Unmanned aerial vehicle group adaptive dynamic rescheduling method and system under space constraint

Abstract

The application discloses a kind of unmanned aerial vehicle group adaptive dynamic rescheduling method and system under space constraint, by obtaining the real-time running state of whole network unmanned aerial vehicle, confirming predicted track and projecting to space safety constraint partition to obtain global potential constraint state;Based on the time-space overlap degree of predicted track, conflict value is calculated, when exceeding threshold, pre-trained priority allocation neural network is called, and control priority ranking result is output by integrating task urgency and real-time running state;Then, through adaptive control model, ranking result is converted into differentiated speed, height and heading correction instruction, and closed-loop deviation feedback correction is supplemented, to realize whole-link adaptive rescheduling.The application solves the problem of static solidification of safety boundary and the problem that scheduling strategy cannot adapt to dynamic task demand by constructing dynamic space constraint partition of static-dynamic fusion and priority intelligent decision based on deep reinforcement learning, and significantly improves the cooperative efficiency and flight safety of unmanned aerial vehicle group in complex dynamic environment.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to an adaptive dynamic rescheduling method and system for UAV swarms under spatial constraints. Background Technology

[0002] With the widespread application of drone technology in fields such as power line inspection, logistics transportation, and emergency rescue, multi-drone collaborative operations have become an important means to improve mission efficiency. In actual operational scenarios, drone swarms need to perform tasks in a complex spatial constraint environment, which includes static obstacles such as high-voltage power equipment and buildings, as well as dynamic factors such as airflow wakes and electromagnetic interference generated by the drones themselves. At the same time, different inspection or transportation tasks often have different time requirements, and some high-urgency tasks require priority access to airspace. How to dynamically coordinate the flight order of multiple drones according to the urgency of the task while ensuring flight safety and communication connectivity is a key technical challenge for achieving efficient drone swarm collaboration.

[0003] Existing UAV swarm scheduling methods primarily employ a combination of offline global planning and online local response. Before mission execution, the scheduling system generates a global flight path and mission execution sequence for each UAV based on known airspace information and a mission list, using a centralized optimization algorithm. When encountering unexpected situations during flight, the system typically employs rule-based local collision avoidance algorithms, such as artificial potential field methods or velocity obstacle methods, allowing the relevant UAVs to adjust their flight states autonomously to avoid collisions. Some schemes introduce the concept of mission priority, generating a fixed job sequence by prioritizing tasks before execution, but the online operation phase is still dominated by a pre-set plan.

[0004] The existing solutions have the following shortcomings: First, offline planning is difficult to adapt to dynamic and complex spatial constraints. When the UAV generates wake disturbances and electromagnetic interference in real time due to differences in UAV models during flight, the preset safety boundaries cannot be updated dynamically, resulting in scheduling decisions being out of sync with the actual flight environment. Second, in the face of time window requirements for high-urgency tasks, the existing scheduling mechanisms based on fixed rules or preset sequences lack global adaptive capabilities and cannot dynamically adjust the control priority of multiple UAVs according to the severity of real-time conflicts, which can easily cause delays in high-priority tasks or a decrease in overall scheduling efficiency. Summary of the Invention

[0005] The purpose of this application is to provide an adaptive dynamic rescheduling method and system for unmanned aerial vehicle (UAV) swarms under spatial constraints, which solves the problem of difficult adaptive scheduling of multiple UAVs under dynamic composite constraints.

[0006] To address the aforementioned technical problems, this application provides the following technical solutions: The first aspect of this application provides an adaptive dynamic rescheduling method for unmanned aerial vehicle (UAV) swarms under spatial constraints, the method comprising: The system obtains the real-time operational status of the entire network of flying drones, confirms the predicted flight path of each drone in the future based on the real-time operational status, and projects the predicted flight path of each drone onto the spatial safety constraint partition to obtain the global potential constraint status. Based on the degree of spatiotemporal overlap between the predicted flight paths of each UAV within the spatial safety constraint zone, the spatiotemporal conflict value between each UAV is calculated, and the spatiotemporal conflict value is compared with a preset spatiotemporal conflict threshold. If the spatiotemporal conflict value exceeds the spatiotemporal conflict threshold, a pre-trained priority allocation neural network is invoked, taking the task urgency and real-time operating status of the drones with spatiotemporal conflicts as input features, and outputting the control priority ranking result of each drone with spatiotemporal conflict. Based on the control priority ranking result, the motion state control parameters of the UAV with spatiotemporal conflict when entering the spatial safety constraint zone are adjusted online, and a dynamic rescheduling execution sequence is generated and issued to guide the UAV with high priority tasks to pass through safely while ensuring autonomous avoidance by multiple UAVs.

[0007] A second aspect of this application provides an adaptive dynamic rescheduling system for unmanned aerial vehicle (UAV) swarms under spatial constraints, the system comprising: The constraint state acquisition module is used to acquire the real-time operating status of the entire network of flying drones, determine the predicted trajectory of each drone in the future time period based on the real-time operating status, and project the predicted trajectory of each drone onto the spatial safety constraint partition to obtain the global potential constraint state. The conflict value comparison module is used to calculate the spatiotemporal conflict value between each UAV based on the degree of spatiotemporal overlap between the predicted trajectories of each UAV within the spatial safety constraint zone, and compare the spatiotemporal conflict value with a preset spatiotemporal conflict threshold. The control priority output module is used to call a pre-trained priority allocation neural network if the spatiotemporal conflict value exceeds the spatiotemporal conflict threshold. The module takes the task urgency and real-time operating status of the UAVs with spatiotemporal conflicts as input features and outputs the control priority ranking result of each UAV with spatiotemporal conflict. The rescheduling module is used to adjust the motion state control parameters of the UAV with spatiotemporal conflicts when it enters the spatial safety constraint zone according to the control priority sorting result, generate a dynamic rescheduling execution sequence and issue it, so as to guide the UAV with high priority tasks to pass through safely while ensuring the autonomous avoidance of multiple UAVs.

[0008] Compared to existing technologies, the adaptive dynamic rescheduling method and system for UAV swarms under spatial constraints provided in this application achieves accurate prediction of potential conflicts by constructing a static-dynamic integrated spatial safety constraint partition and projecting the future flight paths of UAVs in four dimensions. When the conflict value exceeds a threshold, a deep reinforcement learning network is invoked to prioritize tasks based on their urgency and real-time status. The ranking result is then transformed into differentiated correction instructions through an adaptive control model, supplemented by closed-loop deviation feedback to form a full-link adaptive closed loop. This invention effectively solves two major defects in existing technologies: the disconnect between static safety boundaries and the actual flight environment, and the inability of fixed-rule scheduling to adapt to dynamic task requirements. Through dynamically updated composite spatial constraint partitions, scheduling decisions are synchronized with the real flight environment in real time. Through a priority allocation mechanism based on deep reinforcement learning, a global optimal trade-off between second-level priority response to high-urgency tasks and multi-aircraft collaborative avoidance is achieved. Through priority-driven adaptive control and closed-loop correction, reliable execution of rescheduling instructions under external interference such as wind disturbances is ensured. The overall solution significantly improves the collaborative efficiency, task timeliness, and flight safety of UAV swarms in complex dynamic environments. Attached Figure Description

[0009] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein: Figure 1 A flowchart illustrating an adaptive dynamic rescheduling method for UAV swarms under spatial constraints is shown schematically. Figure 2 A schematic diagram of an adaptive dynamic rescheduling system for unmanned aerial vehicle (UAV) swarms under spatial constraints is shown. Detailed Implementation

[0010] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0011] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0012] The methods described in the embodiments of this application will be explained in detail below.

[0013] This embodiment provides an adaptive dynamic rescheduling method for UAV swarms under spatial constraints. This method can be applied to collaborative operations involving multiple UAVs in scenarios such as power line inspection, oil and gas pipeline inspection, emergency rescue, and logistics transportation. In this embodiment, the algorithm processing task can be flexibly deployed on a cloud server or an onboard edge computing terminal of the UAV, depending on the actual needs of the scenario. The task is completed collaboratively by the cloud server and the UAV terminal, or by the UAV terminals themselves, through a wireless communication network.

[0014] In this embodiment, the algorithm processing task can be flexibly deployed on a cloud server or an onboard edge computing terminal of a drone, depending on the actual scenario requirements. The task is completed collaboratively by the cloud server and the drone terminal, or between the drone terminals, via a wireless communication network. The execution entity of the method is a drone swarm adaptive dynamic rescheduling system deployed on the cloud server or the onboard edge computing terminal of the drone. This system is used to perceive the operating status and spatial constraint changes of the drone swarm in real time, execute dynamic rescheduling decisions online, and issue control commands to each drone.

[0015] Step 110: Obtain the real-time operating status of the entire network of flying drones, confirm the predicted flight path of each drone in the future time period based on the real-time operating status, and project the predicted flight path of each drone onto the spatial safety constraint partition to obtain the global potential constraint status.

[0016] This step provides a precise spatiotemporal data foundation for subsequent conflict detection and rescheduling. The system first establishes a digital twin of the physical world, synchronizing and processing three types of key information in real time: the real-time operational status of the UAV swarm, complex spatial safety constraints, and the four-dimensional spatiotemporal mapping relationship between the two.

[0017] The system acquires the real-time operational status of the entire network of unmanned aerial vehicles (UAVs) through airborne sensors and communication links. This real-time operational status may include, but is not limited to, flight status parameters, mission attribute parameters, platform status parameters, equipment status parameters, and communication status parameters.

[0018] The flight status parameters may include, but are not limited to: spatial position (longitude, latitude, altitude); three-dimensional velocity vectors (horizontal velocity component, vertical velocity component); attitude angles (roll angle, pitch angle, yaw angle); and three-axis angular velocities (roll rate, pitch rate, yaw rate).

[0019] Task attribute parameters may include, but are not limited to: the identifier of the currently executing task; the priority label of the currently executing task; the task deadline; the execution time; and the task progress.

[0020] Platform status parameters may include, but are not limited to: remaining battery power; remaining battery life; current load status (no load / full load / load weight).

[0021] Equipment status parameters may include, but are not limited to: the operating status of airborne sensors (normal / abnormal / degraded).

[0022] Communication status parameters may include, but are not limited to: communication link quality (signal strength, bandwidth, delay).

[0023] In a specific implementation, determining the predicted flight path of each UAV in the future time period based on the real-time operating status can be achieved through the following steps: Flight status parameters are extracted from the real-time operating status of each UAV, and micro-meteorological data of the operating area are obtained; The flight state parameters and the micro-meteorological data are constructed into time series features, which are then input into a pre-trained time series trajectory prediction model. The model extrapolates the continuous three-dimensional spatial position set of each UAV within a future set time window to generate the predicted trajectory.

[0024] Specifically, flight status parameters and micro-meteorological data are collected synchronously at a fixed sampling frequency. Micro-meteorological data can be obtained in real-time from micro-meteorological stations deployed within the operational area, airborne wind speed sensors, or through a meteorological data service interface. This data may include, but is not limited to, three-dimensional wind speed, wind direction, air pressure, and temperature. Next, the current flight status parameters and micro-meteorological data are aligned according to their timestamps, and the flight status parameters and micro-meteorological data at the same moment are horizontally concatenated to form a multi-dimensional feature vector. Finally, the feature vectors from N consecutive sampling moments within a preset time window are stacked vertically in temporal order to form an N×M two-dimensional feature matrix, where N is the number of time steps and M is the total number of feature dimensions. This matrix serves as the multi-dimensional time-series feature input to the time-series trajectory prediction model, used by the model to capture the temporal coupling pattern between the UAV's flight status and the meteorological environment, and to extrapolate the three-dimensional spatial position sequence for future periods. This time-series feature is then input into a pre-trained time-series trajectory prediction model.

[0025] The pre-trained temporal trajectory prediction model is constructed using deep learning architectures such as Long Short-Term Memory (LSTM) or Transformer to capture the temporal coupling pattern between UAV flight status and meteorological environment, extrapolating the three-dimensional spatial position sequence for future periods. The model is trained offline based on a historical flight database. Training samples consist of multi-dimensional time-series features collected during actual UAV flight missions and their corresponding future real-world trajectories. During training, the multi-dimensional time-series features are used as model input, and the continuous set of three-dimensional spatial positions within a set future time window is used as the prediction target. The model parameters are iteratively optimized by minimizing the mean squared error between the predicted and actual trajectories until the model converges. After training, the model can quickly extrapolate the predicted trajectories of each UAV for future periods based on real-time collected flight status parameters and micro-meteorological data during the inference phase.

[0026] In a specific implementation, spatial safety constraint partitioning is obtained through the following method: Three-dimensional safety boundaries are extracted from existing high-voltage equipment and buildings within the target airspace to generate static physical boundaries. Based on the model parameters of the UAVs in flight and combined with the real-time operating status, the influence range of airflow wakes and electromagnetic communication interference generated by each UAV during flight is calculated in real time, and dynamic derivative boundaries are dynamically generated. The static physical boundary and the dynamically derived boundary are fused in real time to dynamically update the spatial safety constraint partition.

[0027] Specifically, the "spatial constraint" in this solution is a composite concept, encompassing not only static physical boundaries but also dynamic, invisible boundaries generated in real time during UAV flight. Before system initialization or mission commencement, the target airspace is first modeled. This involves extracting three-dimensional safety boundaries from known hardware entities such as existing high-voltage power equipment, buildings, and communication towers, and then wrapping these entities with an extended safety margin (e.g., extending 15 meters beyond high-voltage lines) to generate a three-dimensional static physical boundary. Simultaneously, based on the UAV's flight parameters (such as rotor size, power type, communication power, and overall weight) and real-time operational status, the system uses aerodynamic wake models (such as rotor downwash models based on blade element theory or simplified computational fluid dynamics models) and electromagnetic propagation attenuation models (such as free-space path loss models) to calculate in real time the influence range of the airflow wake generated by the propeller downwash during each UAV's flight, as well as the electromagnetic communication interference envelope generated by the operation of onboard communication equipment. These real-time calculated influence ranges are then dynamically used to generate corresponding dynamic derived boundaries. Finally, the system performs real-time superposition and union calculation of the static physical boundary and the dynamically derived boundary in a unified three-dimensional spatial coordinate system, dynamically updating the spatial safety constraint partition. This forms a unified, multi-dimensional spatial safety constraint partition that is continuously and dynamically updated as the UAV flies.

[0028] In a specific implementation, projecting the predicted trajectories of each UAV onto the spatial safety constraint partition can be achieved through the following steps: The spatial safety constraint partition updated at the current moment is divided into multiple time slices along the time axis to form a spatiotemporal partition model covering spatial range and time series; Extract the spatial location of each UAV within each time slice from its predicted flight path; The spatial locations of all drones within each time slice are filled into the spatiotemporal partitioning model.

[0029] Specifically, after obtaining the predicted trajectories of each UAV and the updated spatial safety constraint partitions, it is necessary to perform a correlation analysis between the two under a unified spatiotemporal dimension. The control system divides the dynamically updated three-dimensional spatial safety constraint partitions at the current moment into multiple equally spaced time slices along the time axis (e.g., with a step size of 0.5 seconds, divided into...). Multiple consecutive time slices are used to construct a spatiotemporal partitioning model covering a three-dimensional spatial range and a one-dimensional time series. Then, for each UAV, its spatial position coordinates within each time slice are extracted from its generated predicted trajectory. Finally, the spatial positions of all UAVs within each time slice are filled into the spatiotemporal grid corresponding to the spatiotemporal partitioning model. This achieves the projection and alignment of the future trajectories of multiple UAVs under the same spatiotemporal reference frame.

[0030] After projecting the trajectory onto the spatiotemporal partitioning model, the system needs to automatically identify which areas are at risk of collision and which areas are subject to environmental constraints in order to form a global potential constraint state for scheduling decisions.

[0031] Therefore, in a specific implementation, obtaining the global potential constraint state can be achieved through the following steps: Traverse the spatiotemporal partitioning model, mark the regions where two or more UAVs have overlapping spatial positions in the same time slice as dynamic conflict nodes, and mark the regions where the UAV spatial positions overlap with the static physical boundary or the dynamic derived boundary as static restricted nodes. The global potential constraint state is formed by the spatial distribution of the dynamic conflict nodes and the static restricted nodes and the corresponding time slice set.

[0032] Specifically, the system will traverse the entire spatiotemporal partitioning model and identify and mark two potential constraint states: regions where two or more UAVs have overlapping spatial positions within the same time slice will be marked as dynamic conflict nodes; regions where UAVs have overlapping spatial positions with static physical boundaries or dynamically derived boundaries will be marked as static constrained nodes.

[0033] For example, if the projected spatial positions of two or more drones overlap in the same grid area within the same time slice (such as time slice t3), the area is marked as a "dynamic conflict node" (indicating that there is a risk of multi-drone collision at this spatiotemporal point). If the spatial location of a drone within a certain time slice is detected to overlap with the static physical boundary (such as the boundary of a high-voltage transmission tower) or the dynamic derived boundary (such as the strong wake region of an adjacent drone) within that time slice, then the overlapping area is marked as a "static confined node".

[0034] Ultimately, the control system collects all dynamically conflicting nodes and statically constrained nodes, and combines them into a global potential constraint state based on their spatial distribution coordinates, corresponding constraint type labels, and corresponding time slice sets. This state is expressed in the form of a matrix or graph structure, intuitively reflecting the distribution of all security conflicts and environmental constraints across the entire network in future time periods, providing a direct decision-making basis for subsequent priority ranking and adaptive avoidance rescheduling.

[0035] Step 120: Based on the degree of spatiotemporal overlap between the predicted flight paths of each UAV within the spatial safety constraint zone, calculate the spatiotemporal conflict value between each UAV, and compare the spatiotemporal conflict value with a preset spatiotemporal conflict threshold.

[0036] After acquiring the global potential constraint state and the multi-drone projected trajectories, the control system does not simply determine whether two drones will collide. Instead, it quantitatively assesses the severity of safety conflicts between drone pairs from the perspectives of probability and the persistence of spatiotemporal overlap. By comparing the quantitative conflict value with a preset conflict threshold, the system can automatically and intelligently identify which drone pairs have a significant collision risk and use this as the control boundary condition to trigger the adaptive rescheduling mechanism.

[0037] In a specific implementation, the spatiotemporal conflict value between each UAV is calculated based on the degree of spatiotemporal overlap between their predicted flight paths within the spatial safety constraint zone. This can be achieved through the following steps: For any two UAVs, calculate the number of time slices in which their spatial locations overlap in the spatiotemporal partitioning model. The ratio of the number of time slices to the total number of time slices of the predicted trajectory is used as the spatiotemporal conflict value.

[0038] Specifically, based on the spatiotemporal partitioning model constructed above, for any two UAVs (e.g., UAV A and UAV B) within the operational airspace, the control system extracts their respective projected predicted trajectories and performs the following conflict value calculation steps: The first step is to count overlapping time slices: Traverse all time slices covering the predicted flight path in the spatiotemporal partitioning model. In all consecutive time slices (let the total number of time slices be M), retrieve and calculate the number (let m) of time slices in which the spatial positions (i.e., the spatiotemporal grid cells they occupy) of UAV A and UAV B overlap. If, within a certain time slice, the spatial grid cells occupied by the two UAVs overlap, or the distance between the two UAVs is less than a preset safe distance threshold, then it is determined that spatial position overlap has occurred within that time slice.

[0039] The second step is to calculate the spatiotemporal conflict value: Calculate the ratio of the number of overlapping time slices m to the total number of predicted track time slices M, and then... ( This serves as the spatiotemporal conflict value between drone A and drone B. This conflict value... The value of is between [0, 1]. like =0 indicates that the two machines will have no spatiotemporal overlap during the future prediction period, and there is no risk of conflict. like A value close to 1 indicates that the two aircraft will have a long-term and continuous trajectory overlap during the future forecast period, resulting in extremely severe conflict.

[0040] The third step is threshold comparison and trigger determination: the calculated spatiotemporal conflict value is... With the preset spatiotemporal conflict threshold (For example, setting) =0.15) for comparison.

[0041] like This indicates that although the two aircraft briefly met, the risk of conflict was within a safe and controllable range, and no intervention was necessary. like If the conflict value exceeds the conflict threshold, it indicates that the conflict duration of the two machines is too high. The system determines that there is a high-risk spatiotemporal conflict between the two machines and then triggers the subsequent control priority allocation and dynamic rescheduling process.

[0042] Step 130: If the spatiotemporal conflict value exceeds the spatiotemporal conflict threshold, call the pre-trained priority allocation neural network, take the task urgency and real-time operating status of the UAVs with spatiotemporal conflicts as input features, and output the control priority ranking result of each UAV with spatiotemporal conflict.

[0043] Specifically, when the spatiotemporal conflict value calculated in step S120 exceeds a preset threshold, it indicates that there is an unacceptable safety hazard in the current flight order, and the system triggers an active rescheduling mechanism. The core task of this step is to use an artificial intelligence model to dynamically prioritize the control of drones with spatiotemporal conflicts, in order to determine who has priority to pass and who should cooperate to avoid them.

[0044] The system invokes a pre-trained deep reinforcement learning policy network, which is a pre-trained priority allocation neural network. Its input features are divided into two parts: first, the task urgency, which specifically includes the priority label of the task currently being performed by each UAV, the task deadline, and the execution time. These three indicators together characterize the timeliness requirements of the task; second, the real-time operating status, which specifically includes the current remaining battery power of each UAV, as well as flight status parameters extracted from the real-time operating status (which may include, but are not limited to, spatial position, flight speed, and heading angle). These indicators reflect the current capabilities and motion status of the UAV.

[0045] After receiving the aforementioned input features, the deep reinforcement learning policy network performs a forward inference through its internal policy function, directly outputting the control priority ranking of each UAV with spatiotemporal conflict. This ranking result clarifies the right-of-way order of each UAV within the conflict airspace—the UAV ranked higher enjoys the higher priority right-of-way.

[0046] This deep reinforcement learning policy network requires offline training before deployment. The training scenarios are constructed based on historical flight data and randomly generated large-scale multi-aircraft conflict scenarios in a simulation environment. During training, the network aims to maximize long-term cumulative rewards. The reward function is designed as a composite reward system comprising the following three components: task completion timeliness reward, which incentivizes the priority completion of high-urgency tasks; high-priority task priority passage reward, which guides the network to prioritize the right-of-way for high-urgency drones during conflicts; and conflict resolution speed reward, which encourages rapid and efficient conflict resolution decisions. Through large-scale offline training guided by the above composite reward function, the policy network gradually learns to comprehensively weigh task timeliness, drone real-time capabilities, and conflict situation in complex multi-aircraft conflict scenarios, outputting a globally optimal control priority ranking scheme.

[0047] The above solution upgrades traditional rule-based priority ranking to learning-based intelligent decision-making by introducing a deep reinforcement learning policy network. The feature design at the network input covers two dimensions: task urgency (priority label, task deadline, and execution time) and real-time operational status (remaining battery power and flight status parameters), ensuring that the decision considers both task timeliness and the actual execution capabilities of the UAV. The composite reward function designed during the training phase integrates three optimization objectives: timeliness, priority, and conflict resolution speed, ensuring that the trained network can output globally optimal ranking results under multi-objective constraints.

[0048] Suppose that within a certain inspection airspace, the predicted trajectories of UAVs A, B, and C overlap within the same time slice, triggering a rescheduling. The input features of the three UAVs are as follows: Drone A: Performing a routine inspection mission, with a priority tag of "general". The mission deadline is 30 minutes later, and the execution time is 10 minutes. The current remaining battery is 65%. It is located 50 meters east of the conflict airspace at an altitude of 80 meters and is approaching the conflict airspace at a speed of 8 m / s along a heading angle of 90°.

[0049] Drone B: Executing a fault repair mission, with the priority label "urgent". The mission deadline is 5 minutes later, and the execution time is 2 minutes. The current remaining battery is 40%. It is located 30 meters south of the conflict airspace at an altitude of 75 meters and is approaching the conflict airspace at a speed of 6 m / s along a heading angle of 180°.

[0050] Drone C: Performing a routine inspection mission, with a priority tag of "general". The mission deadline is 20 minutes later, and the execution time is 15 minutes. The current remaining battery is 80%. It is located 40 meters west of the conflict airspace at an altitude of 85 meters and is approaching the conflict airspace at a speed of 10 m / s along a heading angle of 90°.

[0051] After comprehensive evaluation by the deep reinforcement learning policy network, the output control priority ranking result is: B > C > A.

[0052] The ranking is based on the following criteria: Drone B is performing a fault repair mission, with the highest priority tag, and the mission deadline is only 5 minutes away, making the time requirement extremely urgent. Although its remaining battery power is low, it is still within its capabilities to complete the current repair mission, therefore it is given the highest control priority and given priority passage through the conflict airspace. Drone C and A are both performing routine inspections, but C's mission deadline is more urgent, its remaining battery power is more abundant, and its flight speed is faster, giving it a stronger maneuverability and avoidance capability, therefore its ranking is higher than A's. Drone A has been performing the longest mission and has moderate battery power, playing a major avoidance role in conflict resolution.

[0053] This example demonstrates that the network does not simply rank tasks based on a single metric, but rather comprehensively considers multiple characteristics across two dimensions—task urgency and real-time operational status—to output the globally optimal control priority scheme.

[0054] Therefore, in a specific implementation, the pre-trained priority allocation neural network is a deep reinforcement learning policy network. The step of calling the pre-trained priority allocation neural network, taking the task urgency and real-time operating status of the UAVs with spatiotemporal conflicts as input features, and outputting the control priority ranking results for each of the UAVs with spatiotemporal conflicts, can be specifically implemented through the following steps: Obtain the priority tag, deadline, and execution time of the currently executing task of the UAV with spatiotemporal conflict, and use this as the task urgency. Obtain the current remaining battery power and flight status parameters of each of the aforementioned drones with spatiotemporal conflicts, and use the remaining battery power and flight status parameters as the real-time operating status. The task urgency and the real-time operating status are input into the deep reinforcement learning policy network; the deep reinforcement learning policy network outputs the control priority ranking results of each UAV with spatiotemporal conflict. The deep reinforcement learning policy network is obtained through offline training based on a composite reward function that includes rewards for timely task completion, rewards for prioritizing high-priority tasks, and rewards for conflict resolution speed.

[0055] Step 140: Based on the control priority ranking result, adjust the motion state control parameters of the UAV with spatiotemporal conflict when it enters the spatial safety constraint zone online, generate a dynamic rescheduling execution sequence and issue it, so as to guide the UAV with high priority task to pass through safely while ensuring multi-aircraft autonomous avoidance.

[0056] After obtaining the control priority ranking results for each UAV with spatiotemporal conflicts, the system enters the adaptive trajectory adjustment phase at the execution level. This phase requires converting the discrete priority ranking into underlying physical control actions to guide high-priority UAVs through safely while ensuring safe avoidance of multiple UAVs. This step transforms the priority ranking results into specific, executable motion control commands for each UAV, ensuring that high-priority UAVs pass smoothly through the conflict airspace first, while low-priority UAVs cooperate to avoid it.

[0057] First, the system converts the control priority ranking of each UAV into corresponding control weight coefficients. The higher the ranking and priority of the UAV, the larger its control weight coefficient. This control weight coefficient is a continuous value (e.g., between 0 and 1), used to adjust the strength of the UAV's adherence to the original flight path in subsequent optimization solutions.

[0058] Next, the trajectory deviation calculation unit in the control system obtains the original and actual flight paths of the UAVs experiencing spatiotemporal conflicts, and calculates the deviation between the original and actual flight paths. The actual flight path is generated by sensors onboard the UAVs, such as GPS modules, inertial measurement units (IMUs), and barometric altimeters, which collect their three-dimensional spatial coordinates in real time at a fixed frequency and continuously transmit the position data back to the rescheduling system via a wireless communication link. The system arranges the position data from each consecutive sampling moment by timestamp, forming a time series of the actual flight path. This trajectory reflects the actual flight path of the UAVs after being affected by factors such as wind disturbances, control delays, and positioning errors in the real environment. The original flight path is the ideal cruising path preset by the mission planning system before the conflict occurs.

[0059] The planned trajectory and the actual trajectory are each represented in the system as a set of three-dimensional spatial coordinate data arranged in chronological order. When calculating the deviation, the control system extracts the target position coordinates from the dataset of the planned trajectory at the same corresponding moment, and simultaneously extracts the current real position coordinates of the UAV from the dataset of the actual trajectory. Then, it compares and calculates the difference between these two sets of three-dimensional coordinates to determine the specific distance and direction of the UAV's deviation from the planned trajectory at the current moment. This three-dimensional difference result is the required deviation.

[0060] Considering the impact of environmental disturbances such as wind fields on control accuracy, the control system acquires micro-meteorological data of the operating area (such as ambient wind speed, wind direction, and airflow turbulence intensity), and calculates estimated values ​​of environmental disturbances applied to the UAV body (such as wind disturbance force vectors) based on the micro-meteorological data. This estimated value is used to perform feedforward compensation in advance during control solving to ensure the stability of the UAV during obstacle avoidance maneuvers.

[0061] The core principle of calculating environmental disturbance estimates based on micrometeorological data lies in transforming the dynamic effects of external airflow relative to the UAV body into specific disturbance forces through an aerodynamic drag model. The system first measures the absolute wind speed... with absolute wind direction Convert to three-dimensional absolute wind speed vector And combined with the drone's current velocity vector The relative wind speed vector of the airflow with respect to the UAV body is calculated by vector subtraction. .

[0062] Subsequently, the system uses a nonlinear aerodynamic drag formula to convert the relative wind speed vector Mapped to a three-dimensional environmental disturbance force vector acting on the organism In specific calculations, the algorithm combines the current air density... The inherent air drag coefficient of drones and the frontal area that changes dynamically with flight attitude Through the formula: ; The magnitude and direction of air resistance experienced by the UAV along each axis in its current attitude are determined. This is the three-dimensional environmental disturbance force vector. The final environmental disturbance estimate is input into the adaptive control model to implement active feedforward compensation in the control solution, ensuring the trajectory tracking accuracy and flight stability of the UAV under complex wind field disturbances.

[0063] Subsequently, the system inputs the above four data points—the deviation between the planned and actual flight paths, flight state parameters, estimated environmental disturbances, and control weight coefficients—into the adaptive control model. This model, built upon model predictive control or neural network self-tuning control, can solve for the corrected speed, altitude, and heading commands for each UAV in real time within a rolling time-domain optimization framework, based on the input multidimensional information.

[0064] In the optimization process, the control weight coefficient plays a key role: for high-priority UAVs with larger control weight coefficients, the optimization algorithm will give them a higher hold weight on their original trajectory, so that the trajectory adjustment range is as small as possible, in order to maintain stable flight and quickly pass through the conflict airspace; for low-priority UAVs with smaller control weight coefficients, the optimization algorithm allows for larger adjustments to their trajectory (such as deceleration, altitude fine-tuning or heading deflection), so as to actively create safe avoidance space for high-priority UAVs.

[0065] The adaptive control model outputs corrected instructions in real time. Based on this, the system generates a dynamic rescheduling execution sequence specifically for the drones with spatiotemporal conflicts, and sends it to each drone with a spatiotemporal conflict for execution via wireless communication. The remaining drones in the network that are not involved in the spatiotemporal conflict maintain their original flight plans. The entire adjustment process ensures that multiple drones autonomously avoid collisions, while prioritizing the safe passage of drones with high-priority tasks, demonstrating precise and efficient hierarchical scheduling capabilities.

[0066] The adaptive control model is built upon model predictive control or neural network self-tuning control and operates within a rolling time-domain optimization framework. Within each control cycle, the model, starting from the current actual trajectory, combines flight state parameters and environmental disturbance estimates to predict the UAV's motion state over a finite time domain. Then, with the objective function of minimizing the deviation between the planned and actual trajectories and the control cost, the optimal control sequence is solved while satisfying flight performance constraints (such as maximum speed, maximum turning rate, and safe distance). The control weight coefficients directly adjust the trajectory-keeping weight in the optimization objective function. UAVs with larger control weight coefficients have a greater weight in the optimization objective, resulting in smaller adjustments to their trajectories and prioritizing their original flight plan. Conversely, UAVs with smaller control weight coefficients have lower trajectory-keeping weights, allowing for larger adjustments to their speed, altitude, or heading, proactively creating safety space for higher-priority UAVs.

[0067] Therefore, in a specific implementation, adjusting the motion state control parameters of the UAV with spatiotemporal conflicts when entering the spatial safety constraint zone based on the control priority ranking result can be achieved online by the following steps: The control priority ranking results of the UAVs with spatiotemporal conflicts are converted into corresponding control weight coefficients, wherein the UAV with higher priority has a larger control weight coefficient. The original and actual flight paths of the UAV with spatiotemporal conflict are obtained, the deviation between the original and actual flight paths is calculated, and flight status parameters are extracted from the real-time operating status of the UAV with spatiotemporal conflict. Acquire micro-meteorological data of the work area, and calculate the estimated value of environmental disturbance based on the micro-meteorological data; The deviation between the original flight path and the actual flight path, the flight state parameters, the estimated environmental disturbance value, and the control weight coefficient are input into the adaptive control model, and the adaptive control model outputs the corrected speed, altitude, and heading commands for each UAV with spatiotemporal conflict in real time. The adaptive control model is constructed based on model predictive control or neural network self-tuning control. The control weight coefficient is used to adjust the retention weight of each UAV with spatiotemporal conflict on its original trajectory in the optimization solution. The larger the control weight coefficient, the smaller the adjustment range of its original trajectory.

[0068] This invention also provides a closed-loop adaptive correction mechanism based on position deviation feedback, used to continuously monitor the execution effect after the dynamic rescheduling command is issued, and to eliminate the potential for unexpected conflicts caused by external environmental interference. This mechanism can be implemented through the following steps: Real-time capture of the actual flight trajectories of each of the aforementioned UAVs with spatiotemporal conflicts when executing the corrected commands, and calculation of the trajectory position deviation between the actual flight trajectory and the desired trajectory; The trajectory position deviation is fed back to the adaptive control model to correct the deviation between the desired trajectory and the actual trajectory in the next control cycle, so as to eliminate the potential for unexpected conflicts caused by external environmental interference.

[0069] In step S140, the adaptive control model outputs the corrected speed, altitude, and heading commands for each relevant UAV in real time, based on the deviation between the original and actual flight paths, flight state parameters, estimated environmental disturbances, and control weight coefficients. However, due to uncertainties in the real operating environment, such as sudden changes in wind speed, positioning signal drift, and control command execution delays, secondary deviations may still occur between the actual flight trajectory of the UAV after executing the corrected commands and the expected effect predicted by the model. If such deviations are not corrected in a timely manner, the accumulated tracking errors may create new spatiotemporal conflicts in subsequent time slices, weakening the overall effect of rescheduling.

[0070] To this end, this closed-loop adaptive correction mechanism performs the following operations in each control cycle: The system continuously collects real-time position data of each relevant UAV after receiving and executing correction commands through sensors such as the GPS module, inertial measurement unit (IMU), and barometric altimeter mounted on each UAV. After being transmitted back via wireless communication link, the data is arranged by timestamp to construct the actual flight trajectory. Subsequently, the system compares the actual flight trajectory with the original trajectory on which the adaptive control model in step S140 is based, time-slice by time, and calculates the position deviation between the two in each axis to form the trajectory position deviation.

[0071] The trajectory position deviation is fed back to the adaptive control model in real time. In the next control cycle, the adaptive control model incorporates the accumulated trajectory position deviation from the previous cycle into the input item "deviation between the original trajectory and the actual trajectory," thereby dynamically adjusting the deviation input value for the current cycle. Through this rolling optimization closed loop of "execution-monitoring-feedback-correction," the adaptive control model can continuously perceive the real impact of external disturbances on the flight trajectory and actively compensate for deviations in subsequent command generation, ensuring that the actual flight paths of each relevant UAV gradually converge to the vicinity of the original trajectory, eliminating the risk of secondary conflicts caused by accumulated errors, and achieving dynamic closed-loop safety in the rescheduling process.

[0072] This application achieves accurate prediction of potential conflicts by constructing a static-dynamic integrated spatial safety constraint partition and projecting the future flight path of UAVs into a four-dimensional spatiotemporal plane. When the conflict value exceeds a threshold, a deep reinforcement learning network is invoked to prioritize tasks based on their urgency and real-time status. An adaptive control model then transforms the prioritization results into differentiated correction commands, supplemented by closed-loop deviation feedback to form a full-link adaptive closed loop. This invention effectively solves two major shortcomings of existing technologies: the disconnect between static safety boundaries and the actual flight environment, and the inability of fixed-rule scheduling to adapt to dynamic task requirements. Through dynamically updated composite spatial constraint partitions, scheduling decisions are synchronized with the real flight environment in real time. A priority allocation mechanism based on deep reinforcement learning achieves a globally optimal trade-off between second-level priority response to high-urgency tasks and multi-aircraft collaborative avoidance. Priority-driven adaptive control and closed-loop correction ensure reliable execution of rescheduling commands under external interference such as wind disturbances. The overall solution significantly improves the collaborative efficiency, mission timeliness, and flight safety of UAV swarms in complex dynamic environments.

[0073] Based on the same inventive concept, as an implementation of the above-mentioned adaptive dynamic rescheduling method for UAV swarms under spatial constraints, this application also provides an adaptive dynamic rescheduling system for UAV swarms under spatial constraints. Figure 2 This is a structural diagram of the system in an embodiment of this application. See also... Figure 2 As shown, the system includes: The constraint state acquisition module is used to acquire the real-time operating status of the entire network of flying drones, determine the predicted trajectory of each drone in the future time period based on the real-time operating status, and project the predicted trajectory of each drone onto the spatial safety constraint partition to obtain the global potential constraint state. The conflict value comparison module is used to calculate the spatiotemporal conflict value between each UAV based on the degree of spatiotemporal overlap between the predicted trajectories of each UAV within the spatial safety constraint zone, and compare the spatiotemporal conflict value with a preset spatiotemporal conflict threshold. The control priority output module is used to call a pre-trained priority allocation neural network if the spatiotemporal conflict value exceeds the spatiotemporal conflict threshold. The module takes the task urgency and real-time operating status of the UAVs with spatiotemporal conflicts as input features and outputs the control priority ranking result of each UAV with spatiotemporal conflict. The rescheduling module is used to adjust the motion state control parameters of the UAV with spatiotemporal conflicts when it enters the spatial safety constraint zone according to the control priority sorting result, generate a dynamic rescheduling execution sequence and issue it, so as to guide the UAV with high priority tasks to pass through safely while ensuring the autonomous avoidance of multiple UAVs.

[0074] This application achieves accurate prediction of potential conflicts by constructing a static-dynamic integrated spatial safety constraint partition and projecting the future flight path of UAVs into a four-dimensional spatiotemporal plane. When the conflict value exceeds a threshold, a deep reinforcement learning network is invoked to prioritize tasks based on their urgency and real-time status. An adaptive control model then transforms the prioritization results into differentiated correction commands, supplemented by closed-loop deviation feedback to form a full-link adaptive closed loop. This invention effectively solves two major shortcomings of existing technologies: the disconnect between static safety boundaries and the actual flight environment, and the inability of fixed-rule scheduling to adapt to dynamic task requirements. Through dynamically updated composite spatial constraint partitions, scheduling decisions are synchronized with the real flight environment in real time. A priority allocation mechanism based on deep reinforcement learning achieves a globally optimal trade-off between second-level priority response to high-urgency tasks and multi-aircraft collaborative avoidance. Priority-driven adaptive control and closed-loop correction ensure reliable execution of rescheduling commands under external interference such as wind disturbances. The overall solution significantly improves the collaborative efficiency, mission timeliness, and flight safety of UAV swarms in complex dynamic environments.

[0075] This application also provides an electronic device, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a program for an adaptive dynamic rescheduling method for UAV swarms under spatial constraints. When the processor executes the computer program, it implements all steps of the aforementioned adaptive dynamic rescheduling method for UAV swarms under spatial constraints, such as S110 to S140 and their sub-steps. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the aforementioned system embodiments.

[0076] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete this application. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0077] The electronic device may be a computing device such as a server or cloud server deployed in an adaptive dynamic rescheduling system for UAV swarms under spatial constraints. The electronic device may include, but is not limited to, processors and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or a combination of certain components, or different components. For example, the electronic device may also include data acquisition cards and output cards for connecting sensors and actuators, network communication devices, buses, etc.

[0078] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (OPGs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the entire electronic device via various interfaces and lines.

[0079] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created according to system usage, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart memory card, secure digital card, flash memory card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0080] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0081] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. An adaptive dynamic rescheduling method for unmanned aerial vehicle (UAV) swarms under spatial constraints, characterized in that, The method includes: The system obtains the real-time operational status of the entire network of flying drones, confirms the predicted flight path of each drone in the future based on the real-time operational status, and projects the predicted flight path of each drone onto the spatial safety constraint partition to obtain the global potential constraint status. Based on the degree of spatiotemporal overlap between the predicted flight paths of each UAV within the spatial safety constraint zone, the spatiotemporal conflict value between each UAV is calculated, and the spatiotemporal conflict value is compared with a preset spatiotemporal conflict threshold. If the spatiotemporal conflict value exceeds the spatiotemporal conflict threshold, a pre-trained priority allocation neural network is invoked, taking the task urgency and real-time operating status of the drones with spatiotemporal conflicts as input features, and outputting the control priority ranking result of each drone with spatiotemporal conflict. Based on the control priority ranking result, the motion state control parameters of the UAV with spatiotemporal conflict when entering the spatial safety constraint zone are adjusted online, and a dynamic rescheduling execution sequence is generated and issued to guide the UAV with high priority tasks to pass through safely while ensuring autonomous avoidance by multiple UAVs.

2. The method according to claim 1, characterized in that, The process of determining the predicted flight paths of each UAV in the future time period based on the real-time operating status includes: Flight status parameters are extracted from the real-time operating status of each UAV, and micro-meteorological data of the operating area are obtained; The flight state parameters and the micro-meteorological data are constructed into time series features, which are then input into a pre-trained time series trajectory prediction model. The model extrapolates the continuous three-dimensional spatial position set of each UAV within a future set time window to generate the predicted trajectory.

3. The method according to claim 1, characterized in that, The spatial safety constraint partition is obtained through the following method: Three-dimensional safety boundaries are extracted from existing high-voltage equipment and buildings within the target airspace to generate static physical boundaries. Based on the model parameters of the UAVs in flight and combined with the real-time operating status, the influence range of airflow wakes and electromagnetic communication interference generated by each UAV during flight is calculated in real time, and dynamic derivative boundaries are dynamically generated. The static physical boundary and the dynamically derived boundary are fused in real time to dynamically update the spatial safety constraint partition.

4. The method according to claim 3, characterized in that, The step of projecting the predicted trajectories of each UAV onto the spatial safety constraint partition includes: The spatial safety constraint partition updated at the current moment is divided into multiple time slices along the time axis to form a spatiotemporal partition model covering spatial range and time series; Extract the spatial location of each UAV within each time slice from its predicted flight path; The spatial locations of all drones within each time slice are filled into the spatiotemporal partitioning model.

5. The method according to claim 4, characterized in that, The process of obtaining the global potential constraint state includes: Traverse the spatiotemporal partitioning model, mark the regions where two or more UAVs have overlapping spatial positions in the same time slice as dynamic conflict nodes, and mark the regions where the UAV spatial positions overlap with the static physical boundary or the dynamic derived boundary as static restricted nodes. The global potential constraint state is formed by the spatial distribution of the dynamic conflict nodes and the static restricted nodes and the corresponding time slice set.

6. The method according to claim 4, characterized in that, The calculation of the spatiotemporal conflict value between UAVs, based on the degree of spatiotemporal overlap between the predicted trajectories of each UAV within the spatial safety constraint zone, includes: For any two UAVs, calculate the number of time slices in which their spatial locations overlap in the spatiotemporal partitioning model. The ratio of the number of time slices to the total number of time slices of the predicted trajectory is used as the spatiotemporal conflict value.

7. The method according to claim 1, characterized in that, The pre-trained priority allocation neural network is a deep reinforcement learning policy network. The invocation of the pre-trained priority allocation neural network uses the task urgency and real-time operating status of the UAVs with spatiotemporal conflicts as input features, and outputs the control priority ranking results for each of the UAVs with spatiotemporal conflicts, including: Obtain the priority tag, deadline, and execution time of the currently executing task of the UAV with spatiotemporal conflict, and use this as the task urgency. Obtain the current remaining battery power and flight status parameters of each of the aforementioned drones with spatiotemporal conflicts, and use the remaining battery power and flight status parameters as the real-time operating status. The task urgency and the real-time operating status are input into the deep reinforcement learning policy network; the deep reinforcement learning policy network outputs the control priority ranking results of each UAV with spatiotemporal conflict. The deep reinforcement learning policy network is obtained through offline training based on a composite reward function that includes rewards for timely task completion, rewards for prioritizing high-priority tasks, and rewards for conflict resolution speed.

8. The method according to claim 1, characterized in that, The step of adjusting the motion state control parameters of the UAV with spatiotemporal conflicts when entering the spatial safety constraint zone based on the control priority ranking result includes: The control priority ranking results of the UAVs with spatiotemporal conflicts are converted into corresponding control weight coefficients, wherein the UAV with higher priority has a larger control weight coefficient. The original and actual flight paths of the UAV with spatiotemporal conflict are obtained, the deviation between the original and actual flight paths is calculated, and flight status parameters are extracted from the real-time operating status of the UAV with spatiotemporal conflict. Acquire micro-meteorological data of the work area, and calculate the estimated value of environmental disturbance based on the micro-meteorological data; The deviation between the original flight path and the actual flight path, the flight state parameters, the estimated environmental disturbance value, and the control weight coefficient are input into the adaptive control model, and the adaptive control model outputs the corrected speed, altitude, and heading commands for each UAV with spatiotemporal conflict in real time. The adaptive control model is constructed based on model predictive control or neural network self-tuning control. The control weight coefficient is used to adjust the retention weight of each UAV with spatiotemporal conflict on its original trajectory in the optimization solution. The larger the control weight coefficient, the smaller the adjustment range of its original trajectory.

9. The method according to claim 8, characterized in that, After generating and issuing the dynamic rescheduled execution sequence, the method further includes: Real-time capture of the actual flight trajectories of each of the aforementioned UAVs with spatiotemporal conflicts when executing the corrected commands, and calculation of the trajectory position deviation between the actual flight trajectory and the desired trajectory; The trajectory position deviation is fed back to the adaptive control model to correct the deviation between the desired trajectory and the actual trajectory in the next control cycle, so as to eliminate the potential for unexpected conflicts caused by external environmental interference.

10. An adaptive dynamic rescheduling system for unmanned aerial vehicle (UAV) swarms under spatial constraints, characterized in that, The system includes: The constraint state acquisition module is used to acquire the real-time operating status of the entire network of flying drones, determine the predicted trajectory of each drone in the future time period based on the real-time operating status, and project the predicted trajectory of each drone onto the spatial safety constraint partition to obtain the global potential constraint state. The conflict value comparison module is used to calculate the spatiotemporal conflict value between each UAV based on the degree of spatiotemporal overlap between the predicted trajectories of each UAV within the spatial safety constraint zone, and compare the spatiotemporal conflict value with a preset spatiotemporal conflict threshold. The control priority output module is used to call a pre-trained priority allocation neural network if the spatiotemporal conflict value exceeds the spatiotemporal conflict threshold. The module takes the task urgency and real-time operating status of the UAVs with spatiotemporal conflicts as input features and outputs the control priority ranking result of each UAV with spatiotemporal conflict. The rescheduling module is used to adjust the motion state control parameters of the UAV with spatiotemporal conflicts when it enters the spatial safety constraint zone according to the control priority sorting result, generate a dynamic rescheduling execution sequence and issue it, so as to guide the UAV with high priority tasks to pass through safely while ensuring the autonomous avoidance of multiple UAVs.

Images