A multi-unmanned aerial vehicle task allocation and path planning method based on machine learning

Abstract

The application discloses a kind of multi-unmanned aerial vehicle task allocation and path planning method based on machine learning, comprising the following steps: collecting cooperative operation data, forming multi-source fusion features;Perform reachability screening, construct energy time limit reachable domain;Divide node type and establish hyperedge, construct task track dynamic hypergraph;Improved dynamic hypergraph neural network model is input, and task matching score and track evaluation result are output;Parameter optimization is carried out based on frost ice optimization algorithm, and improved dynamic hypergraph neural network model is updated;According to model output result, candidate execution scheme is constructed;Candidate execution scheme is executed joint optimization and constraint repair;Multi-unmanned aerial vehicle task allocation and path planning result is output.The application introduces task track dynamic hypergraph, improved dynamic hypergraph neural network model and frost ice optimization mechanism, improves multi-unmanned aerial vehicle cooperative task allocation accuracy, path planning safety and complex constraint adaptability.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for unmanned aerial vehicles (UAVs), and in particular to a multi-UAV task allocation and path planning method based on machine learning. Background Technology

[0002] With the development of intelligent control and autonomous collaborative technologies for unmanned aerial vehicles (UAVs), multi-UAV systems have been widely applied in scenarios such as regional inspection, disaster search and rescue, logistics delivery, environmental monitoring, and target reconnaissance. Existing multi-UAV collaborative operations typically require establishing correspondences between multiple UAVs and multiple task targets, and planning flight paths for each UAV. Traditional methods often determine task assignment based on fixed indicators such as the distance between the UAV and the task point, remaining battery power, and task priority, and use algorithms such as A*Sat algorithm, genetic algorithm, particle swarm optimization, or ant colony optimization to generate flight tracks. While these methods can yield usable results in simple scenarios, they struggle to meet the needs of multi-UAV collaborative operations when there are many tasks, complex airspace constraints, and dynamically changing environments.

[0003] With the development of machine learning and intelligent optimization technologies, existing methods have begun to utilize neural networks, reinforcement learning, or graph learning models for task allocation and path planning. Current methods typically use UAV state, mission requirements, and environmental constraints as general feature inputs to the model, lacking a unified expression for the high-order relationships between UAVs, missions, tracks, energy states, time windows, and airspace constraints. The executability of multi-UAV missions depends not only on the state of individual UAVs but also on the combined influence of mission time limits, candidate tracks, obstacle distribution, no-fly zone boundaries, communication coverage, and the occupancy of tracks by other UAVs. Ordinary pairwise relationship modeling methods are insufficient to accurately describe this type of multi-faceted coupling.

[0004] Existing technologies generally handle task allocation and path planning separately, that is, first determining the task assignment, and then planning the trajectory separately for the assigned task. This can easily lead to a lack of linkage between the task allocation result and path cost, energy consumption constraints, obstacle avoidance requirements, and safety intervals, resulting in problems such as excessively long trajectories, excessive energy consumption, task timeouts, or multi-aircraft trajectory conflicts. When a new task is added, the UAV's battery power decreases, communication is abnormal, or the airspace environment changes, traditional solutions usually require a global recalculation, and dynamic replanning is inefficient.

[0005] Therefore, how to provide a machine learning-based method for multi-UAV task allocation and path planning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a multi-UAV task allocation and path planning method based on machine learning. This invention fully utilizes dynamic hypergraph neural networks, frost-ice optimization algorithms, and UAV intelligent control technology to perform multi-source fusion modeling of UAV status, task requirements, and environmental constraints. By improving the dynamic hypergraph neural network, it generates task matching scores and trajectory evaluation results, and combines the constraint-guided frost-ice optimization algorithm to achieve joint optimization of task assignment, task order, trajectory nodes, and arrival time slots. It has the advantages of high intelligence in task matching, strong collaborative path planning, good adaptability to energy consumption and time constraints, strong collaborative obstacle avoidance capability, and high efficiency in dynamic replanning.

[0007] A multi-UAV task allocation and path planning method based on machine learning according to an embodiment of the present invention includes: Collect data on collaborative operations of multiple UAVs, and standardize, spatiotemporally align, and encode the data to form multi-source fusion features; The flight capability, mission constraints, and airspace constraints in the multi-source fusion features are screened for accessibility to form an energy-time-limited reachable domain. Based on the multi-source fusion characteristics and energy time-limited reachable domain, node types are divided, hyperedges are established and weights are initialized to construct a dynamic hypergraph of the mission trajectory. Input the dynamic hypergraph of the mission trajectory into the improved dynamic hypergraph neural network model, perform node encoding, reachability-gated hyperedge update, mission trajectory feedback convolution and conflict shadow prediction, and output mission matching score and trajectory evaluation results. For the node encoding dimension, hypergraph convolution layer number, feature fusion weight, hyperedge weight and decision threshold in the improved dynamic hypergraph neural network model, the frost optimization algorithm with energy-time coupling metric function and conflict shadow penalty factor is used for iterative optimization. The fitness is constructed with model error and cooperative execution cost, and the target parameter combination is generated and the improved dynamic hypergraph neural network model is updated. Based on the updated model output, construct candidate execution schemes that include task affiliation, task order, track nodes, and arrival time slots; Joint optimization and constraint repair are performed on candidate execution schemes to generate multi-UAV task allocation and path planning results that meet constraints of energy, time limit, communication, obstacle avoidance and safety interval.

[0008] Optionally, the formation of multi-source fusion features includes: Collect drone status data, mission requirement data, and environmental constraint data; The flight capability information is obtained by extracting location status, battery status, payload status, perception status, communication status, and mission load status from the UAV status data. The mission airspace constraint information is obtained by extracting mission location, mission attributes, mission time limit, payload requirements, sensor requirements, obstacle areas, no-fly zone boundaries, communication coverage areas, candidate waypoints, and safety intervals from mission requirement data and environmental constraint data. Missing value filling, coordinate system unification, and encoding are performed on the flight capability information and mission airspace constraint information. The data is then spliced ​​together according to the UAV number, mission number, and time window to form multi-source fusion features.

[0009] Optionally, the reachable domain of the formation energy time limit includes: Extract flight capability information, mission constraint information, and airspace constraint information from multi-source fusion features; Calculate the estimated arrival time and estimated mission completion time based on the drone's current location, mission location, flight speed, mission duration, and mission time limit. Mark combinations where the estimated mission completion time does not exceed the mission time limit as time-reachable. The energy margin is calculated based on the drone's remaining battery power, estimated flight energy consumption, mission execution energy consumption, and return-to-home reserve power. Combinations that meet the safe return-to-home power requirements are marked as energy-available. Capability matching and airspace selection are performed based on payload capacity, sensor type, payload requirements, sensor requirements, obstacle areas, no-fly zone boundaries, communication coverage areas, and safety intervals. The flight paths of UAV missions that simultaneously meet the requirements of time reachability, energy reachability, capability matching, and airspace selection are written into the energy-time reachability domain, and the energy margin, time margin, and reachability marker are recorded.

[0010] Optionally, the construction of the mission trajectory dynamic hypergraph includes: Based on the characteristics of multi-source fusion and the reachability of energy time limits, node types are divided into UAV nodes, mission nodes, track nodes, environment nodes, time window nodes, and energy status nodes. Write the UAV status characteristics, mission constraint characteristics, candidate waypoint characteristics, airspace constraint characteristics, mission execution time period, and energy status characteristics into the corresponding nodes; Based on reachability markers, candidate tracks, airspace constraints, expected arrival time slots, and mission execution constraints, we establish UAV-mission-energy hyperedges, mission-track-environment hyperedges, UAV-track-time window hyperedges, multi-UAV-track conflict hyperedges, and mission-priority-execution time limit hyperedges. The weights of various hyperedges are initialized based on reachability, energy margin, time margin, track cost, airspace risk, and safety interval margin to form a dynamic hypergraph of the mission track.

[0011] Optionally, the output task matching score and track evaluation results include: An improved dynamic hypergraph neural network model is constructed, which includes a multi-source node encoding layer, a reachability-gated hyperedge update layer, a mission trajectory feedback convolutional layer, and a joint scoring output layer. The multi-source node coding layer introduces a node type scaling factor. The reciprocal of the square root of the proportion of node types appearing in the mission track dynamic hypergraph is calculated to obtain the node type scaling factor. All node coding features are multiplied by the corresponding node type scaling factor and then written into the improved dynamic hypergraph neural network model. The reachable domain gated superedge update layer introduces a gated temperature coefficient. The gated temperature coefficient is obtained by logarithmically scaling the sum of the energy margin percentage and the time margin percentage. The gated temperature coefficient is used to adjust the weight of reachable combination superedges and reset the weight of unreachable combination superedges to zero. The task trajectory feeds back to the convolutional layer to introduce a risk decay coefficient. Vertex-to-superedge feature aggregation is performed to obtain temporary superedge features. Superedge-to-vertex feature backpropagation is performed to generate task matching intermediate features and trajectory cost intermediate features. The average value of the trajectory cost intermediate features of the current batch is obtained by exponential decay mapping to obtain the risk decay coefficient. The trajectory cost intermediate features are scaled according to the risk decay coefficient and embedded into the task matching intermediate features. Based on the trajectory node sequence, UAV flight speed, arrival time slot sequence and safety interval radius, temporal position inference is performed to construct the three-dimensional occupancy volume of future multi-time windows. Then, the occupancy volume is expanded in time dimension to generate a four-dimensional occupancy block. The four-dimensional occupancy block is written as a conflict shadow into the multi-UAV trajectory conflict superedge. The joint scoring output layer introduces a conflict penalty coefficient, which is obtained by linear combination of the spatial overlap rate of conflict shadows within the same time window and a fixed weight. The conflict penalty coefficient is used to perform risk weighting on the joint representation of the mission trajectory, and outputs the mission matching score as well as trajectory cost, energy consumption risk and conflict risk scores. The improved dynamic hypergraph neural network model is trained by constructing a training sample set. The training objective is the weighted sum of task matching classification loss, trajectory cost regression loss, energy consumption risk regression loss, and conflict risk classification loss. Backpropagation is used to simultaneously update the multi-source node encoding layer, reachability-gated hyperedge update layer, task trajectory feedback convolutional layer, joint scoring output layer, as well as node type scaling coefficient, gating temperature coefficient, risk decay coefficient, and conflict penalty coefficient.

[0012] Optionally, generating the target parameter combination and updating the improved dynamic hypergraph neural network model includes: The set of parameters to be optimized is determined, including the number of node encoding dimensions, the number of hypergraph convolutional layers, feature fusion weights, UAV mission energy hyperedge weights, mission trajectory environment hyperedge weights, UAV trajectory time window hyperedge weights, and multi-UAV trajectory conflict hyperedge weights. The decision thresholds include mission matching thresholds, trajectory risk thresholds, and replanning trigger thresholds. An energy-time coupling metric function is constructed, mapping the percentage of energy margin and the percentage of time margin to the same numerical interval. The energy-time coupling metric is obtained by weighted fusion according to the energy-sensitive weight and the time-sensitive weight. The energy-time coupling metric is divided into high-constraint interval, medium-constraint interval and low-constraint interval according to the energy-time coupling metric value. The set of parameters to be optimized is sampled hierarchically, and the sampling results are encoded as frost individuals to generate the initial population. The parameter combination corresponding to the individual frost ice is written into the improved dynamic hypergraph neural network model. The model error and collaborative execution cost are calculated. The conflict shadow penalty factor is constructed based on the spatial overlap rate, the degree of security interval intrusion and the proportion of conflict duration in the conflict shadow. The conflict shadow penalty factor is weighted and fused with the multi-UAV trajectory conflict hyperedge weight and the conflict risk score to obtain the conflict shadow penalty term. The fitness value is formed by weighted summation of model error, cooperative execution cost and conflict shadow penalty term. The initial population is then updated by frost search according to the fitness value, and the set of parameters to be optimized and the decision threshold are iteratively adjusted. When a frost-ice individual violates the energy margin limit, time margin limit, communication coverage limit, no-fly zone boundary limit, obstacle safety boundary limit, or safety interval limit, the infeasible solution repair operator is invoked to repair the parameters and recalculate the fitness value. When the number of iterations reaches the maximum number of iterations or the fitness difference between two consecutive generations is lower than the convergence threshold, the frost individual with the smallest fitness value is selected as the target parameter combination, and the dynamic hypergraph neural network model is updated and improved with the target parameter combination.

[0013] Optionally, the construction of candidate execution schemes, including task attribution, task order, track nodes, and arrival time slots, includes: The drone task combinations are filtered according to the task matching threshold to form a candidate task affiliation set; The tasks are sorted according to their priority, time limit, and duration to form a candidate task order set; Candidate waypoints are selected based on track cost score, energy consumption risk score and conflict risk score to form a set of candidate track nodes; Arrival time windows are divided based on the UAV's flight speed, mission execution time, and candidate track node set to form a candidate arrival time slot set; The candidate task affiliation set, candidate task order set, candidate track node set, and candidate arrival time slot set are encoded and combined, and schemes with missing tasks, duplicate tasks, track breaks, and time slot conflicts exceeding a preset threshold are eliminated to form candidate execution schemes.

[0014] Optionally, the joint optimization and constraint repair of candidate execution schemes to generate multi-UAV task allocation and path planning results that satisfy energy, time limit, communication, obstacle avoidance, and safety interval constraints includes: Read the candidate execution plans and calculate the comprehensive evaluation value based on the task completion rate, trajectory cost, energy consumption risk, conflict risk, and task load balance. The candidate execution plans are ranked according to the comprehensive evaluation value, and the candidate execution plans that meet the evaluation conditions are selected as the execution plans to be repaired. The repair execution plan includes supplementing unassigned tasks, deleting duplicate tasks, transferring low-power tasks, detouring in no-fly zones, detouring around obstacles, and adjusting conflict time slots. The corrected execution plan was determined as the result of multi-UAV task allocation and path planning.

[0015] The beneficial effects of this invention are: This invention introduces an improved dynamic hypergraph neural network model into the multi-UAV task allocation and path planning process, achieving unified modeling and intelligent evaluation of UAV status, task requirements, and environmental constraints. By constructing multi-source fusion features and energy-time-limited reachability domains, this invention can perform reachability screening of UAV flight capabilities, task constraints, and airspace constraints before task allocation, avoiding assigning tasks to UAVs with insufficient power, unmet time limits, mismatched payloads, or unreachable paths, thus improving the executability and accuracy of task matching results.

[0016] This invention constructs a dynamic hypergraph of mission tracks, incorporating UAVs, missions, tracks, environment, time windows, and energy states into the same dynamic hypergraph structure. Through reachability-gated hyperedge updates and mission track-feedback convolutions, mission matching scores are simultaneously adjusted by track cost, energy consumption risk, airspace risk, and cooperative conflict risk. Compared to traditional methods that separate mission allocation from path planning, this invention considers path planning constraints concurrently during the mission allocation phase, reducing flight path cost, mission timeout risk, and energy waste, and improving the consistency between mission allocation and path planning results.

[0017] This invention uses a conflict shadow prediction mechanism to pre-assess the airspace overlap status of multiple UAVs within a future time window. It then combines this with a constraint-guided frost-ice optimization algorithm to jointly optimize and correct constraints on model parameters, hyperedge weights, decision thresholds, and candidate execution schemes, generating task allocation and path planning results that satisfy constraints related to energy, time limits, communication, obstacle avoidance, and safety intervals. When new tasks are added, UAV battery power decreases, communication fails, or the airspace environment changes, this invention can quickly adjust task assignment, task order, track nodes, and arrival time slots based on the updated model output, improving the collaborative obstacle avoidance capability, dynamic replanning efficiency, and overall operational stability of multi-UAV systems. 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 multi-UAV task allocation and path planning method based on machine learning proposed in this invention; Figure 2 This is a schematic diagram of the dynamic hypergraph structure of the task trajectory of the multi-UAV task allocation and path planning method based on machine learning proposed in this invention. Figure 3 This is a schematic diagram of the improved dynamic hypergraph neural network model structure for a multi-UAV task allocation and path planning method based on machine 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 Figure 1 , Figure 2 and Figure 3 A multi-UAV task allocation and path planning method based on machine learning, comprising: Collect data on collaborative operations of multiple UAVs, and standardize, spatiotemporally align, and encode the data to form multi-source fusion features; The flight capability, mission constraints, and airspace constraints in the multi-source fusion features are screened for accessibility to form an energy-time-limited reachable domain. Based on the multi-source fusion characteristics and energy time-limited reachable domain, node types are divided, hyperedges are established and weights are initialized to construct a dynamic hypergraph of the mission trajectory. Input the dynamic hypergraph of the mission trajectory into the improved dynamic hypergraph neural network model, perform node encoding, reachability-gated hyperedge update, mission trajectory feedback convolution and conflict shadow prediction, and output mission matching score and trajectory evaluation results. For the node encoding dimension, hypergraph convolution layer number, feature fusion weight, hyperedge weight and decision threshold in the improved dynamic hypergraph neural network model, the frost optimization algorithm with energy-time coupling metric function and conflict shadow penalty factor is used for iterative optimization. The fitness is constructed with model error and cooperative execution cost, and the target parameter combination is generated and the improved dynamic hypergraph neural network model is updated. Based on the updated model output, construct candidate execution schemes that include task affiliation, task order, track nodes, and arrival time slots; Joint optimization and constraint repair are performed on candidate execution schemes to generate multi-UAV task allocation and path planning results that meet constraints of energy, time limit, communication, obstacle avoidance and safety interval.

[0021] In this embodiment, the formation of multi-source fusion features includes: Collect drone status data, mission requirement data, and environmental constraint data; The flight capability information is obtained by extracting location status, battery status, payload status, perception status, communication status, and mission load status from the UAV status data. The mission airspace constraint information is obtained by extracting mission location, mission attributes, mission time limit, payload requirements, sensor requirements, obstacle areas, no-fly zone boundaries, communication coverage areas, candidate waypoints, and safety intervals from mission requirement data and environmental constraint data. Missing value filling, coordinate system unification, and encoding are performed on the flight capability information and mission airspace constraint information. The data is then spliced ​​together according to the UAV number, mission number, and time window to form multi-source fusion features.

[0022] In this embodiment, the reachable domain of the energy formation time limit includes: Extract flight capability information, mission constraint information, and airspace constraint information from multi-source fusion features; Calculate the estimated arrival time and estimated mission completion time based on the drone's current location, mission location, flight speed, mission duration, and mission time limit. Mark combinations where the estimated mission completion time does not exceed the mission time limit as time-reachable. The energy margin is calculated based on the drone's remaining battery power, estimated flight energy consumption, mission execution energy consumption, and return-to-home reserve power. Combinations that meet the safe return-to-home power requirements are marked as energy-available. Capability matching and airspace selection are performed based on payload capacity, sensor type, payload requirements, sensor requirements, obstacle areas, no-fly zone boundaries, communication coverage areas, and safety intervals. Specifically, capability matching is performed based on payload capacity, sensor type, payload requirements, and sensor requirements. The system reads the UAV's payload capacity limit, loadable payload categories, sensor type set, sensor operating status, and sensor availability duration. It also reads the corresponding payload requirements, payload category requirements, sensor type requirements, data acquisition accuracy requirements, and mission duration to establish a capability matching table between the UAV and the mission. The system compares the payload capacity limit with the payload requirements, verifies the consistency between loadable payload categories and payload category requirements, verifies the coverage relationship between the sensor type set and sensor type requirements, and verifies the availability of sensors and their operating status and availability duration against the mission duration. If the payload capacity limit is not less than the payload requirements, the loadable payload categories meet the payload category requirements, the sensor type set covers the sensor type requirements, and the sensor availability duration is not less than the mission duration, the UAV-mission combination is marked as a capability-matched combination. If any verification result does not meet the corresponding requirements, the UAV-mission combination is marked as a capability-mismatched combination. Airspace selection is performed based on obstacle areas, no-fly zone boundaries, communication coverage areas, and safety intervals, specifically as follows: The system reads candidate waypoints, candidate flight segments, obstacle area boundaries, no-fly zone boundaries, communication coverage area boundaries, and safety interval thresholds. Based on the candidate waypoints and flight segments, it generates candidate track buffers. It performs intersection detection between the candidate track buffers and obstacle area boundaries, marking intersecting UAV mission track combinations as obstacle conflict combinations. It performs boundary crossing detection between the candidate track buffers and no-fly zone boundaries, marking UAV mission track combinations entering the no-fly zone boundary as no-fly conflict combinations. It performs coverage relationship detection between the candidate track buffers and communication coverage area boundaries, marking UAV mission track combinations that do not meet communication coverage requirements as communication unreachable combinations. It calculates the safety interval margin based on the minimum distance between different UAV candidate track buffers, marking UAV mission track combinations with a safety interval margin less than the safety interval threshold as insufficient safety interval combinations. The flight paths of UAV missions that simultaneously meet the requirements of time reachability, energy reachability, capability matching, and airspace selection are written into the energy-time reachability domain, and the energy margin, time margin, and reachability marker are recorded.

[0023] In this embodiment, constructing the dynamic hypergraph of the mission trajectory includes: Based on multi-source fusion characteristics and the energy-time-limited reachability domain, node types are classified into UAV nodes, mission nodes, track nodes, environment nodes, time window nodes, and energy status nodes. Specifically, the node types are classified according to multi-source fusion characteristics and the energy-time-limited reachability domain as follows: Read the UAV number, mission number, candidate waypoint number, environmental constraint area number, time window number, remaining power status, mission constraint information and airspace constraint information from the multi-source fusion features, and read the reachability marker, energy margin and time margin in the energy time limit reachability domain; A drone node is generated by merging its location status, battery status, payload status, sensor status, communication status, and mission load status according to its drone number. Task nodes are generated by merging task location, task attributes, task time limit, payload requirements, sensor requirements, and task priority according to task number; By merging candidate waypoints, candidate segments, track costs, and estimated arrival time slots according to the candidate waypoint numbers, track nodes are generated. Based on the environmental constraint area numbering, obstacle areas, no-fly zone boundaries, communication coverage areas, and airspace risk information are merged to generate environmental nodes; Merge the task start time slot, task end time slot, expected arrival time slot and executable period according to the time window number to generate time window nodes; Based on the drone number, time window number, and remaining power status, the remaining power, estimated flight energy consumption, mission execution energy consumption, return-to-home power reserve, and energy margin are merged to generate an energy status node; Write the reachability marker, energy margin, and time margin into the association attributes between the corresponding UAV node, mission node, track node, time window node, and energy status node to complete the node type classification. Write the UAV status characteristics, mission constraint characteristics, candidate waypoint characteristics, airspace constraint characteristics, mission execution time period, and energy status characteristics into the corresponding nodes; Based on reachability markers, candidate tracks, airspace constraints, expected arrival time slots, and mission execution constraints, we establish UAV-mission-energy hyperedges, mission-track-environment hyperedges, UAV-track-time window hyperedges, multi-UAV-track conflict hyperedges, and mission-priority-execution time limit hyperedges. Specifically, the establishment of UAV-mission-energy hyperedges, mission-track-environment hyperedges, UAV-track-time window hyperedges, multi-UAV-track conflict hyperedges, and mission-priority-execution time limit hyperedges is as follows: Read reachability markers, energy margin, time margin, candidate track numbers, airspace constraint markers, estimated arrival time slots, mission priority, and mission execution time limits; The drone nodes, mission nodes, and energy status nodes that meet the reachability requirements are combined into a drone-mission-energy superedge, and the energy margin, expected flight energy consumption, mission execution energy consumption, and return-to-home reserve power are written into the superedge attributes. The task node, candidate track node, and environment node are combined to form a task-track-environment hyperedge, and obstacle constraints, no-fly zone constraints, communication coverage constraints, and airspace risk level are written into the hyperedge attributes. The UAV node, track node, and time window node are combined into a UAV-track-time window hyperedge, and the expected arrival time slot, task execution period, track cost, and time margin are written into the hyperedge attributes. Multiple UAV nodes and track nodes that have overlapping time windows, intersecting track buffers, or insufficient safety intervals are combined into a multi-UAV-track conflict superedge, and the conflict location, conflict time slot, minimum spacing, and safety interval margin are written into the superedge attributes. The task node, task priority attribute, and execution time limit attribute are combined to form a task-priority-execution time limit hyperedge. The task priority, latest completion time, task duration, and time limit margin are written into the hyperedge attribute to complete the hyperedge establishment of the task track dynamic hypergraph. The weights of various hyperedges are initialized based on reachability, energy margin, time margin, track cost, airspace risk, and safety interval margin to form a dynamic hypergraph of the mission track. Specifically, the weight initialization of various hyperedges based on reachability, energy margin, time margin, track cost, airspace risk, and safety interval margin is as follows: Read the reachability state, energy margin, time margin, track cost, airspace risk level and safety interval margin corresponding to each hyperedge, convert the reachability state into reachability weight base value, convert the energy margin and time margin into positive constraint weights, and convert the track cost, airspace risk level and safety interval insufficiency into negative constraint weights. For the UAV-mission-energy superedge, the initial weights are determined according to the reachable weight base value, energy margin, and return-to-home reserved power. For the mission-track-environment hyperedge, the initial weights are determined according to track cost, airspace risk level, obstacle distance, and no-fly zone boundary distance. For the UAV-track-time window superedge, the initial weight is determined according to the expected arrival time slot, mission execution time limit, and time margin; For multi-UAV-track conflict superedges, the initial weights are determined according to the safety interval margin, track buffer overlap status, and conflict time slot overlap status. For the task-priority-execution time limit superedge, determine the initial weight according to the task priority, the latest completion time, and the time limit margin; The initial weights of various hyperedges are written into the corresponding hyperedge attribute fields, and the nodes, hyperedges, and hyperedge weights are associated and stored according to the time window to form a dynamic hypergraph of the mission track.

[0024] In this embodiment, the output task matching score and trajectory evaluation results include: An improved dynamic hypergraph neural network model is constructed, which includes a multi-source node encoding layer, a reachability-domain gated hyperedge update layer, a mission trajectory feedback convolutional layer, and a joint scoring output layer. Specifically, the construction of the improved dynamic hypergraph neural network model is as follows: A network structure is established consisting of a multi-source node encoding layer, a reachability-gated superedge update layer, a mission trajectory feedback convolutional layer, and a joint scoring output layer connected sequentially. In the multi-source node coding layer, node type identifiers and node attribute mapping units are set up to map the attribute features of UAV nodes, mission nodes, track nodes, environment nodes, time window nodes and energy state nodes into node coding features of a unified dimension. The node coding features output by the multi-source node coding layer are input into the reachability-gated hyperedge update layer. The reachability-gated hyperedge update layer is then connected to the reachability marker, energy margin, and time margin in the energy-limited reachability domain to update the hyperedge association strength and hyperedge weight in the mission trajectory dynamic hypergraph. The updated superedge features output from the reachability-gated superedge update layer are input into the task trajectory feedback convolutional layer. The node-to-superedge feature aggregation and superedge-to-node feature backpropagation are performed in the task trajectory feedback convolutional layer to generate the joint representation of the task trajectory. The conflict shadow representation is generated based on the trajectory node, time window node and safety interval information. The task trajectory fed back into the convolutional layer's output joint representation of the task trajectory and the conflict shadow representation are input into the joint scoring output layer. Risk weighting and multi-objective scoring are performed in the joint scoring output layer, and the task matching score, trajectory cost score, energy consumption risk score and conflict risk score are output to form an improved dynamic hypergraph neural network model. The multi-source node encoding layer introduces a node type scaling factor. This factor is calculated by taking the inverse square root of the percentage of node types appearing in the mission trajectory dynamic hypergraph. All node encoding features are multiplied by their corresponding scaling factors and then written into the improved dynamic hypergraph neural network model. The multi-source node encoding layer includes: Node category identification unit: Reads the node type identifier and node number in the mission track dynamic hypergraph, and classifies the nodes into UAV nodes, mission nodes, track nodes, environment nodes, time window nodes, and energy status nodes; Node attribute mapping unit: performs numerical normalization, state embedding, linear mapping and nonlinear activation on the attribute features of different types of nodes to generate initial encoding features of nodes with a unified dimension; Node type scaling unit: Calculates the proportion of each node type in the mission track dynamic hypergraph, calculates the square root of the proportion of the number of nodes, generates the node type scaling coefficient, and scales the initial encoding features of the corresponding nodes. Encoding output alignment unit: Arranges and aligns the scaled node encoding features according to node number, time window number and hyperedge association relationship, and generates a node encoding feature matrix for use by the reachability gated hyperedge update layer; The node category identification unit reads the node type identifiers from the task trajectory dynamic hypergraph, assigning UAV status features to UAV nodes, task constraint features to task nodes, candidate waypoint features to trajectory nodes, airspace constraint features to environment nodes, task execution time periods to time window nodes, and remaining battery power, energy margin, and return-to-home reserved battery power to energy status nodes. The node attribute mapping unit normalizes continuous attribute features, embeds and encodes discrete attribute features, and converts the attribute features of different types of nodes into initial node encoding features of a unified dimension through linear mapping and nonlinear activation. The node type scaling unit calculates the node quantity ratio according to the occurrence of UAV nodes, task nodes, trajectory nodes, environment nodes, time window nodes, and energy status nodes, and uses the reciprocal of the square root of the node quantity ratio as the node type scaling coefficient, so that node types with fewer occurrences receive higher weight compensation during the encoding stage. The encoding output alignment unit writes the scaled node encoding features into the node encoding feature matrix according to the node number and time window number, so that the node encoding features output by the multi-source node encoding layer maintain the differences in node types, the correspondence of time windows, and the hyperedge association relationship. The reachable domain gated hyperedge update layer introduces a gating temperature coefficient. The gating temperature coefficient is obtained by logarithmically scaling the sum of the energy margin percentage and the time margin percentage. This gating temperature coefficient is used to adjust the weights of reachable combination hyperedges and reset the weights of unreachable combination hyperedges to zero. The reachable domain gated hyperedge update layer includes: Reachability information access unit: Reads the reachability marker, energy margin, time margin, energy margin percentage and time margin percentage in the energy time limit reachability domain, and establishes a correspondence with the hyperedge number in the mission track dynamic hypergraph; Gated temperature calculation unit: sums up the percentage of energy margin and the percentage of time margin, performs logarithmic scaling and numerical range constraints to generate a gated temperature coefficient, which is used to characterize the achievable margin of UAV mission combinations under energy constraints and time constraints. Superedge weight gating unit: Adjusts the weight of reachable combination superedges according to the reachability flag and gating temperature coefficient, and resets the weight of unreachable combination superedges to zero, so that unreachable combinations do not participate in superedge feature transfer; Updated output unit: Outputs the updated hyperedge weight matrix and hyperedge correlation matrix according to the hyperedge number and time window number, which are used by the task trajectory to feed back into the convolutional layer; The reachability information access unit reads the reachability markers, energy margins, and time margins corresponding to the UAV-mission-energy hyperedge, mission-track-environment hyperedge, and UAV-track-time window hyperedge. It converts the energy margin into an energy margin percentage and the time margin into a time margin percentage. The gating temperature calculation unit sums the energy margin percentage and the time margin percentage, performs logarithmic scaling and range constraints on the summation result, and generates a gating temperature coefficient. This coefficient ensures that reachable combinations with higher energy and time margins have stronger hyperedge retention, while reachable combinations with lower energy and time margins have weaker hyperedge retention. The hyperedge weight gating unit writes the gating temperature coefficient into the weight update process of the corresponding reachable combination hyperedge, adjusts the weights of reachable combination hyperedges, and resets the weights of unreachable combination hyperedges to zero. The update result output unit inputs the gating-processed hyperedge weight matrix and hyperedge correlation matrix into the mission track feedback convolutional layer to restrict information propagation from unreachable hyperedges and enhance the feature representation of reachable hyperedges. The task trajectory feedback convolutional layer introduces a risk attenuation coefficient, performs vertex-to-hyperedge feature aggregation to obtain temporary hyperedge features, performs hyperedge-to-vertex feature backpropagation to generate intermediate features for task matching and intermediate features for trajectory cost, obtains the risk attenuation coefficient by applying an exponential decay mapping to the average value of the intermediate features for trajectory cost in the current batch, and embeds the intermediate features for trajectory cost into the intermediate features for task matching after scaling according to the risk attenuation coefficient. Based on the trajectory node sequence, UAV flight speed, arrival time slot sequence, and safety interval radius, it performs temporal position extrapolation to construct a three-dimensional occupancy volume for future multi-time windows, and then performs temporal dimension expansion on the occupancy volume to generate a four-dimensional occupancy block. The four-dimensional occupancy block is written as a conflict shadow into the multi-UAV trajectory conflict hyperedge. The task trajectory feedback convolutional layer includes: Vertex-to-hyperedge aggregation unit: Read the node encoding features belonging to the same hyperedge based on the hyperedge correlation matrix, and perform weighted aggregation on the node encoding features to generate temporary hyperedge features; Hyperedge-to-vertex backhaul unit: Based on the temporary features and weights of the hyperedge, it backhauls feature information to the associated nodes to generate intermediate features for task matching and intermediate features for trajectory cost. Risk attenuation embedding unit: Generates a risk attenuation coefficient based on the average value of intermediate features of current batch track cost, and embeds the intermediate features of task matching after scaling the intermediate features of track cost; Conflict Shadow Generation Unit: Performs temporal position deduction based on track node sequence, UAV flight speed, arrival time slot sequence and safety interval radius, generates four-dimensional duty blocks, and writes the four-dimensional duty blocks as conflict shadows into the multi-UAV track conflict superedge; The vertex-to-hyperedge aggregation unit reads the node encoding features output by the multi-source node encoding layer, the hyperedge correlation matrix output by the reachability-gated hyperedge update layer, and the hyperedge weight matrix. It then aggregates the encoding features of UAV nodes, task nodes, track nodes, environment nodes, time window nodes, and energy state nodes connected by the same hyperedge according to the hyperedge weights, obtaining temporary hyperedge features. The hyperedge-to-vertex backpropagation unit propagates these temporary hyperedge features back to the associated nodes, enabling UAV task reachability relationships, track-environment constraint relationships, time window occupancy relationships, and energy state relationships to participate in node feature updates, generating intermediate features for task matching and intermediate features for track cost. The risk attenuation embedding unit calculates the intermediate features of the current batch of track costs. The average value is calculated, and an exponential decay mapping is performed on the average value to obtain the risk decay coefficient. The intermediate features of the track cost are scaled according to the risk decay coefficient, and the scaled intermediate features of the track cost are embedded into the intermediate features of task matching, so that high-risk tracks have an inhibitory effect on task matching results. The conflict shadow generation unit calculates the position status of the UAV in the future multiple time windows based on the track node sequence, UAV flight speed and arrival time slot sequence, and expands it with the safety interval radius to form a three-dimensional occupancy volume. The three-dimensional occupancy volume is expanded in the time dimension to generate a four-dimensional occupancy block representing the spatiotemporal occupancy range of the UAV. The four-dimensional occupancy block is written into the multi-UAV track conflict superedge to form a conflict shadow for conflict risk calculation. The joint scoring output layer introduces a conflict penalty coefficient, which is obtained by linearly combining the spatial overlap rate of conflict shadows within the same time window with fixed weights. This conflict penalty coefficient is used to risk-weight the joint representation of the mission trajectory, outputting a mission matching score as well as trajectory cost, energy consumption risk, and conflict risk scores. The joint scoring output layer includes: Joint Representation Fusion Unit: Receives intermediate features of task matching, intermediate features of track cost, energy state features, and conflict shadow representation, and generates a joint representation of the task track; Conflict penalty calculation unit: Reads the spatial overlap rate of conflict shadows corresponding to different UAVs within the same time window, and performs linear combination according to fixed weights to obtain the conflict penalty coefficient; Risk-weighted unit: The conflict penalty coefficient is written into the weighting process of the joint representation of mission tracks, and risk enhancement processing is performed on the joint representation of mission tracks with spatial overlap, insufficient safety interval or high conflict risk. Scoring output unit: Based on the risk-weighted joint characterization of mission tracks, output mission matching score, track cost score, energy consumption risk score and conflict risk score; The joint representation fusion unit reads the intermediate features of task matching, intermediate features of track cost, and conflict shadow representations in the hyperedges of multi-UAV track conflicts from the output of the task track feed convolutional layer. It fuses the features of task executability, track passage cost, energy consumption status, and spatiotemporal conflict relationship to form a joint task track representation. The conflict penalty calculation unit performs overlap detection on the four-dimensional occupancy blocks corresponding to different UAVs within the same time window, calculates the spatial overlap rate of the conflict shadows, and linearly combines the spatial overlap rate with fixed weights to obtain the conflict penalty coefficient. The risk weighting unit applies the conflict penalty coefficient to the joint task track representation, so that UAV task track combinations with track spatial overlap or insufficient safety intervals receive higher risk weights. The scoring output unit performs multi-objective mapping on the risk-weighted joint task track representation to generate task matching score, track cost score, energy consumption risk score, and conflict risk score, so that the model output simultaneously reflects the task suitability, track execution cost, energy consumption risk, and multi-UAV conflict risk. The improved dynamic hypergraph neural network model is trained by constructing a training sample set. The training objective is a weighted sum of task matching classification loss, trajectory cost regression loss, energy consumption risk regression loss, and conflict risk classification loss. Backpropagation is used to simultaneously update the multi-source node encoding layer, reachability-gated hyperedge update layer, task trajectory feedback convolutional layer, joint scoring output layer, as well as node type scaling coefficient, gating temperature coefficient, risk decay coefficient, and conflict penalty coefficient. Specifically, the training of the improved dynamic hypergraph neural network model involves: Acquire historical multi-UAV collaborative operation records, simulation task execution records, and manually annotated task planning records to construct a training sample set. The training sample set includes multi-source fusion features, energy time-limit reachability domain, task trajectory dynamic hypergraph, task ownership annotation value, trajectory cost annotation value, energy consumption risk annotation value, and conflict risk annotation value. The training sample set is input into the improved dynamic hypergraph neural network model. After passing through the multi-source node encoding layer, the reachability-gated hyperedge update layer, the mission trajectory feedback convolutional layer, and the joint scoring output layer, the mission matching prediction value, trajectory cost prediction value, energy consumption risk prediction value, and conflict risk prediction value are obtained. The task matching classification loss is calculated based on the task matching prediction value and the task affiliation label value; the track cost regression loss is calculated based on the track cost prediction value and the track cost label value; the energy consumption risk regression loss is calculated based on the energy consumption risk prediction value and the energy consumption risk label value; and the conflict risk classification loss is calculated based on the conflict risk prediction value and the conflict risk label value. The comprehensive training loss is obtained by weighting and summing the task matching classification loss, trajectory cost regression loss, energy consumption risk regression loss and conflict risk classification loss according to their respective weights. Backpropagation is performed based on the comprehensive training loss to update the network parameters of the multi-source node encoding layer, reachability-gated superedge update layer, mission trajectory feedback convolutional layer, and joint scoring output layer, and the node type scaling factor, gating temperature factor, risk decay factor, and conflict penalty factor are updated simultaneously. When the comprehensive training loss meets the convergence condition or the training epochs reach the set number of epochs, the improved dynamic hypergraph neural network model is obtained after training is completed.

[0025] In this embodiment, generating the target parameter combination and updating the improved dynamic hypergraph neural network model includes: The set of parameters to be optimized is determined, including the number of node encoding dimensions, the number of hypergraph convolutional layers, feature fusion weights, UAV mission energy hyperedge weights, mission trajectory environment hyperedge weights, UAV trajectory time window hyperedge weights, and multi-UAV trajectory conflict hyperedge weights. The decision thresholds include mission matching thresholds, trajectory risk thresholds, and replanning trigger thresholds. An energy-time coupling metric function is constructed, mapping the energy margin percentage and time margin percentage to the same numerical interval. Weighted fusion is performed according to energy-sensitive and time-sensitive weights to obtain the energy-time coupling metric value. Based on this metric value, high-constraint, medium-constraint, and low-constraint intervals are defined. The set of parameters to be optimized is then stratified and sampled. The sampling results are encoded as frost individuals to generate the initial population. Specifically, the high-constraint, medium-constraint, and low-constraint intervals are defined according to the energy-time coupling metric value as follows: Read the energy-time coupling metric value corresponding to each UAV mission combination, and set the energy-time coupling metric value as the constraint strength value. The value range is from 0 to 1. The larger the value, the less sufficient the energy margin and time margin, and the stronger the task execution constraint. Set a first stratification threshold and a second stratification threshold, where the first stratification threshold is greater than the second stratification threshold; When the energy-time coupling metric is greater than or equal to the first stratification threshold, the corresponding UAV mission combination is classified into the high constraint interval. The high constraint interval indicates that at least one of the energy margin percentage and time margin percentage is close to the safety lower limit, or the weighted result of the two indicates that the mission execution urgency is relatively high. When the energy-time coupling metric is less than the first stratification threshold and greater than or equal to the second stratification threshold, the corresponding UAV mission combination is classified into the medium constraint interval. The medium constraint interval indicates that the energy margin percentage and time margin percentage are within the executable range and there is limited room for adjustment. When the energy-time coupling metric is less than the second stratification threshold, the corresponding UAV mission combination is classified into the low constraint interval. The low constraint interval indicates that both the energy margin percentage and the time margin percentage are in a generous state, and the mission execution constraints are relatively weak. The parameter sampling ratios are set separately for high-constraint, medium-constraint, and low-constraint intervals. The sampling density of trajectory risk threshold, replanning trigger threshold, and multi-UAV trajectory conflict super-edge weight is increased in the high-constraint interval. The sampling density of various parameters is balanced in the medium-constraint interval. The sampling density of task matching threshold and feature fusion weight is increased in the low-constraint interval, resulting in a hierarchical sampling result. The sampling results are encoded as frost individuals to generate the initial population, specifically: Read the hierarchical sampling results corresponding to the high-constraint, medium-constraint, and low-constraint intervals. Use the node encoding dimension number, hypergraph convolution layer number, feature fusion weight, UAV mission energy hyperedge weight, mission trajectory environment hyperedge weight, UAV trajectory time window hyperedge weight, multi-UAV trajectory conflict hyperedge weight, mission matching threshold, trajectory risk threshold, and replanning trigger threshold as the encoding fields for the frost individual. Use integer encoding for the node encoding dimension number and hypergraph convolution layer number, and real number encoding for the feature fusion weight, various hyperedge weights, and decision thresholds. Limit the boundaries of each encoding field according to a preset value range. Combine the encoding fields corresponding to the same hierarchical sampling result in a fixed order to form a frost individual. Generate multiple frost individuals according to the sampling ratio of the high-constraint, medium-constraint, and low-constraint intervals, and perform deduplication on duplicate frost individuals. Write the deduplicated frost individuals into the initial population set, and establish an individual number, interval source label, and fitness field to be calculated for each frost individual to generate the initial population for the frost optimization algorithm. The parameter combinations corresponding to individual frost ice objects are written into an improved dynamic hypergraph neural network model. Model error and collaborative execution cost are calculated. A conflict shadow penalty factor is constructed based on the spatial overlap rate, safety interval intrusion degree, and conflict duration proportion within the conflict shadow. This conflict shadow penalty factor is then weighted and fused with the multi-UAV trajectory conflict hyperedge weights and conflict risk scores to obtain the conflict shadow penalty term. Specifically, the conflict shadow penalty factor, constructed based on the spatial overlap rate, safety interval intrusion degree, and conflict duration proportion within the conflict shadow, is as follows: Read the four-dimensional occupied blocks corresponding to different UAVs within the same time window in the multi-UAV trajectory conflict hyperedge, and calculate the spatial overlap rate by comparing the spatial overlap volume of the four-dimensional occupied blocks with the baseline value of the occupied volume. Read the minimum track spacing of different drones within the same time window, and convert the insufficient amount between the safety interval radius and the minimum track spacing into the safety interval intrusion degree. When the safety interval meets the requirements, the safety interval intrusion degree is recorded as zero. The number of time windows with spatial overlap or insufficient safety intervals is counted, and the ratio is calculated to the total number of time windows participating in the evaluation to obtain the proportion of conflict duration. The spatial overlap rate, the degree of security interval intrusion, and the proportion of conflict duration are mapped to the same numerical range, and weighted and fused according to the overlap risk weight, the interval intrusion weight, and the duration weight to obtain the basic risk value of the conflict shadow. The basic risk value of the conflict shadow is subject to range constraints and penalty amplification processing to generate a conflict shadow penalty factor, so that the frost individuals with higher spatial overlap rate, greater degree of safety interval intrusion, or higher proportion of conflict duration receive greater conflict penalty intensity. The fitness value is formed by weighted summation of model error, cooperative execution cost, and conflict shadow penalty term. The initial population is then updated using a frost search based on this fitness value, iteratively adjusting the set of parameters to be optimized and the decision threshold. Specifically, the fitness value is formed by weighted summation of model error, cooperative execution cost, and conflict shadow penalty term. Read the model error, cooperative execution cost, and conflict shadow penalty term for each frost individual, map the model error, cooperative execution cost, and conflict shadow penalty term to the same numerical range, and perform a weighted sum according to the error weight, execution cost weight, and conflict penalty weight to obtain the fitness value. The fitness value is used to characterize the overall advantages and disadvantages of the parameter combination corresponding to the frost individual in terms of model prediction accuracy, cooperative execution cost, and trajectory conflict risk. The smaller the fitness value, the better the parameter combination.

[0026] The initial population is updated using a frost search based on its fitness value. The set of parameters to be optimized and the decision threshold are iteratively adjusted as follows: Frost individuals are sorted according to their fitness values, and the individual with the smallest fitness value is selected as the current best individual. A search update is generated based on the encoding differences between the current best individual and other frost individuals, updating the number of node encoding dimensions, the number of hypergraph convolutional layers, feature fusion weights, various hyperedge weights, and decision thresholds. Boundary correction and value type correction are performed on the updated frost individuals, and their fitness values ​​are recalculated. The current best individual is updated according to the recalculated fitness value, and the frost search update is iteratively performed until the termination condition is met. When a frost-ice individual violates the energy margin limit, time margin limit, communication coverage limit, no-fly zone boundary limit, obstacle safety boundary limit, or safety interval limit, the infeasible solution repair operator is invoked to repair the parameters and recalculate the fitness value. When the number of iterations reaches the maximum number of iterations or the fitness difference between two consecutive generations is lower than the convergence threshold, the frost individual with the smallest fitness value is selected as the target parameter combination, and the dynamic hypergraph neural network model is updated and improved with the target parameter combination.

[0027] In this embodiment, the construction of a candidate execution scheme that includes task attribution, task order, track nodes, and arrival time slots includes: The drone task combinations are filtered according to the task matching threshold to form a candidate task affiliation set; The tasks are sorted according to their priority, time limit, and duration to form a candidate task order set; Candidate waypoints are selected based on track cost score, energy consumption risk score and conflict risk score to form a set of candidate track nodes; Arrival time windows are divided based on the UAV's flight speed, mission execution time, and candidate track node set, forming a candidate arrival time slot set. Specifically: The system reads the UAV's current position, flight speed, candidate path node set, mission start time, mission execution time, and mission time limit; calculates the segment distance between adjacent path points according to the order of waypoints in the candidate path node set, and obtains the estimated flight time for each segment by combining the UAV's flight speed; sums the estimated flight time of each segment with the mission execution time to obtain the estimated arrival time and estimated completion time corresponding to the candidate path; divides the mission time limit range into multiple continuous time windows according to a preset time granularity, and marks the time window into which the estimated arrival time falls as a candidate arrival time slot; when the estimated completion time does not exceed the mission time limit and the candidate arrival time slot meets the requirements of communication coverage, energy margin, and safety interval, the corresponding candidate arrival time slot is written into the candidate arrival time slot set. The candidate task affiliation set, candidate task order set, candidate track node set, and candidate arrival time slot set are encoded and combined, and schemes with missing tasks, duplicate tasks, track breaks, and time slot conflicts exceeding a preset threshold are eliminated to form candidate execution schemes.

[0028] In this embodiment, the joint optimization and constraint repair of candidate execution schemes to generate multi-UAV task allocation and path planning results that satisfy energy, time limit, communication, obstacle avoidance, and safety interval constraints includes: Read the candidate execution plans and calculate the comprehensive evaluation value based on task completion rate, trajectory cost, energy consumption risk, conflict risk, and task load balance. Specifically, the calculation of the comprehensive evaluation value based on these factors is as follows: Read the task ownership, task order, track nodes, arrival time slots and remaining battery status from the candidate execution schemes, calculate the ratio of the number of assigned tasks to the total number of tasks, and obtain the task completion rate; calculate the track cost based on the track nodes and segment distances in the candidate execution schemes. Energy consumption risk is calculated based on remaining battery status, estimated flight energy consumption, mission execution energy consumption, and battery reserve upon return. Calculate the conflict risk based on the overlap and safety interval status of different UAVs within the same time window in the candidate execution schemes; The task load balance is calculated based on the differences in the number of tasks assigned to each drone, the task execution time, and the energy consumption. The task completion rate is converted into a positive evaluation value, while the trajectory cost, energy consumption risk, conflict risk, and task load balance are converted into negative evaluation values. The results are then weighted and fused according to the completion rate weight, trajectory cost weight, energy consumption risk weight, conflict risk weight, and load balance weight to obtain the comprehensive evaluation value of the candidate execution scheme. The candidate execution plans are ranked according to the comprehensive evaluation value, and the candidate execution plans that meet the evaluation conditions are selected as the execution plans to be repaired. The repair execution plan includes supplementing unassigned tasks, deleting duplicate tasks, transferring low-power tasks, detouring in no-fly zones, detouring around obstacles, and adjusting conflict time slots. The corrected execution plan is defined as the result of multi-UAV task allocation and path planning. Specifically, the corrected execution plan is defined as the result of multi-UAV task allocation and path planning. The system reads the task affiliation, task execution order, track node sequence, arrival time slot, remaining battery status, and constraint repair records from the repaired execution plan. It then performs integrity and constraint compliance checks on the repaired execution plan. Integrity checks confirm that there are no missing task affiliations, task order, track nodes, or arrival time slots. Constraint compliance checks confirm that energy margin, task time limits, communication coverage, no-fly zone boundaries, obstacle safety boundaries, and safety intervals meet the requirements. If the checks pass, the system writes the task affiliation, task execution order, track node sequence, arrival time slot, and remaining battery status into the result field, forming the multi-UAV task allocation and path planning result. If the checks fail, the failed items are returned to the constraint repair step for re-repair, and the multi-UAV task allocation and path planning result is output after the re-check passes.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a post-disaster emergency inspection scenario in a certain industrial park. The test area was a semi-open area of ​​2.6km × 1.9km, including office buildings, storage areas, waterways, temporary isolation zones, and road intersections. A total of 8 drones participated in the operation, of which 4 were equipped with visible light cameras, 2 with infrared thermal imaging equipment, 1 with a loudspeaker, and 1 with a small material delivery device. The initial battery level of the drones ranged from 72% to 96%, the maximum flight speed was 12m / s, the minimum return-to-home battery threshold was 15%, and the safe interval between multiple drones was 30m. There were 32 mission objectives, including road access inspection, rooftop anomaly verification, heat source search, personnel calling, and small material delivery; the mission time limit was 5min–28min. 14 fixed obstacles, 2 temporary no-fly zones, and 3 areas with weak communication coverage were set up in the operation area. During the execution, scenarios such as the addition of new tasks, changes in no-fly zone boundaries, and abnormal battery drops in a single drone were simulated.

[0030] At the start of the mission, the system first collects UAV status data, including position, speed, battery level, payload status, sensor status, communication status, and mission load information. Simultaneously, it collects mission data, including mission location, mission type, priority, mission time limit, mission duration, payload requirements, and sensor requirements. This data is then combined with information on obstacle areas, no-fly zone boundaries, communication coverage, candidate waypoints, and multi-aircraft safety intervals to form multi-source fusion features. These multi-source fusion features include UAV flight capability information (position accuracy ±0.5m, remaining battery percentage, payload weight, and effective sensor status), mission constraint information (mission coordinate accuracy ±0.5m, mission time limit deviation ±10s), and airspace constraint information (no-fly zone boundary accuracy ±1m, weak communication coverage area radius 20m, obstacle height 2m–8m).

[0031] The system performs reachability filtering based on UAV flight capabilities, mission constraints, and airspace constraints. The filtering rules include: mission combinations that cannot be reached due to insufficient battery power (approximately 12%) are marked as unreachable; combinations with mismatched payload or sensor types (8%); combinations that cannot be reached due to mission time constraints (15%); and combinations whose flight paths cross no-fly zones (5%). After filtering, a total of 420 feasible UAV mission flight path combinations remain, with each UAV having at least four selectable combinations, covering all mission points.

[0032] Subsequently, the system constructs a dynamic hypergraph of mission and trajectory nodes, including UAVs, missions, tracks, environment, time windows, and energy status nodes. There are a total of 182 nodes: 8 UAV nodes, 32 mission nodes, 96 track nodes, 14 environment nodes, 16 time window nodes, and 16 energy status nodes. Node characteristics include position vector, battery percentage, mission priority, estimated trajectory cost, obstacle distance, airspace risk score (0–1), and multi-UAV safety separation. The system establishes 168 hyperedges across 5 categories: 32 UAV-mission-energy hyperedges, 96 mission-track-environment hyperedges, 24 UAV-track-time window hyperedges, 12 multi-UAV-track conflict hyperedges, and 4 mission-priority-execution time limit hyperedges.

[0033] The system inputs the mission trajectory dynamic hypergraph into an improved dynamic hypergraph neural network model. The model includes a multi-source node encoding layer, an reachability-gated hyperedge update layer, a mission trajectory feedback convolutional layer, and a joint scoring output layer. In the multi-source node encoding layer, the system encodes UAV state nodes, mission nodes, and trajectory nodes separately, with a node feature dimension of 64. Each node encoding incorporates a node type scaling factor. The reachability-gated hyperedge update layer adjusts the hyperedge weights based on node reachability labels, resetting the weights of inaccessible combinations to zero. The mission trajectory feedback convolutional layer maps trajectory cost, energy consumption risk, and airspace risk back to the node encoding through convolution, constructing a four-dimensional conflict shadow that occupies three future time windows, each with a 2×2m² rasterized space occupancy rate. Conflict shadow prediction can identify potential airspace overlap risks in advance. The joint scoring output layer calculates the mission matching score (0–1), trajectory cost score (0–100), energy consumption risk score (0–1), and conflict risk score (0–1).

[0034] The constraint-guided frost-ice optimization algorithm optimizes model parameters, hyperedge weights, and decision thresholds. The initial population consists of 32 individuals, each with node encoding dimensions, convolutional layer number, feature fusion weights, and various hyperedge weights. Fitness is weighted by model error, cooperative execution cost, and conflict shadow penalty term, with weights of 0.4, 0.4, and 0.2, respectively. During iteration, the proportion of individuals with energy margins below 20% decreased to 0, the track conflict rate decreased from 9 times to 1 time, and the task timeout rate decreased from 6 times to 1 time.

[0035] Based on the updated model output, the system generates 64 candidate execution plans, each including task affiliation, task order, track nodes, and arrival time slots. The system filters and sorts the candidate plans according to the task matching threshold (0.75), task priority, track cost, energy consumption risk, conflict risk, and task load balancing. Plans involving low-battery task detours, task duplication, and time slot conflicts are adjusted through the constraint repair module, ultimately generating one optimal execution plan.

[0036] During execution, dynamic changes were simulated: three new heat source search tasks were added at the 18th minute; the temporary no-fly zone expanded eastward by 70 meters at the 31st minute; and the battery level of an infrared UAV dropped to 28% at the 43rd minute. The system was able to recalculate task assignments, track nodes, and arrival time slots based on the updated dynamic hypergraph of mission tracks, maintaining a safe distance of ≥30 meters between multiple aircraft, controlling the average delay of affected tasks to within 2.1 minutes, and ensuring a 96.88% completion rate for all critical tasks.

[0037] In the comparative experiment, the method of this invention was compared with the "distance-priority task allocation + A-satellite path planning" method. Under the same task conditions, the traditional method completed 26 tasks, with a task completion rate of 81.25%, a total flight distance of 37.8 km, an average single task completion time of 14.6 min, 6 task timeouts, 9 potential track conflicts, a minimum return battery level of 10.8%, and an average replanning time of 21.4 s after adding a new task. The present invention completed 31 tasks, with a task completion rate of 96.88%, a total flight distance of 31.9 km, an average single task completion time of 11.2 min, 1 task timeout, 1 potential track conflict, a minimum return battery level of 18.6%, and an average replanning time of 7.3 s after adding a new task.

[0038] As can be seen from this embodiment, the present invention provides quantitative data support for each step of the operation, including multi-source feature construction, reachability screening, dynamic hypergraph construction of mission trajectory, calculation of improved dynamic hypergraph neural network model, parameter optimization of frost and ice optimization algorithm and constraint repair, which verifies the feasibility and superiority of the present invention in multi-UAV mission allocation and path planning in complex dynamic environments.

Claims

1. A multi-UAV task allocation and path planning method based on machine learning, characterized in that, include: Collect data on collaborative operations of multiple UAVs, and standardize, spatiotemporally align, and encode the data to form multi-source fusion features; The flight capability, mission constraints, and airspace constraints in the multi-source fusion features are screened for accessibility to form an energy-time-limited reachable domain. Based on the multi-source fusion characteristics and energy time-limited reachable domain, node types are divided, hyperedges are established and weights are initialized to construct a dynamic hypergraph of the mission trajectory. Input the dynamic hypergraph of the mission trajectory into the improved dynamic hypergraph neural network model, perform node encoding, reachability-gated hyperedge update, mission trajectory feedback convolution and conflict shadow prediction, and output mission matching score and trajectory evaluation results. For the node encoding dimension, hypergraph convolution layer number, feature fusion weight, hyperedge weight and decision threshold in the improved dynamic hypergraph neural network model, the frost optimization algorithm with energy-time coupling metric function and conflict shadow penalty factor is used for iterative optimization. The fitness is constructed with model error and cooperative execution cost, and the target parameter combination is generated and the improved dynamic hypergraph neural network model is updated. Based on the updated model output, construct candidate execution schemes that include task attribution, task order, track nodes, and arrival time slots; Joint optimization and constraint repair are performed on candidate execution schemes to generate multi-UAV task allocation and path planning results that meet constraints of energy, time limit, communication, obstacle avoidance and safety interval; The output task matching score and trajectory evaluation results include: An improved dynamic hypergraph neural network model is constructed, which includes a multi-source node encoding layer, a reachability-gated hyperedge update layer, a mission trajectory feedback convolutional layer, and a joint scoring output layer. The multi-source node coding layer introduces a node type scaling factor. The reciprocal of the square root of the proportion of node types appearing in the mission track dynamic hypergraph is calculated to obtain the node type scaling factor. All node coding features are multiplied by the corresponding node type scaling factor and then written into the improved dynamic hypergraph neural network model. The reachable domain gated superedge update layer introduces a gated temperature coefficient. The gated temperature coefficient is obtained by logarithmically scaling the sum of the energy margin percentage and the time margin percentage. The gated temperature coefficient is used to adjust the weight of reachable combination superedges and reset the weight of unreachable combination superedges to zero. The task trajectory feeds back to the convolutional layer to introduce a risk decay coefficient. Vertex-to-superedge feature aggregation is performed to obtain temporary superedge features. Superedge-to-vertex feature backpropagation is performed to generate task matching intermediate features and trajectory cost intermediate features. The average value of the trajectory cost intermediate features of the current batch is obtained by exponential decay mapping to obtain the risk decay coefficient. The trajectory cost intermediate features are scaled according to the risk decay coefficient and embedded into the task matching intermediate features. Based on the trajectory node sequence, UAV flight speed, arrival time slot sequence and safety interval radius, temporal position inference is performed to construct the three-dimensional occupancy volume of future multi-time windows. Then, the occupancy volume is expanded in time dimension to generate a four-dimensional occupancy block. The four-dimensional occupancy block is written as a conflict shadow into the multi-UAV trajectory conflict superedge. The joint scoring output layer introduces a conflict penalty coefficient, which is obtained by linear combination of the spatial overlap rate of conflict shadows within the same time window and a fixed weight. The conflict penalty coefficient is used to perform risk weighting on the joint representation of the mission trajectory, and outputs the mission matching score as well as trajectory cost, energy consumption risk and conflict risk scores. The improved dynamic hypergraph neural network model is trained by constructing a training sample set. The training objective is the weighted sum of task matching classification loss, trajectory cost regression loss, energy consumption risk regression loss, and conflict risk classification loss. Backpropagation is used to simultaneously update the multi-source node encoding layer, reachability-gated hyperedge update layer, task trajectory feedback convolutional layer, joint scoring output layer, as well as node type scaling coefficient, gating temperature coefficient, risk decay coefficient, and conflict penalty coefficient.

2. The multi-UAV task allocation and path planning method based on machine learning according to claim 1, characterized in that, The formation of multi-source fusion features includes: Collect drone status data, mission requirement data, and environmental constraint data; The flight capability information is obtained by extracting location status, battery status, payload status, perception status, communication status, and mission load status from the UAV status data. The mission airspace constraint information is obtained by extracting mission location, mission attributes, mission time limit, payload requirements, sensor requirements, obstacle areas, no-fly zone boundaries, communication coverage areas, candidate waypoints, and safety intervals from mission requirement data and environmental constraint data. Missing value filling, coordinate system unification, and encoding are performed on the flight capability information and mission airspace constraint information. The data is then spliced ​​together according to the UAV number, mission number, and time window to form multi-source fusion features.

3. The multi-UAV task allocation and path planning method based on machine learning according to claim 1, characterized in that, The reachable range of the energy formation time limit includes: Extract flight capability information, mission constraint information, and airspace constraint information from multi-source fusion features; Calculate the estimated arrival time and estimated mission completion time based on the drone's current location, mission location, flight speed, mission duration, and mission time limit. Mark combinations where the estimated mission completion time does not exceed the mission time limit as time-reachable. The energy margin is calculated based on the drone's remaining battery power, estimated flight energy consumption, mission execution energy consumption, and return-to-home reserve power. Combinations that meet the safe return-to-home power requirements are marked as energy-available. Capability matching and airspace selection are performed based on payload capacity, sensor type, payload requirements, sensor requirements, obstacle areas, no-fly zone boundaries, communication coverage areas, and safety intervals. The flight paths of UAV missions that simultaneously meet the requirements of time reachability, energy reachability, capability matching, and airspace selection are written into the energy-time reachability domain, and the energy margin, time margin, and reachability marker are recorded.

4. The multi-UAV task allocation and path planning method based on machine learning according to claim 1, characterized in that, The construction of the dynamic hypergraph of the mission trajectory includes: Based on the characteristics of multi-source fusion and the reachability of energy time limits, node types are divided into UAV nodes, mission nodes, track nodes, environment nodes, time window nodes, and energy status nodes. Write the UAV status characteristics, mission constraint characteristics, candidate waypoint characteristics, airspace constraint characteristics, mission execution time period, and energy status characteristics into the corresponding nodes; Based on reachability markers, candidate tracks, airspace constraints, expected arrival time slots, and mission execution constraints, we establish UAV-mission-energy hyperedges, mission-track-environment hyperedges, UAV-track-time window hyperedges, multi-UAV-track conflict hyperedges, and mission-priority-execution time limit hyperedges. The weights of various hyperedges are initialized based on reachability, energy margin, time margin, track cost, airspace risk, and safety interval margin to form a dynamic hypergraph of the mission track.

5. The multi-UAV task allocation and path planning method based on machine learning according to claim 1, characterized in that, The process of generating target parameter combinations and updating and improving the dynamic hypergraph neural network model includes: The set of parameters to be optimized is determined, including the number of node encoding dimensions, the number of hypergraph convolutional layers, feature fusion weights, UAV mission energy hyperedge weights, mission trajectory environment hyperedge weights, UAV trajectory time window hyperedge weights, and multi-UAV trajectory conflict hyperedge weights. The decision thresholds include mission matching thresholds, trajectory risk thresholds, and replanning trigger thresholds. An energy-time coupling metric function is constructed, mapping the percentage of energy margin and the percentage of time margin to the same numerical interval. The energy-time coupling metric is obtained by weighted fusion according to the energy-sensitive weight and the time-sensitive weight. The energy-time coupling metric is divided into high-constraint interval, medium-constraint interval and low-constraint interval according to the energy-time coupling metric value. The set of parameters to be optimized is sampled hierarchically, and the sampling results are encoded as frost individuals to generate the initial population. The parameter combination corresponding to the individual frost ice is written into the improved dynamic hypergraph neural network model. The model error and collaborative execution cost are calculated. The conflict shadow penalty factor is constructed based on the spatial overlap rate, the degree of security interval intrusion and the proportion of conflict duration in the conflict shadow. The conflict shadow penalty factor is weighted and fused with the multi-UAV trajectory conflict hyperedge weight and the conflict risk score to obtain the conflict shadow penalty term. The fitness value is formed by weighted summation of model error, cooperative execution cost and conflict shadow penalty term. The initial population is then updated by frost search according to the fitness value, and the set of parameters to be optimized and the decision threshold are iteratively adjusted. When a frost-ice individual violates the energy margin limit, time margin limit, communication coverage limit, no-fly zone boundary limit, obstacle safety boundary limit, or safety interval limit, the infeasible solution repair operator is invoked to repair the parameters and recalculate the fitness value. When the number of iterations reaches the maximum number of iterations or the fitness difference between two consecutive generations is lower than the convergence threshold, the frost individual with the smallest fitness value is selected as the target parameter combination, and the dynamic hypergraph neural network model is updated and improved with the target parameter combination.

6. The multi-UAV task allocation and path planning method based on machine learning according to claim 1, characterized in that, The construction of candidate execution schemes, including task attribution, task order, track nodes, and arrival time slots, includes: The drone task combinations are filtered according to the task matching threshold to form a candidate task affiliation set; The tasks are sorted according to their priority, time limit, and duration to form a candidate task order set; Candidate waypoints are selected based on track cost score, energy consumption risk score and conflict risk score to form a set of candidate track nodes; Arrival time windows are divided based on the UAV's flight speed, mission execution time, and candidate track node set to form a candidate arrival time slot set; The candidate task affiliation set, candidate task order set, candidate track node set, and candidate arrival time slot set are encoded and combined, and schemes with missing tasks, duplicate tasks, track breaks, and time slot conflicts exceeding a preset threshold are eliminated to form candidate execution schemes.

7. The multi-UAV task allocation and path planning method based on machine learning according to claim 1, characterized in that, The process of jointly optimizing and constraining candidate execution schemes to generate multi-UAV task allocation and path planning results that satisfy energy, time, communication, obstacle avoidance, and safety interval constraints includes: Read the candidate execution plans and calculate the comprehensive evaluation value based on the task completion rate, trajectory cost, energy consumption risk, conflict risk and task load balance. The candidate execution plans are ranked according to the comprehensive evaluation value, and the candidate execution plans that meet the evaluation conditions are selected as the execution plans to be repaired. The repair execution plan includes supplementing unassigned tasks, deleting duplicate tasks, transferring low-battery tasks, detouring in no-fly zones, detouring around obstacles, and adjusting conflict time slots. The corrected execution plan was determined as the result of multi-UAV task allocation and path planning.

Images