the trajectory balance deviation is used as the trajectory balance loss. For training-generated trajectories with closure indication values ​​that are not closed, the termination reward feedback samples are not included, and the debt non-closure loss is calculated. The debt non-closure loss is obtained by adding the number of unsettled state resource debt records, the number of non-closable type task commitment anchor candidates, and the number of unmatched debt closure records. The total training loss is calculated per training batch. For each training trajectory generated in a training batch, the trajectory balancing loss is added to the weighted debt non-closure loss to obtain the single trajectory loss. The sum of all single trajectory losses in the training batch is divided by the number of training trajectories to obtain the total loss. The training convergence condition is that the absolute value of the difference in total loss between adjacent training batches is not greater than the training loss change threshold, and the batch average of the number of unsettled state resource debt records, the number of non-closable type task commitment anchor candidates, and the number of unmatched debt closure records are all zero. When the conditions are met, the debt flow GFlowNet scheduling algorithm with completed training is output. The debt flow GFlowNet scheduling algorithm transforms the single termination object generation process in the GFlowNet generation flow network into retaining settlementable but not directly executable tasks during the task commitment phase. It then delivers the task commitment anchor to the subsequent resource debt ledger through the commitment flow type and debt retention reason. Subsequently, the debt settlement flow calls the replenishment action, battery swap return action, drift avoidance action, generator takeover action, or time period migration action in the settlement action set along the commitment-settlement coupling edge. The corresponding resource debt in the resource debt ledger forms a debt closure record as the basis for the generation flow to continue to expand or be allowed to terminate. This limits the termination object output by the debt flow GFlowNet scheduling algorithm to a debt closure candidate spraying scheduling chain where all resource debts have been closed.

[0025] In this embodiment, the generation of the resource debt ledger specifically includes: The debt identification unit reads the task commitment anchor set, the spraying operation status set, and the commitment-settlement coupling edge in the two-layer generation graph to establish a debt identification base table; The debt identification unit reads the spraying task unit, committed drone, candidate drone set, required liquid volume, operation time window, candidate execution period, return point, return power requirement, committed flow type, debt retention reason and scope of callable settlement action from the task commitment anchor item by item. It also reads the task debt identification status, drone operation status set, supply node status set, drift risk status set and initial route occupation status from the spraying operation status set, and writes the data corresponding to the same task commitment anchor into the same debt identification base table record. The debt identification unit performs resource debt identification according to the debt identification base table to obtain the resource debt identification results; Resource debt identification includes: marking a liquid supply debt when the remaining liquid level of the committed drone in the task commitment anchor is less than the required liquid level in the task commitment anchor; marking a return-to-home power debt when the remaining battery power of the committed drone in the task commitment anchor is less than the return-to-home power requirement in the task commitment anchor; marking a drift risk debt when the drift impact range corresponding to the spraying task unit of the task commitment anchor in the drift risk status set overlaps with the no-spray zone, or when the debt retention reason of the task commitment anchor includes a drift risk debt retention reason; and marking a drift risk debt when the candidate execution period of the task commitment anchor is not within the operation time window of the task commitment anchor, or When the reason for retaining the debt of the task commitment anchor includes the reason for retaining the debt due to time period migration, it is marked as operation time window debt. When there is a replenishment action or a battery swap return action in the scope of the callable settlement action of the task commitment anchor, and the end time of the replenishment service period of the replenishment action or the battery swap service period of the battery swap return action is later than the end time of the operation time window of the task commitment anchor, it is marked as supply queuing debt. When there is a time overlap and spatial overlap between the candidate execution period of the task commitment anchor and the initial route occupancy status, it is marked as airspace conflict debt. When the same task commitment anchor triggers more than two debt markings, the debt markings are retained in the resource debt identification results. Based on the resource debt identification results, the resource debt is written into the resource debt ledger; The resource debt ledger establishes a resource debt record for each debt marker in the resource debt identification result, and writes the corresponding task commitment anchor, debt type, debt retention reason, committed drone, candidate drone set, corresponding commitment-settlement coupling edge, scope of callable settlement actions and unsettled status into the same resource debt record. When the same task commitment anchor has more than two debt markers in the resource debt identification result, two or more resource debt records are established respectively and associated with the same task commitment anchor. Configure debt-settlement mapping records according to the debt type and corresponding commitment-settlement coupling edge in the resource debt record, including a set of primary settlement actions and standby settlement actions; Resource debt records and debt-settlement mapping records are encapsulated into a resource debt ledger.

[0026] In this embodiment, the configuration of the debt-settlement mapping record is based on the settlement priority sequence corresponding to the debt type. The main settlement action and the set of backup settlement actions are determined from the scope of callable settlement actions. The settlement action that is ranked first in the settlement priority sequence and exists in the scope of callable settlement actions is determined as the main settlement action. The other settlement actions that exist in the scope of callable settlement actions and belong to the same debt type settlement priority sequence are written into the set of backup settlement actions. Settlement actions that are not written into the scope of callable settlement actions do not enter the main settlement action and the set of backup settlement actions. The settlement priority sequence includes: liquid supply debt corresponds to replenishment action and aircraft takeover action in sequence; return power debt corresponds to battery swap return action and aircraft takeover action in sequence; drift risk debt corresponds to drift avoidance action; operation time window debt corresponds to time period migration action; replenishment queuing debt corresponds to replenishment action after service period adjustment, battery swap return action after service period adjustment, and time period migration action in sequence; and airspace conflict debt corresponds to time period migration action and aircraft takeover action in sequence.

[0027] In this embodiment, the generation of the debt closure record table specifically includes: The debt settlement unit reads resource debt records and debt-settlement mapping records from the resource debt ledger to form a sequence of resource debts to be settled. The pending resource debt sequence is generated according to the task commitment anchor, debt type, unsettled status, main settlement action, set of standby settlement actions and corresponding commitment-settlement coupling edge in the resource debt record. The unsettled status restricts the resource debt record from entering the debt settlement flow. Resource debt records that have been written into the debt closure record table will no longer enter the pending resource debt sequence. Based on the sequence of resources and liabilities to be settled, execute the main settlement action along the corresponding commitment-settlement coupling edge to generate the main settlement flow record; The main settlement actions include: when the main settlement action is a replenishment action, the amount of medicine after replenishment is written into the execution medicine amount corresponding to the task commitment anchor; when the main settlement action is a battery swap and return action, the amount of electricity after battery swap is written into the execution electricity corresponding to the task commitment anchor; when the main settlement action is a drift avoidance action, the execution time period after avoidance and the spraying parameters after avoidance are written into the spraying execution field corresponding to the task commitment anchor; when the main settlement action is a drone takeover action, the replacement drone is written into the committed drone field corresponding to the task commitment anchor; when the main settlement action is a time period migration action, the execution time period after migration is written into the candidate execution time period corresponding to the task commitment anchor. Perform debt closure verification based on the master settlement flow record and generate master settlement verification results; The debt closure verification includes: when the amount of liquid medicine executed corresponding to the liquid medicine supply debt is not less than the amount of liquid medicine required in the mission commitment anchor, it is recorded as liquid medicine supply debt closure; when the amount of electricity executed corresponding to the return power debt is not less than the return power demand in the mission commitment anchor, it is recorded as return power debt closure; when the drift risk debt's drift impact range no longer overlaps with the no-spray zone, it is recorded as drift risk debt closure; when the execution period corresponding to the operation time window debt is within the operation time window in the mission commitment anchor, it is recorded as operation time window debt closure; when the end time of the supply service period corresponding to the supply queuing debt is not later than the end time of the operation time window in the mission commitment anchor, or when the execution period after migration is within the operation time window in the mission commitment anchor, it is recorded as supply queuing debt closure; when the execution period corresponding to the airspace conflict debt does not overlap in time or space with the initial route occupancy status, it is recorded as airspace conflict debt closure; resource debts that do not meet the corresponding debt closure conditions are recorded as main settlement not closed. When the main settlement verification result is that the main settlement is not closed, the debt settlement unit generates a standby settlement flow record according to the standby settlement action set; The standby settlement flow record is read by the debt settlement unit according to the settlement priority sequence in the debt-settlement mapping record. The standby settlement actions are executed one by one. After each standby settlement action is executed, the execution liquid volume, execution electricity, spraying execution field, commitment drone field or candidate execution time period in the corresponding task commitment anchor are updated. The debt settlement unit performs debt closure verification again on the updated task commitment anchor. The standby settlement action that makes the resource debt reach the corresponding debt closure condition for the first time is written into the standby settlement flow record. When all standby settlement actions in the standby settlement action set do not make the resource debt reach the corresponding debt closure condition, the resource debt record is kept in an unsettled state. A debt settlement flow is generated based on the main settlement flow record, the main settlement verification result, and the standby settlement flow record; The debt settlement flow includes the task commitment anchor, debt type, corresponding commitment-settlement coupling edge, main settlement action to be executed, backup settlement action to be executed, liquid volume to be executed after settlement, electricity volume to be executed after settlement, execution period after settlement, committed drone to be executed after settlement, and settlement status, including closed status and unsettled status. Write the records in the debt settlement flow that are in a closed state into the debt closure record table.

[0028] In this embodiment, the generation of the pesticide spraying drone scheduling scheme specifically includes: Establish a commitment chain skeleton based on the set of task commitment anchors, and use each task commitment anchor as a commitment node in the commitment chain skeleton; The commitment chain skeleton reads the spraying task unit, committed drone, candidate drone set, required amount of liquid, operation time window, candidate execution period, return point, return power requirement, forward commitment flow record, commitment flow type and scope of callable settlement action from each task commitment anchor, writes them into the same commitment node, and arranges the commitment nodes according to the operation time window of the spraying task unit. Based on the correspondence between commitment nodes and resource debt records, the resource debt ledger is written into the commitment chain skeleton to form a commitment debt corresponding chain; The commitment debt corresponding chain reads the resource debt record with the same task commitment anchor as the commitment node in the resource debt ledger for each commitment node. When the commitment node does not have a corresponding resource debt record, the commitment node is marked as a debt-free commitment node. When the commitment node has a corresponding resource debt record, the debt type, corresponding commitment-settlement coupling edge, main settlement action, set of backup settlement actions and unsettled status in the corresponding resource debt record are written to the commitment node. Based on the correspondence between resource debt records and debt closure record tables, debt closure verification is performed on the corresponding chain of committed debts to form a closed commitment chain segment; Among them, the debt closure verification reads the debt closure record table that has the same task commitment anchor, the same debt type and the same commitment-settlement coupling edge as the resource debt record for the commitment node written into the resource debt record. When all resource debt records corresponding to the commitment node have corresponding debt closure records, the commitment node, the corresponding resource debt record and the corresponding debt closure record are written into the same closed commitment chain segment. Construct a set of candidate spraying scheduling chains for closed-loop commitment chains based on the operational time sequence between segments of the closed commitment chain and the ownership relationship of the committed drones. The construction of the debt closure candidate spraying scheduling chain set includes: arranging the closed commitment chain segments belonging to the same committed drone according to the candidate execution time period or the execution time period after settlement in the debt closure record; when the closed commitment chain segment contains a drone takeover action, the committed drone after settlement in the debt closure record is taken as the committed drone of the closed commitment chain segment; when the closed commitment chain segment contains a refill action, a battery swap return action, a drift avoidance action, or a time period migration action, the executed liquid volume after settlement, the executed electricity volume after settlement, the executed time period after settlement, and the settlement action to close the resource debt in the debt closure record are written into the corresponding closed commitment chain segment; and constructing the closed commitment chain segments arranged continuously under the same committed drone as the debt closure candidate spraying scheduling chain. Determine the target spraying schedule chain from the set of candidate spraying schedule chains with closed debt; The target spraying scheduling chain is determined by removing candidate spraying scheduling chains that contain unsettled resource debt records, removing candidate spraying scheduling chains that contain non-closable task commitment anchors, and retaining candidate spraying scheduling chains that have closed all resource debts. The retained candidate spraying scheduling chains are sorted in the following order: the number of spraying task units covered, the continuity of candidate execution time, the number of calls to replenishment and battery swapping return actions, and the number of airspace conflict debts. The candidate spraying scheduling chain with the highest ranking is determined as the target spraying scheduling chain. Generate a pesticide spraying drone scheduling scheme based on the target spraying scheduling chain; The pesticide spraying drone scheduling scheme includes drone task allocation results, execution order of spraying task units, committed drones, candidate drone set, execution time period corresponding to spraying task unit, refilling action, battery swapping and return action, drifting and avoidance action, drone takeover action, time period migration action, settled execution liquid volume, settled execution electricity volume, and debt closure record table. Each closed commitment chain segment in the target spraying scheduling chain is written into the pesticide spraying drone scheduling scheme according to the committed drone and execution time period.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the scheduling scenario of pesticide spraying drones in a contiguous rice planting base. During pest and disease control operations, this base is characterized by scattered plots, numerous field ridges and ditches, some areas adjacent to fishponds and residential access roads, and frequent drone refills and battery swaps. Previously, scheduling methods primarily arranged operations according to plot order and remaining drone battery power. When wind direction changed, refill points were queued, a drone ran out of pesticide, or adjacent flight paths were occupied, problems easily arose such as tasks being temporarily skipped, unclear takeover procedures, conflicting execution times after refills, and the need for repeated manual adjustments to already generated flight paths. This resulted in spraying tasks, although assigned, not landing continuously during actual execution.

[0030] In this scenario, data is first collected on farmland boundaries, spraying tasks, drone status, resupply nodes, no-spray zones, wind speed and direction, and operation time windows to form a spraying operation status set. The system determines the operation area corresponding to the spraying task based on the farmland boundaries and divides the spraying task into units according to the drone's executable spray width, ensuring that each spraying task unit corresponds to a specific operation time window, required pesticide volume, and adjacent operation strips. Subsequently, the candidate commitment relationships between spraying task units and drones are written into the task commitment layer, and resupply actions, battery swapping and return actions, drift avoidance actions, drone takeover actions, and time-slot migration actions are written into the resource settlement layer. Finally, a correspondence between task commitment anchor options and callable settlement actions is established through commitment-settlement coupling edges, thus forming a two-layer generation graph.

[0031] Before deployment, the debt flow GFlowNet scheduling algorithm was trained using the base's historical flight logs, resupply records, weather records, no-spray zone records, and manual adjustment records. Training samples included directly executable spraying tasks, as well as task trajectories requiring resupply, battery swapping, drift avoidance, aircraft takeover, or time-slot migration due to insufficient pesticide, insufficient return-to-base power, drift risk, resupply queuing, and airspace conflicts. During training, each generated trajectory was divided into a task commitment segment, a resource debt segment, and a debt settlement segment, respectively learning the generation relationships between task commitment anchor generation, resource debt writing, and settlement action invocation. For trajectories with resource debt that had formed a closed debt record, the algorithm used them as termination reward feedback samples; for trajectories with unsettled resource debt records, non-closable task commitment anchor candidates, or unmatched debt closure records, the algorithm calculated the debt non-closure loss and reduced the sampling bias of the corresponding generated path. After training convergence, the debt flow GFlowNet scheduling algorithm can prioritize generating spraying scheduling chains with closable resource debt.

[0032] In practical applications, the GFlowNet scheduling algorithm reads the two-layer generated graph, reads the task commitment anchor candidates item by item, and binds the spraying task with the commitment drone or candidate drone set to form a task commitment anchor set. For tasks that can be executed with sufficient liquid and power, the commitment flow type is directly written; for tasks that cannot be executed directly but involve replenishment, battery swapping and return, drift avoidance, drone takeover, or time period migration, they are not immediately deleted, but the debt retention reason is written. The debt identification unit then writes the liquid supply, return power, drift risk, operation time window, replenishment queue, and airspace conflict into the resource debt ledger according to the task commitment anchor set and the spraying operation status set. The debt settlement unit calls the main settlement action according to the debt type and commitment-settlement coupling edge in the resource debt ledger; if the main settlement action fails to close the corresponding resource debt, the set of backup settlement actions is called, and the closed resource debt is written into the debt closure record table.

[0033] On-site operation records show that during continuous spraying operations in the same work season, flight control logs, resupply node queuing records, weather station wind speed and direction records, and scheduling backend records are mutually consistent, forming a complete traceability chain between task commitment anchors, resource debt ledgers, debt settlement flows, and debt closure record tables. When the drone's pesticide solution is insufficient, the scheduling scheme can invoke resupply actions or drone takeover actions from the corresponding commitment-settlement coupling edge; when wind direction poses a drift risk to the spraying task unit, the task can be written into the drift risk debt and a drift avoidance action can be invoked; when resupply node queuing affects the original execution time, the execution time can be rescheduled through time period migration actions. Thus, the final pesticide spraying drone scheduling scheme is no longer just a simple plot sequence or flight path result, but an executable scheduling chain that includes task allocation, resupply arrangement, avoidance handling, drone takeover, time period migration, and debt closure records, which can effectively reduce temporary interruptions and repeated manual adjustments caused by insufficient resources or operational conflicts.

[0034] Table 1 Comparison of Dispatch Performance of Resource Debt-Based Closed-Loop Pesticide Spraying Drones

[0035] Table 1 is based on a spraying operation scenario within the same contiguous rice planting base. The test objects include farmland boundaries, spraying tasks, drone status, resupply nodes, no-spraying zones, wind speed and direction, and scheduling inputs with identical operation time windows. The resource debt closure rate in the table is the percentage of resource debt with closed records, calculated after uniformly converting pesticide supply, return-to-base power, drift risk, operation time window, resupply queue, and airspace conflicts into resource debt. As shown in Table 1, the genetic algorithm for small-plot sequential scheduling has a low average scheduling computation time of 6.8 seconds, but its effective spraying task unit completion rate is 88.7%, and its resource debt closure rate is 80.9%, with 8.6 unclosed resource debts per day. This indicates that it is more suitable for quickly arranging plot sequences, but its ability to handle the linkage between pesticide resupply, power swapping, drift avoidance, resupply queue, and airspace conflicts is insufficient, easily leading to resupply waiting and manual adjustments during on-site execution.

[0036] OR-Tools CP-SAT task scheduling demonstrates strong stability in constraint solving, achieving an effective spraying task unit completion rate of 91.2% and a resource debt closure rate of 84.6%, outperforming the genetic algorithm's small-plot sequential scheduling. However, the average scheduling computation time reaches 18.6 seconds, and the number of unclosed resource debts remains at 6.9 per day. This is because while the method handles time windows, task allocation, and resource constraints well, it typically addresses tasks that cannot be directly executed by using constraint satisfaction or constraint conflict methods. It lacks a commitment-settlement coupling edge between task commitment anchors and settlement actions, and cannot retain subsequently callable replenishment actions, battery swapping and return actions, drift avoidance actions, aircraft takeover actions, and time-segment migration actions during the task generation phase.

[0037] DQN (Drone-Quality Noise) drone task allocation and MAPPO (Multi-Drone Collaborative Scheduling) can improve task allocation among multiple drones through learning strategies. MAPPO multi-Drone collaborative scheduling achieved an effective spraying task unit completion rate of 92.5%, a resource debt closure rate of 86.8%, and reduced the number of temporary manual adjustments to 4.9 times per day, demonstrating the advantages of multi-agent collaborative methods in task assignment and multi-drone coordination. However, these two methods typically treat insufficient pesticide, insufficient power, drift risk, and resupply queues as rewards / penalties or state constraints, failing to explicitly create a resource debt ledger or assign a set of primary and backup settlement actions for each resource debt. Therefore, when multiple resource problems overlap on-site, unclosed resource debts and resupply waiting still occur.

[0038] The original GFlowNet combinatorial scheduling outperforms genetic algorithms, DQN, and MAPPO in generating diverse candidate scheduling chains. Its effective spraying task unit completion rate is 93.1%, resource debt closure rate is 88.2%, and average job completion time is 11.3 min / ha, indicating that generative flow networks are suitable for handling the spraying task combinatorial scheduling problem. However, the original GFlowNet combinatorial scheduling mainly samples based on the generation probability and termination reward of candidate scheduling chains, failing to explicitly separate the generation trajectory into task commitment segments, resource debt segments, and debt settlement segments. It also fails to include unsettled resource debt records, non-closable task commitment anchor candidates, and unmatched debt closure records in the debt closure loss, thus still resulting in 5.1 unclosed resource debts per day.

[0039] The present invention achieves an effective spraying task unit completion rate of 95.4%, a resource debt closure rate of 92.7%, reduces the number of unclosed resource debts to 3.2 per day, reduces the average resupply waiting time to 7.6 minutes per sortie, and reduces the number of temporary manual adjustments to 2.8 times per day. Its performance improvement mainly comes from three aspects: First, by coupling the task commitment layer and the resource settlement layer to form a two-layer generative graph, the task commitment anchor candidates retain the range of callable settlement actions before entering the debt flow GFlowNet scheduling algorithm; second, by using the resource debt ledger to uniformly write pesticide supply, return-to-base power, drift risk, operation time window, resupply queuing, and airspace conflicts into traceable resource debt records, and by using debt-settlement mapping records to determine the main settlement action and backup settlement action set; third, by using the debt settlement flow and debt closure record table to perform closure verification on the settlement results of resupply actions, battery swapping and return-to-base actions, drift avoidance actions, aircraft takeover actions, and time period migration actions, ensuring that only commitment chain segments with closed resource debts enter the debt closure candidate spraying scheduling chain set.

[0040] As shown in Table 1, compared with genetic algorithm small plot sequential scheduling, OR-Tools CP-SAT task scheduling, DQN UAV task allocation, MAPPO multi-UAV collaborative scheduling, and the original GFlowNet combined scheduling, the present invention performs better in terms of effective spraying task unit completion rate, resource debt closure rate, number of unclosed resource debts, average resupply waiting time, number of temporary manual adjustments, drift risk handling success rate, and average operation completion time. Its advantage does not come from simply increasing the scheduling search intensity, but from forming a complete link from task commitment to resource replenishment, avoidance, UAV swapping, time period migration, and debt closure verification through task commitment anchors, resource debt ledgers, debt settlement flow, and debt closure record tables. This makes the pesticide spraying UAV scheduling scheme more executable, stable, and traceable in complex farmland operation environments.

[0041] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for scheduling pesticide spraying drones based on deep learning, characterized in that, Includes the following steps: Generate a set of spraying operation statuses based on farmland boundaries, spraying tasks, drone status, resupply nodes, no-spraying zones, wind speed and direction, and operation time windows; Based on the operational relationships between spraying tasks, drones, and supply nodes, a task commitment layer and a resource settlement layer are constructed. The task commitment layer forms a candidate set of task commitment anchors, and the resource settlement layer forms a set of settlement actions. The task commitment layer and the resource settlement layer are coupled to obtain a two-layer generative graph. The two-layer generative graph is input into the debt flow GFlowNet scheduling algorithm. The commitment generation unit reads the task commitment anchor candidates one by one and binds the spraying task with the commitment drone or candidate drone set to form the task commitment anchor set. The debt identification unit reads the task commitment anchor set and the spraying operation status set, identifies task commitment anchors that cannot be directly executed, and writes them into the resource debt ledger. The debt settlement unit selects the corresponding settlement action from the settlement action set according to the debt type in the resource debt ledger, generates a debt settlement flow, and forms a debt closure record table after the debt settlement flow completes the replenishment, avoidance, waiting, machine replacement or time period adjustment of the corresponding resource debt; The task commitment anchor set, resource debt ledger, and debt closure record table are concatenated into a debt closure candidate spraying scheduling chain set. The target spraying scheduling chain is determined from the debt closure candidate spraying scheduling chain set where all resource debts have been closed, and a pesticide spraying drone scheduling plan is generated.

2. The method for scheduling pesticide spraying drones based on deep learning according to claim 1, characterized in that, The generation of the spraying operation status set specifically includes: Within the farmland boundary, the corresponding work area for the spraying task is determined. The work area is divided into strips according to the executable spray width in the drone status and bound to the corresponding work time window to form a set of spraying task units. The system reads the operating area, required amount of pesticide, operating time window, and adjacent operating strips of each spraying task unit according to the spraying task unit set, and forms the task debt identification status. Record the current position, remaining battery power, remaining liquid volume, executable spray width, and return point of each drone according to its status, and form a set of drone operation status records; A set of supply node states is formed based on the location of the supply node, its replenishment capacity, its battery swapping capacity, and its queuing time. Based on the positional relationship between the no-spraying area, wind speed and direction, and the spraying task unit set, a drift risk status set is formed, and an initial flight path occupancy status is formed based on the current position of the drone and the occupied flight path segment; The spraying task unit set, task debt identification status, UAV operation status set, supply node status set, drift risk status set, and initial flight path occupancy status are encapsulated into a spraying operation status set.

3. The method for scheduling pesticide spraying drones based on deep learning according to claim 1, characterized in that, The generation of the two-layer generated graph specifically includes: The spraying task units in the spraying task unit set are used as task nodes, and the drones in the drone operation status set are used as drone nodes. The task commitment layer is constructed according to the correspondence between task nodes and drone nodes. According to the task commitment layer, the task node, the drone node and the task commitment candidate edge are encapsulated into task commitment anchor candidates, and then summarized to form a task commitment anchor candidate set. A resource settlement layer is constructed using the supply node status set, drift risk status set, UAV operation status set, operation time window, and initial flight path occupancy status as the source of actions; Settlement action sets are formed according to the resource settlement conditions in the resource settlement layer. The settlement action sets include replenishment action, battery swapping and return action, drift avoidance action, engine swapping and takeover action, and time period migration action. Match the task commitment anchor candidates with the action records in the settlement action set one by one, generate commitment-settlement coupling edges according to the matching results, and form the correspondence between task commitment anchor candidates and callable settlement actions; The task commitment layer, resource settlement layer, task commitment anchor candidate set, settlement action set, and commitment-settlement coupling edge are encapsulated into a two-layer generative graph.

4. The method for scheduling pesticide spraying drones based on deep learning according to claim 1, characterized in that, The generation of the task commitment anchor set specifically includes: The commitment generation unit of the debt flow GFlowNet scheduling algorithm reads the task commitment anchor candidate set and commitment-settlement coupling edge in the two-layer generation graph to establish the commitment generation state; The commitment generation unit extracts the range of callable settlement actions for each task commitment anchor candidate based on the settlement entry field and commitment-settlement coupling edge in the commitment generation status; The forward commitment flow allocation unit reads the commitment generation status and the scope of callable settlement actions to generate a forward commitment flow record, including the commitment flow type and the reason for debt retention; The types of commitment flows include direct commitments, debt retention, and non-closable types; The commitment sorting unit generates the commitment sorting result according to the forward commitment flow record corresponding to the same spraying task unit; Based on the commitment ranking results, the committed drones and the candidate drone set are bound together to form a mission commitment anchor; The task commitment anchors corresponding to each spraying task unit are summarized to form a task commitment anchor set.

5. The method for scheduling pesticide spraying drones based on deep learning according to claim 4, characterized in that, The debt flow GFlowNet scheduling algorithm includes a commitment generation unit, a forward commitment flow allocation unit, a commitment sorting unit, a debt identification unit, and a debt settlement unit; The commitment generation unit reads the task commitment anchor candidates and their corresponding commitment-settlement coupling edges in the two-layer generation graph, forms a commitment relationship between each spraying task unit and its corresponding UAV, and writes the range of callable settlement actions of the task commitment anchor candidate into the commitment relationship. The forward commitment flow allocation unit reads the remaining liquid volume, required liquid volume, remaining power, return power requirement, candidate execution time period, and range of callable settlement actions in the commitment relationship, and writes the commitment flow type and debt retention reason into the forward commitment flow record. The commitment flow type includes direct commitment type, debt retention type, and non-closable type. The commitment sorting unit determines the commitment UAV and candidate UAV set in the order of priority of direct commitment type over debt retention type, priority of candidate execution time period, and priority of larger remaining power under the same spraying task unit. The debt identification unit converts the task commitment anchor of debt retention type into resource debt in the resource debt ledger. The debt settlement unit selects the settlement action along the commitment-settlement coupling edge corresponding to the resource debt and writes it into the debt closure record table after the resource debt is settled.

6. The method for scheduling pesticide spraying drones based on deep learning according to claim 1, characterized in that, The generation of the resource debt ledger specifically includes: The debt identification unit reads the task commitment anchor set, the spraying operation status set, and the commitment-settlement coupling edge in the two-layer generation graph to establish a debt identification base table; The debt identification unit performs resource debt identification according to the debt identification base table to obtain the resource debt identification results; Based on the resource debt identification results, the resource debt is written into the resource debt ledger; Configure debt-settlement mapping records according to the debt type and corresponding commitment-settlement coupling edge in the resource debt record, including a set of primary settlement actions and standby settlement actions; Resource debt records and debt-settlement mapping records are encapsulated into a resource debt ledger.

7. A method for scheduling pesticide spraying drones based on deep learning according to claim 6, characterized in that, The configuration of the debt-settlement mapping record determines the main settlement action and the set of backup settlement actions from the scope of callable settlement actions according to the settlement priority sequence corresponding to the debt type. The settlement action that is ranked first in the settlement priority sequence and exists in the scope of callable settlement actions is determined as the main settlement action. The remaining settlement actions that exist in the scope of callable settlement actions and belong to the same debt type settlement priority sequence are written into the set of backup settlement actions. Settlement actions that are not written into the scope of callable settlement actions do not enter the main settlement action and backup settlement action sets.

8. The method for scheduling pesticide spraying drones based on deep learning according to claim 1, characterized in that, The generation of the debt closure record table specifically includes: The debt settlement unit reads resource debt records and debt-settlement mapping records from the resource debt ledger to form a sequence of resource debts to be settled. Based on the sequence of resources and liabilities to be settled, execute the main settlement action along the corresponding commitment-settlement coupling edge to generate the main settlement flow record; Perform debt closure verification based on the master settlement flow record and generate master settlement verification results; When the main settlement verification result is that the main settlement is not closed, the debt settlement unit generates a standby settlement flow record according to the standby settlement action set; A debt settlement flow is generated based on the main settlement flow record, the main settlement verification result, and the standby settlement flow record; Write the records in the debt settlement flow that are in a closed state into the debt closure record table.

9. A method for scheduling pesticide spraying drones based on deep learning according to claim 1, characterized in that, The generation of the pesticide spraying drone scheduling scheme specifically includes: Establish a commitment chain skeleton based on the set of task commitment anchors, and use each task commitment anchor as a commitment node in the commitment chain skeleton; Based on the correspondence between commitment nodes and resource debt records, the resource debt ledger is written into the commitment chain skeleton to form a commitment debt corresponding chain; Based on the correspondence between resource debt records and debt closure record tables, debt closure verification is performed on the corresponding chain of committed debts to form a closed commitment chain segment; Construct a set of candidate spraying scheduling chains for closed-loop commitment chains based on the operational time sequence between segments of the closed commitment chain and the ownership relationship of the committed drones. Determine the target spraying schedule chain from the set of candidate spraying schedule chains with closed debt; A pesticide spraying drone scheduling scheme is generated based on the target spraying scheduling chain.

Images