An unmanned aerial vehicle cluster autonomous cooperative control method and system

Abstract

The application discloses a kind of unmanned aerial vehicle cluster autonomous cooperative control method and system, belong to unmanned aerial vehicle cluster control technical field.System is constituted closed loop architecture by data acquisition, pre-processing, space-time evolution prediction, risk assessment, proactive redistribution, execution control, feedback learning seven big modules.The application fuses kinematics physical model and GAT-LSTM graph time sequence hybrid prediction model, adaptively weighted fusion output cluster multi-step future situation and calculate prediction confidence interval;From spatial conflict, task progress, energy exhaustion dimension constructs quantitative risk function, obtains comprehensive risk index by weighting, anticipates potential failure in advance and triggers redistribution.Constructs multi-objective optimization model with task length, energy consumption balance, communication load, minimum adjustment disturbance as target, solves Pareto optimal solution set using NSGA-II, and completes robust safety check output distribution scheme by combining confidence interval sampling.System continuously iterates prediction and optimization model parameters by online learning, realizes proactive active regulation.

Description

Technical Field

[0001] This invention relates to the technical field of unmanned aerial vehicle (UAV) swarm control, and more specifically, to an autonomous and cooperative control method and system for UAV swarms based on spatiotemporal evolution prediction and dynamic reallocation. Background Technology

[0002] With the rapid development of UAV technology and the booming low-altitude economy, single UAVs can no longer meet the increasingly complex mission requirements. UAV swarm collaborative control has become an important development direction in the current UAV technology field. Through multi-UAV collaboration, UAV swarms can accomplish complex tasks that are impossible for a single UAV, such as wide-area parallel search, multi-target collaborative reconnaissance, and swarm saturation strikes.

[0003] In practical engineering applications, the core challenge of collaborative control of UAV swarms lies in how to achieve efficient configuration and dynamic adjustment of swarm resources (including UAV platforms, payloads, onboard energy, communication bandwidth, etc.) in a highly dynamic mission environment, so as to maximize mission completion efficiency and ensure flight safety.

[0004] Currently, the following technical approaches are mainly used for task allocation and collaborative control of UAV swarms:

[0005] (1) Static task allocation method: Before the task begins, the task is pre-assigned to each UAV based on known prior information. This method is suitable for scenarios with small environmental changes and high task determinism (such as inspection tasks with fixed routes), but it cannot cope with sudden situations and dynamic changes.

[0006] (2) Reactive dynamic reallocation method: During task execution, task reallocation is triggered when a conflict, fault, or change in requirements is detected. For example, patent CN120560304B uses conflict detection to trigger reallocation, patent CN121879421A uses "capability or environmental semantic changes" to trigger reallocation, and patent CN120066067A uses a real-time computing paradigm for replanning. This type of method is essentially a reactive response; when reallocation occurs, the problem has already occurred or is in progress, which may lead to task delays, resource waste, or even flight safety accidents.

[0007] (3) Prediction-based trajectory planning method: For example, patent CN114339842A designs UAV trajectories by predicting the time-varying data of ground users, but the prediction object is the external environment (user needs) rather than the internal state of the cluster itself, and the prediction results are not used for task redistribution.

[0008] (4) Task planning methods based on reinforcement learning: In recent years, some task planning methods based on deep reinforcement learning have emerged, which make task allocation decisions by training an agent. However, the policy is usually fixed after training and lacks the ability to adapt to dynamic environments in real time and the ability to continuously evolve.

[0009] (5) Game theory-based cluster behavior prediction methods: For example, patent CN121705619A uses dynamic game coupling modeling to predict long-term cluster behavior, but focuses on game theory modeling itself and does not use the prediction results to guide cluster control decisions and task redistribution. Summary of the Invention

[0010] To address the aforementioned technical problems in related technologies, this invention provides a method and system for autonomous collaborative control of unmanned aerial vehicle (UAV) swarms, which can solve the above problems.

[0011] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows:

[0012] An autonomous collaborative control system for unmanned aerial vehicle (UAV) swarms includes:

[0013] The module includes a data acquisition module, a data preprocessing module, a spatiotemporal evolution prediction module, a bottleneck and risk assessment module, a proactive reallocation decision-making module, an execution control module, and a feedback and online learning module.

[0014] The data acquisition module is used to collect multi-dimensional status data of each UAV in the cluster in real time, including the first... A drone Position vector at time , No. A drone velocity vector at time t , No. A drone Remaining energy at any given moment , No. A drone The current amount of work completed at any given moment , No. The drone and the first Between drones Communication link quality at any moment Environmental perception data;

[0015] The data preprocessing module is used to perform filtering, noise reduction, spatiotemporal registration, and normalization on the collected data.

[0016] The spatiotemporal evolution prediction module is used to predict the spatiotemporal evolution of the cluster at multiple future time steps based on preprocessed data, using a hybrid architecture of physical prediction model and data-driven prediction model.

[0017] The bottleneck and risk assessment module is used to proactively assess the risks at future moments based on the prediction results of the spatiotemporal evolution situation. The risks include spatial conflict risk, mission progress risk, and energy depletion risk.

[0018] The proactive reallocation decision module is used to generate a reallocation scheme before the actual occurrence of the risk when the risk assessment result meets the triggering conditions.

[0019] The execution control module is used to execute the redistribution scheme and update the task queues and tracks of each UAV.

[0020] The feedback and online learning module is used to collect actual execution effect data, calculate prediction error and redistribution effect indicators, and update the parameters of the spatiotemporal evolution prediction module and the proactive redistribution decision module accordingly.

[0021] Furthermore, the spatiotemporal evolution prediction module includes:

[0022] The physical prediction model unit is used to calculate the predicted position and velocity for the next K time steps based on the kinematic equations and motion constraints of the UAV.

[0023] The data-driven prediction model unit is used to learn cluster interaction patterns from historical time series data based on graph attention network GAT and long short-term memory network LSTM, and predict the state sequence of the next K time steps, where K is the number of prediction steps;

[0024] The prediction fusion unit is used to adaptively weight and fuse the prediction results of the physical prediction model and the data-driven prediction model to obtain the final prediction result.

[0025] The confidence interval calculation unit is used to calculate the confidence interval of the final prediction result based on the distribution of historical prediction errors.

[0026] Furthermore, in the data-driven prediction model unit:

[0027] The GAT layer models drone swarms as a graph. ,in For a collection of drone nodes, A set of communication links;

[0028] node exist The feature vector at time step is ;

[0029] The GAT layer calculates the attention weights between neighboring nodes using a multi-head attention mechanism, and then aggregates the features of the neighboring nodes in a weighted manner to output the aggregated features. , Represents a node exist The aggregation characteristics of time;

[0030] LSTM layers with historical time windows Aggregated feature sequences within Given the input, predict the state sequence for the next K time steps.

[0031] Furthermore, the bottleneck and risk assessment module includes:

[0032] The spatial conflict risk assessment unit is used to calculate the predicted distance between any two UAVs based on the predicted location, and to calculate the spatial conflict risk index based on a Gaussian function.

[0033] The task schedule risk assessment unit is used to calculate the task schedule risk index based on the deviation between the predicted task completion schedule and the target schedule.

[0034] The energy depletion risk assessment unit is used to calculate the energy depletion risk index based on the deviation between the predicted remaining energy and the safety threshold.

[0035] The comprehensive risk calculation unit is used to weight and sum the space conflict risk index, mission schedule risk index and energy depletion risk index to obtain a comprehensive risk index.

[0036] The reallocation trigger judgment unit is used to trigger proactive dynamic reallocation when there is a step k such that the comprehensive risk index exceeds a preset risk threshold. .

[0037] Furthermore, the proactive reallocation decision module includes:

[0038] The multi-objective optimization unit is used to establish a multi-objective optimization model with the optimization objectives of minimizing task completion time, maximizing energy consumption balance, balancing communication load, and minimizing redistribution disturbance.

[0039] The constraint processing unit is used to apply task coverage constraints, UAV capability constraints, energy constraints, and communication constraints to the multi-objective optimization model.

[0040] The solution unit is used to solve the multi-objective optimization model using the non-dominated sorting genetic algorithm NSGA-II to obtain the Pareto optimal solution set;

[0041] The robust processing unit is used to verify, by combining the predicted confidence interval with the Pareto optimal solution set and using the scenario sampling method, whether the scheme satisfies the safety constraints in all possible scenarios, and select the comprehensive optimal solution as the final redistribution scheme.

[0042] A method for autonomous and cooperative control of unmanned aerial vehicle (UAV) swarms includes the following steps:

[0043] S1. Real-time collection of status data and environmental data of each UAV, including the first... A drone Position vector at time , No. A drone velocity vector at time t , No. A drone Remaining energy at any given moment , No. A drone The current amount of work completed at any given moment , No. The drone and the first Between drones Communication link quality at any moment Environmental perception data;

[0044] S2. Perform filtering, noise reduction, spatiotemporal registration, and normalization on the collected data;

[0045] S3. Input the preprocessed data into the hybrid prediction model to predict the spatiotemporal evolution of the cluster over the next K time steps, where K is the number of prediction steps.

[0046] S4. Based on the prediction results of the spatiotemporal evolution, calculate the spatial conflict risk, mission progress risk, and energy depletion risk at each future moment, and calculate the comprehensive risk index.

[0047] S5. If there exists a step k such that the comprehensive risk index exceeds the preset risk threshold, then proactive reallocation is triggered, and the process proceeds to S6; otherwise, the process proceeds to S8. ;

[0048] S6. Based on the multi-objective optimization model and the NSGA-II algorithm, a redistribution scheme is generated, and the final scheme is determined after robust constraint verification.

[0049] S7. Distribute the redistribution scheme to each UAV and update the task queue and flight path of each UAV.

[0050] S8. Each UAV executes control commands according to its current task assignment.

[0051] S9. Collect actual execution effect data and calculate prediction error and redistribution effect indicators;

[0052] S10. Update the parameters of the hybrid prediction model and the multi-objective optimization model based on the actual execution effect data;

[0053] S11, return to S1, and enter the next control cycle.

[0054] Furthermore, in step S3:

[0055] The hybrid prediction model includes a physical prediction model and a data-driven prediction model. The physical prediction model makes predictions based on the kinematic equations of the UAV, and the data-driven prediction model makes predictions based on GAT+LSTM.

[0056] No. A drone in Final prediction result of position at time Fusion as follows:

[0057] ,

[0058] in The output of the physical prediction model A drone in Predicted location at time The first output of the data-driven prediction model A drone in Predicted location at time To predict the step size, ;in and For the weighting coefficients, satisfying ,and and It adaptively adjusts with the number of steps k.

[0059] Furthermore, in step S4:

[0060] Space Conflict Risk Index Calculate using the following formula:

[0061] ,

[0062] Among them, the predicted first The drone and the first A drone in Interval of time , Parameters that are sensitive to conflict;

[0063] Task progress risk index Calculate using the following formula:

[0064] ,

[0065] Among them, the predicted first The task is in Progress of completion at any given moment , To execute the first A collection of drones for a single mission. For the predicted first A drone in The amount of work completed at any given time. For the task Total workload For the task exist The target progress at any given time;

[0066] Energy depletion risk index Calculate using the following formula:

[0067] ,

[0068] Among them, the predicted first A drone in Remaining energy at any moment , For the predicted first The drone in Energy consumption per time step For the first Energy safety threshold for drones;

[0069] Comprehensive risk indicators Calculate using the following formula:

[0070] ,

[0071] in For the weighting coefficients, satisfying .

[0072] Furthermore, the multi-objective optimization model described in step S6 is as follows:

[0073] , For task completion time, As an energy consumption balance indicator, For communication load, For redistribution disturbances;

[0074] in This is the redistribution scheme matrix. Indicates the first Are drones assigned to missions? , For the first The drone completed its first The predicted time required for each task. For the first drones in the redistribution scheme Predicted energy consumption This represents the average energy consumption. In the redistribution scheme Downlink Communication load, This represents the current allocation status, where N is the total number of drones. This represents the total number of tasks.

[0075] Furthermore, the robust constraint verification in step S6 includes:

[0076] Considering the uncertainty of prediction, the redistribution scheme must ensure that the safety constraints are met in all possible scenarios within the prediction confidence interval. The safety constraints include at least: the distance between any two drones is greater than or equal to the preset minimum safety distance; and the remaining energy of each drone is greater than or equal to the energy that needs to be reserved.

[0077] By sampling from the confidence interval The scenario requires the solution to satisfy the security constraints in all sampled scenarios.

[0078] The beneficial effects of this invention are:

[0079] (1) Paradigm shift from “passive response” to “proactive pre-adjustment”: By adjusting in advance based on the prediction results before the problem occurs, task delays and resource conflicts can be avoided.

[0080] (2) High prediction accuracy: The hybrid architecture of the physical prediction model and the data-driven prediction model takes into account both the interpretability of physical constraints and the adaptability of data-driven approaches. This application has been verified by a large number of simulation experiments. The average position error of the hybrid prediction model within a 20-step prediction range is reduced by about 35% compared with the pure physical model and by about 22% compared with the pure data-driven model.

[0081] (3) Strong forward-looking decision-making: Multi-step prediction enables the redistribution decision to "see" the situation evolution at multiple future time steps, avoiding short-sighted decision-making. In simulation comparison, the redistribution scheme of the present invention has an overall performance improvement of about 28% over the optimization scheme based on the current time step within a 30-step prediction window.

[0082] (4) Good robustness: Introducing prediction confidence intervals and adopting a conservative redistribution strategy under the condition of prediction uncertainty to ensure the robustness of the system.

[0083] (5) Continuous evolution capability: The closed-loop feedback mechanism enables the system to continuously learn from actual execution and continuously improve its prediction and decision-making capabilities. Attached Figure Description

[0084] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0085] Figure 1 This is a system architecture diagram of an autonomous collaborative control system for unmanned aerial vehicle (UAV) swarms as described in an embodiment of the present invention;

[0086] Figure 2 This is a flowchart of the spatiotemporal evolution prediction process described in an embodiment of the present invention;

[0087] Figure 3 This is a flowchart of the proactive dynamic reallocation decision-making process described in an embodiment of the present invention;

[0088] Figure 4 This is a flowchart of an autonomous collaborative control method for unmanned aerial vehicle (UAV) swarms according to an embodiment of the present invention;

[0089] Figure 5A This is a simulation verification diagram of the initial task allocation and flight trajectory of the UAV swarm described in this embodiment of the invention;

[0090] Figure 5B1 This is a comparison chart of the spatiotemporal evolution prediction and actual trajectory of the UAV-1 in the simulation verification described in this embodiment of the invention;

[0091] Figure 5B2 This is a comparison chart of the spatiotemporal evolution prediction and actual trajectory of the UAV-3 in the simulation verification described in this embodiment of the invention;

[0092] Figure 5B3 This is a comparison chart of the spatiotemporal evolution prediction and actual trajectory of the UAV-5 drone in the simulation verification described in this embodiment of the invention;

[0093] Figure 5B4 This is a comparison chart of the spatiotemporal evolution prediction and actual trajectory of the UAV-7 in the simulation verification described in this embodiment of the invention;

[0094] Figure 5C This is a graph showing the evolution of the comprehensive risk index over time in the simulation verification described in this embodiment of the invention;

[0095] Figure 5D1 This is a bar chart comparing the completion times of proactive redistribution versus reactive redistribution tasks in the simulation verification described in this embodiment of the invention.

[0096] Figure 5D2 This is a line graph comparing the cumulative completion time of proactive redistribution versus reactive redistribution in the simulation verification described in this embodiment of the invention.

[0097] Figure 5E(a) shows the task completion time as described in the embodiment of the present invention. For the horizontal axis, energy consumption balance A schematic diagram of the Pareto front along the vertical axis;

[0098] Figure 5E(b) shows the communication load according to an embodiment of the present invention. With redistribution disturbance A schematic diagram of the solution distribution;

[0099] Figure 5E(c) shows the four objective functions described in the embodiments of the present invention. Completion time Energy balance, Communication load, Radar chart of the redistribution disturbance;

[0100] Figure 5E(d) shows the NSGA-II algorithm described in this embodiment of the invention during 50 iterations. and The convergence curve. Detailed Implementation

[0101] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0102] Example 1 (System Architecture):

[0103] like Figure 1 As shown, this application discloses an autonomous collaborative control system for unmanned aerial vehicle (UAV) swarms. The system consists of the following modules: a data acquisition module 101, a data preprocessing module 102, a spatiotemporal evolution prediction module 103, a bottleneck and risk assessment module 104, a proactive redistribution decision module 105, an execution control module 106, and a feedback and online learning module 107.

[0104] The connection relationships of each module are as follows: the output of the data acquisition module 101 is connected to the input of the data preprocessing module 102; the output of the data preprocessing module 102 is connected to the input of the spatiotemporal evolution prediction module 103; the output of the spatiotemporal evolution prediction module 103 is connected to the input of the bottleneck and risk assessment module 104; the output of the bottleneck and risk assessment module 104 is connected to the input of the proactive reallocation decision module 105; the output of the proactive reallocation decision module 105 is connected to the input of the execution control module 106; the actual execution effect of the execution control module 106 is fed back to the spatiotemporal evolution prediction module 103 and the proactive reallocation decision module 105 through the feedback and online learning module 107, forming a closed loop.

[0105] Example 2 (Data Acquisition and Preprocessing):

[0106] The data collection in this application includes the following:

[0107] With sampling period s collects the following data:

[0108] (1) No. A drone Position at any given moment (3D position vector) ( , , These represent the X-axis geodetic coordinates, Y-axis geodetic coordinates, and Z-axis geodetic coordinates, respectively (unit: m).

[0109] (2) No. A drone Velocity at any given moment (three-dimensional velocity vector) ( , , These represent the velocity components in the X, Y, and Z directions, respectively (unit: m / s).

[0110] (3) No. A drone Remaining energy at any given moment (Unit: J);

[0111] (4) No. A drone The current amount of work completed at any given moment (Units are related to specific tasks, such as search area in m², delivery of supplies in kg, etc.);

[0112] (5) No. The drone and the first Between drones Communication link quality at any moment ;

[0113] (6) Environmental perception data (location of obstacles, weather conditions, etc.).

[0114] The data preprocessing in this application includes the following:

[0115] (1) Filtering and noise reduction: Kalman filtering is used to smooth the position and velocity data and remove sensor measurement noise;

[0116] (2) Spatiotemporal registration: unifying data from different sensors into the same spatiotemporal coordinate system;

[0117] (3) Normalization: Normalize each feature data to Interval:

[0118]

[0119] in The original data values, and These are the minimum and maximum values ​​of this feature in historical data, respectively.

[0120] Example 3 (Spatiotemporal Evolution Prediction Model):

[0121] like Figure 2 As shown, the spatiotemporal evolution prediction model of this invention adopts a hybrid architecture of a physical prediction model and a data-driven prediction model. The design concept is as follows: the physical prediction model provides a prediction benchmark with clear physical meaning based on the kinematic equations of the UAV, ensuring the interpretability of the prediction and its rationality under extreme conditions; the data-driven prediction model learns complex environmental coupling relationships and swarm interaction patterns from historical data based on GAT+LSTM, compensating for the parts that the physical model struggles to accurately model. The two complement each other, achieving high-precision and high-reliability predictions. Specifically, it includes the following:

[0122] 1. Physical prediction model (based on UAV kinematics)

[0123] For the The discrete-time kinematic model of the drone is as follows:

[0124]

[0125]

[0126] in For the first A drone in The acceleration at any given moment (which can be obtained through the inertial measurement unit or calculated from control commands). To predict the step size (preferably 2 seconds). express The position of the UAV was calculated by the kinematic model. express The speed of the drone was calculated from the kinematic model.

[0127] Considering the motion constraints of the drone:

[0128]

[0129] in The maximum speed is determined by the drone model; for example, a certain type of quadcopter has a speed of 15 m / s. This is the maximum acceleration (e.g., 5 m / s² for a certain type of quadcopter).

[0130] The physical prediction model outputs the first A drone in Predicted location at time and prediction speed ,in , The prediction steps are set (preferably 30 steps, corresponding to a 60-second prediction window).

[0131] 2. Data-driven prediction model (GAT+LSTM)

[0132] (21) Graph Attention Network (GAT) layer.

[0133] Modeling drone swarms as graphs ,in Let N be the set of drone nodes, and N be the total number of drones. For the set of communication links (when It is assumed that an edge exists. (Indicates the communication quality threshold).

[0134] node exist The feature vector at time step is:

[0135]

[0136] The attention coefficient of the GAT layer is calculated as follows: First, calculate the node... For neighboring nodes Unnormalized attention score:

[0137]

[0138] in A learnable linear transformation matrix ( For the input feature dimension, For the transformed dimension, the preferred dimension is... ), For learnable attention vectors, This represents a vector concatenation operation. The above formula means: concatenate the nodes... and nodes The transformed features are concatenated into one dimensional vector, then combined with attention vector Perform the inner product, and finally obtain the scalar attention score through the LeakyReLU activation function.

[0139] Softmax normalization of the attention score:

[0140]

[0141] in For nodes The set of neighboring nodes (including itself). For nodes For nodes Attention weights, satisfying .

[0142] node The aggregation characteristics are:

[0143]

[0144] in This is the ReLU activation function. The formula means: the transformed features of each neighbor node are weighted and summed according to attention weights, and then the aggregated features are obtained through the activation function. The purpose of introducing the GAT layer is to allow the model to automatically learn which neighboring drones' state information is more important during prediction (e.g., drones that are close by are more important for predicting spatial conflicts, and drones with high task relevance are more important for predicting task progress).

[0145] (22) LSTM time series prediction layer.

[0146] Feature sequences after GAT aggregation Input LSTM network to predict the future The state at each time step :

[0147]

[0148] in The historical time window length (preferably 20 steps, corresponding to 20 seconds of historical data) includes the predicted UAV spatial position, speed, remaining energy, and completed task workload.

[0149] The core calculation formulas for the LSTM unit are as follows:

[0150] Forget Gate (determines which information to discard from the previous cell state):

[0151]

[0152] Input gate (determines which new information is stored in the cell state):

[0153]

[0154]

[0155] Cell state update (from the previous state to the current state):

[0156]

[0157] Output gate (determines what information is output):

[0158]

[0159]

[0160] in for The input feature vector at time t, for The hidden state at all times for Cellular state at any given moment The weight matrix is ​​a learnable matrix. For bias vectors, It is the Sigmoid activation function. The hyperbolic tangent activation function is used. This indicates element-wise multiplication.

[0161] 3. Predicting fusion and confidence intervals

[0162] The results predicted by the physical prediction model Results of data-driven prediction models Perform adaptive weighted fusion, where Output the first data-driven prediction model A drone in Predicted location at time:

[0163]

[0164] in and These are weighting coefficients, which vary with the number of steps. Adaptive adjustment .

[0165]

[0166] in Adjustment coefficient (preferred) The engineering basis for this design is that: in short-term prediction, the physical prediction model based on the extrapolation of the current motion state is relatively accurate; in long-term prediction, the accumulated physical modeling error increases, while the prediction based on historical pattern learning by the data-driven prediction model is relatively more reliable.

[0167] The prediction confidence interval is calculated based on the historical distribution of prediction errors. Let the most recent... The preferred time period The prediction error is Then the predicted standard deviation estimate is:

[0168]

[0169] in Let be the mean of historical prediction errors. Then the confidence interval is... for:

[0170]

[0171] in Confidence level (Preferred 95%) corresponds to the standard normal distribution quantile ( ).

[0172] The spatiotemporal evolution state of the cluster is defined as the set of predicted states of all UAVs (i.e., the spatiotemporal evolution prediction results):

[0173]

[0174] in For the predicted number in the set A drone in Location at any given moment For the predicted number in the set A drone in The speed of time For the predicted number in the set A drone in Remaining energy at any given moment For the predicted number in the set A drone in The amount of work completed at any given time.

[0175] Example 4 (Bottleneck and Risk Assessment):

[0176] Based on spatiotemporal evolution prediction results We will conduct a forward-looking assessment of the following three types of risks: space conflict risk, mission schedule risk, and energy depletion risk.

[0177] (1) Space conflict risk

[0178] Predicting the first The drone and the first A drone in Interval of time:

[0179]

[0180] The spatial conflict risk index is defined using a Gaussian function:

[0181]

[0182] in For conflict-sensitive parameters (preferred) The value of m is based on the minimum safe distance specification of 10m for a certain type of UAV. Half of the safe distance is taken as the sensitive parameter, so that the risk index increases significantly when the distance is less than 10m. The property of this function is: when the distance is 0, the risk index is 1 (complete overlap); when the distance is much greater than 10m, the risk index increases significantly. The risk index approaches 0 (no risk).

[0183] (2) Task schedule risk

[0184] Predicting the first The task is in Progress of completion at any given moment:

[0185]

[0186] in To execute the first A collection of drones for a single mission. For the predicted first A drone in The amount of work completed at any given time (and) (Dimensional consistency) For the task The total workload. The meaning of this formula: task The completion progress is the percentage of the total workload completed by all drones performing the task.

[0187] Task progress risk index:

[0188]

[0189] in For the task exist The target progress at any given time (obtained by linear interpolation based on the task time schedule).

[0190] (3) Risk of energy depletion.

[0191] Predicting the first A drone in Remaining energy at any given moment:

[0192]

[0193] in For the predicted first The drone in Energy consumption per time step (estimated based on current flight status and predicted mission payload).

[0194] Energy depletion risk index:

[0195]

[0196] in For the first The energy safety threshold for a drone (usually 20% to 30% of full charge to ensure the drone has enough energy to return home).

[0197] (4) Comprehensive risk indicators

[0198] Comprehensive risk indicators:

[0199]

[0200] in Let be the weighting coefficient, satisfying The preferred value is (Space conflicts are directly related to flight safety and have the highest priority.) , .

[0201] Reallocation trigger condition: When any Make At that time, proactive dynamic reallocation is triggered. Preset risk threshold (preferred) This value is determined through Monte Carlo simulation, which achieves a good balance between false alarm rate and missed alarm rate.

[0202] Example 5 (Proactive Dynamic Reallocation Decision):

[0203] like Figure 3 As shown, the redistribution decision-making process of the present invention is as follows. The essential difference from the prior art (which passively responds after a problem occurs) is that the present invention triggers redistribution before the problem actually occurs, based on the prediction and evaluation results.

[0204] 1. Multi-objective optimization model

[0205] The redistribution problem is modeled as a multi-objective optimization problem:

[0206]

[0207] in This is the redistribution scheme matrix. Indicates the first Are drones assigned to missions? .

[0208] The objective functions are defined as follows:

[0209] (1) Minimize task completion time:

[0210]

[0211] in For the first The drone completed its first The predicted time required for each task (estimated based on the drone's current position, speed, and task location). This represents the maximum completion time among all tasks (i.e., makespan). Minimization means making the latest completion time of the entire task set as early as possible.

[0212] (2) Maximizing energy consumption balance (expressed as minimizing energy consumption variance):

[0213]

[0214] in For the first drones in the redistribution scheme Predicted energy consumption This represents the average energy consumption. The smaller the value, the more balanced the energy consumption.

[0215] (3) Communication load balancing:

[0216]

[0217] in In the redistribution scheme Downlink The communication load (the amount of data that needs to be transmitted per unit of time). This represents the peak communication load across all links in the entire network. The smaller the value, the more balanced the communication load and the lower the risk of congestion.

[0218] (4) Minimize redistribution disturbances (minimize interference with ongoing tasks):

[0219]

[0220] in This represents the current allocation status. This indicates the magnitude of the change in the plan; the smaller the value, the less interference it will cause to the ongoing task. N represents the total number of drones. This represents the total number of tasks.

[0221] 2. Constraint Adjustment

[0222] (1) Task coverage constraint: Each task must be assigned to at least one UAV.

[0223]

[0224] (2) Unmanned aerial vehicle (UAV) capability constraints: Each UAV may perform at most one primary mission.

[0225]

[0226] (3) Energy constraint:

[0227]

[0228] in For the first Reserved energy for the drone (usually 25% of the full charge to ensure a safe return).

[0229] (4) Communication constraints:

[0230]

[0231] in Communication quality threshold (preferred) (corresponding to a signal strength of -80dBm or higher).

[0232] 3. Solution Algorithm

[0233] The above multi-objective optimization problem is solved using an improved non-dominated sorting genetic algorithm (NSGA-II):

[0234] Step 1: Initialize the population Each individual represents a redistribution scheme matrix. The encoding method uses binary encoding, which will... The allocation matrix is ​​expanded row-wise to a length of The binary string. Population size. .

[0235] Step 2: Perform non-dominated sorting on the population and calculate the crowding distance for each individual.

[0236] Step 3: Select via tournament (tournament size 2), simulate binary crossover (crossover probability) Distribution index ) and polynomial mutation (probability of mutation) Distribution index Generate offspring population .

[0237] Step 4: Merge the parent and offspring populations The next generation of the population is generated by performing non-dominated sorting and crowding selection. .

[0238] Step 5: Iterate until the maximum number of algebras is reached. Or, the Pareto frontier no longer improves after 20 consecutive generations (convergence criterion).

[0239] Step Six: From the Pareto optimal solution set, select the comprehensive optimal solution using the ideal point distance method: calculate the distance from each solution to the ideal point. The solution with the minimum normalized Euclidean distance is selected as the final redistribution scheme.

[0240] 4. Robust handling of prediction uncertainty

[0241] Considering the uncertainty of prediction, robust constraints are introduced into the optimization:

[0242]

[0243] The reallocation scheme must ensure that the safety constraints are met under all possible scenarios within the prediction confidence interval:

[0244]

[0245] in Indicates in the scene The following plan Do all safety constraints (spatial spacing) meet? ,energy wait).

[0246] To reduce computational complexity, this invention employs a scenario sampling method to approximate the robust constraints described above: random sampling is performed within the confidence interval. The scenario requires the solution to meet safety constraints in all sampling scenarios.

[0247] Example 6 (Closed-Loop Adaptive Control Flow):

[0248] like Figure 4 As shown, the closed-loop adaptive control flow of the UAV swarm autonomous collaborative control system of the present invention is as follows:

[0249] Step 1 (Data Acquisition): Collect real-time status data (position, speed, energy, mission status) and environmental data of each UAV at a 1-second interval.

[0250] Step 2 (Data Preprocessing): Perform Kalman filtering, spatiotemporal registration, and normalization on the collected data.

[0251] Step 3 (Spatiotemporal Evolution Prediction): Input the preprocessed data into the hybrid prediction model (physical model + GAT + LSTM) to predict the future. The spatiotemporal evolution of the cluster at each time step (60 seconds in total) .

[0252] Step 4 (Bottleneck and Risk Assessment): Based on the prediction results, calculate the spatial conflict risk, mission schedule risk, and energy depletion risk using formulas (6.4.1-2), (6.4.2-2), and (6.4.3-2) respectively, and then calculate the comprehensive risk index using formula (6.4.4-1). .

[0253] Step 5 (Reassignment Judgment): If it exists Make If the current allocation is not triggered, proceed to step six to trigger proactive reallocation; otherwise, proceed to step eight to maintain the current allocation and continue execution.

[0254] Step 6 (Proactive Reassignment): Based on the multi-objective optimization model (Equations 6.5.1-1 to 6.5.2-4), the NSGA-II algorithm (Steps 1 to 6 in Example 5) is used to generate a reassignment scheme, which is then verified by robust constraints (Equations 6.5.4-1 to 6.5.4-2).

[0255] Step 7 (Execute Reassignment): Distribute the reassignment plan to each drone and update the task queue and flight path of each drone. Completion Time Target: The time from triggering the reassignment to completing the distribution of all drone commands should be controlled within 3 seconds.

[0256] Step 8 (Execution Control): Each UAV executes control commands according to its current task assignment.

[0257] Step Nine (Effect Evaluation and Feedback): Collect actual execution effect data, calculate prediction error (deviation between predicted location and actual location) and redistribution effect indicators (actual task completion time, actual energy consumption, etc. after redistribution).

[0258] Step 10 (Online Learning Update): Use real data to update the parameters of the hybrid prediction model (LSTM is fine-tuned using online gradient descent, learning rate...). ) and redistribution strategy parameters (adjusting multi-objective optimization weights based on redistribution effect metrics) ).

[0259] Step 11 (Iterative Loop): Return to Step 1 and enter the next control cycle (cycle is 1 second).

[0260] Based on the above steps, to facilitate understanding of the specific implementation process of the present invention, a simplified numerical example is given below:

[0261] Scenario setting: Three drones (UAV1, UAV2, UAV3) are collaboratively performing search missions in two areas (Mission A and Mission B). Current time... The status of each drone is shown in Table 0 below:

[0262] Table 0: Current Time Status of each drone at that time

[0263] ;

[0264] Step 3 Prediction: The hybrid prediction model predicts the next 60 seconds ( The situation of s). The prediction results show:

[0265] UAV3 in At time s, the remaining energy will drop to 22%, which is below the safety threshold of 25%.

[0266] UAV1 and UAV2 will overlap in the mission area, with a predicted minimum spacing of only 8m (below the safe spacing of 10m).

[0267] Step Four Assessment: In (correspond When s), .exist (correspond When s), Exceeding the threshold .

[0268] Step 5: Determine if proactive reallocation is triggered (at this time, the actual time...) (50 seconds remain until the predicted risk occurs).

[0269] Step Six: Reassignment: The reassignment scheme generated by multi-objective optimization is as follows: Task B is transferred from UAV3 to UAV2 (UAV2 will switch to executing Task B after completing Task A), and UAV3 is reassigned to the less loaded sub-region of Task A. This scheme reduces the predicted energy consumption of UAV-3 by 30%. The remaining energy prediction value at time s increased to 28% (above the safety threshold), and the distance between drones in the mission area increased to 18m, eliminating the risk of spatial conflict.

[0270] Steps seven and eight are executed: The reallocation is completed and the plan is immediately executed, with each drone adjusting its flight path according to the new plan.

[0271] Effect Comparison: If the existing "reactive" redistribution technology is used, it will take until... The reassignment was triggered only after UAV3 issued a low battery warning at time s. By then, UAV3 was no longer able to complete the remaining part of Task B, and other drones needed to rush to the Task B area from their current location, adding approximately 90 seconds to the process. This invention completes the reassignment 50 seconds earlier, avoiding mission delays.

[0272] Example 7 (Simulation Verification and Effect Comparison):

[0273] To verify the effectiveness of the proposed method, simulations were performed on relevant data based on practical application experience. The simulation scenario included eight UAVs (UAV-1 to UAV-8) and four heterogeneous missions (mission 1 to mission 4), with the mission area being a 500m × 500m two-dimensional plane. The initial position, velocity, remaining energy, mission allocation, and mission location of the UAVs were all randomly generated. The simulation was performed using the spatiotemporal evolution prediction model (physical model + GAT + LSTM hybrid prediction), comprehensive risk assessment mechanism (spatial conflict / mission progress / energy depletion), and proactive dynamic reallocation method (multi-objective optimization + NSGA-II + robust verification) proposed in this invention. The simulation parameters were consistent with the preferred configuration in Table 1, specifically including: sampling period. Predicting step size Prediction steps K=30, risk threshold The multi-objective optimization weights are β1=0.4, β2=0.3, and β3=0.3.

[0274] Table 1: Preferred parameter configurations for each module of the present invention

[0275] ;

[0276] The results of the five sets of simulation experiments are given below.

[0277] (1) Initial task allocation and flight trajectory ( Figure 5A )

[0278] Figure 5A The diagram illustrates the initial task allocation scheme and flight trajectories of eight UAVs during the simulation period. In the diagram, eight different colored trajectory lines correspond to the movement paths of the eight UAVs (UAV-1 to UAV-8) from their starting positions to their current positions; the circular starting point marker "○" indicates the starting position, and the square ending point marker "□" indicates the current position; red star "★" marks the spatial locations of four task target points, and a red dashed circle represents the task's influence range (radius 20m). Priority levels (1-5, with level 5 being the highest) are indicated next to each task.

[0279] from Figure 5A It can be seen that the initial allocation of UAVs was as follows: UAV-1, UAV-2, UAV-3, and UAV-4 to perform Task 1; UAV-5 and UAV-6 to perform Task 2; UAV-7 to perform Task 3; and UAV-8 to perform Task 4. The distribution of UAV trajectories is reasonable, but some UAVs (such as UAV-3 and UAV-4) show a tendency for their trajectories to intersect, indicating potential future spatial conflicts. Simultaneously, the density of UAVs in Task 1 is high (4 aircraft), while Task 3 and Task 4 each have only 1 aircraft, posing a potential risk of unbalanced task load. These phenomena validate the necessity of this invention—further prediction and reallocation are needed to avoid potential conflicts and optimize load balancing.

[0280] (2) Verification of spatiotemporal evolution prediction effect ( Figures 5B1-5B4 )

[0281] Figures 5B1-5B4 The results demonstrate the trajectory prediction performance of a hybrid prediction model (physical model + GAT + LSTM) on four representative UAVs (UAV-1, UAV-3, UAV-5, and UAV-7). In the figure, the solid blue line represents the actual flight trajectory of the UAV, the dashed red line represents the predicted trajectory output by the hybrid prediction model (prediction step size 10 steps, i.e., 20 seconds), the semi-transparent red circle represents the prediction confidence interval (CI, corresponding to the 95% confidence level, formula 6.3.3-2), and the green diamond "◇" marks the prediction starting point.

[0282] from Figures 5B1-5B4 It can be seen that the predicted trajectory (red dashed line) of the hybrid prediction model generally matches the actual trajectory (blue solid line) well, with an average position error of 8.7m within a 10-step prediction range. The confidence interval (red semi-transparent circle) expands with increasing prediction step size, reflecting the accumulation of prediction uncertainty, consistent with the theoretical expectation of formula 6.3.3-2. The prediction accuracy varies among different UAVs: the predicted trajectories of UAV-1 and UAV-5 match the actual trajectories well (error <10m), while the predicted trajectory of UAV-7 deviates slightly (error approximately 12m), mainly due to the larger maneuverability of this UAV during its mission. The average position error of the pure physics model (without GAT and LSTM corrections) on the same dataset is 13.2m. The hybrid prediction model of this invention reduces the prediction error by 34.1% (from 13.2m to 8.7m), verifying the effectiveness of the hybrid architecture of the physics model and the data-driven model.

[0283] The above results directly support the claim that "the prediction accuracy of the hybrid prediction model is better than that of the single model" in the beneficial effects of this invention, and also verify the practical feasibility of the spatiotemporal evolution prediction method constructed by formulas (6.3.1-1) to (6.3.3-3).

[0284] (3) Evolution of comprehensive risk indicators and timing of redistribution ( Figure 5C )

[0285] Figure 5C This demonstrates the risk of space conflict throughout the entire simulation period (60 seconds, 30 time steps). Task progress risk Risk of energy depletion and comprehensive risks The evolution curve over time. The red dashed line in the graph marks the risk threshold. The purple dashed line marks the moment when the overall risk first exceeds the threshold. The trigger point is marked with a purple diamond "◆".

[0286] from Figure 5C It can be seen that in the initial stage of simulation (0s~20s), the overall risk... The level remains low (0.2~0.4), with the main risk source being spatial conflict risk (the gradual formation of dense drone areas). In the 20s~30s range, there is a risk of energy depletion. The increase is due to the accelerated energy consumption of some drones (UAV-3, UAV-6), leading to a comprehensive risk. It then climbed.

[0287] exist At that time, comprehensive risk First time exceeding the threshold This triggers proactive dynamic reallocation (marked in purple). At this point: spatial conflict risk. (The distance between UAV-3 and UAV-4 is less than the safe distance); mission progress risk. (Task 1 is behind schedule); Risk of energy depletion (UAV-3's remaining energy prediction is 22%, below the safety threshold of 25%). After triggering redistribution (38s~60s), the overall risk... The value gradually fell below 0.45, indicating that the reallocation effectively mitigated various risks.

[0288] The above results verify the effectiveness of the risk assessment mechanism of this invention (Formulas 6.4.1-2, 6.4.2-2, 6.4.3-2, 6.4.4-1). Its core value lies in: At that time, spatial conflicts and energy depletion problems had not yet actually occurred, but the present invention identified and triggered redistribution in advance based on the prediction results, which reflects the essential difference between "proactive" control and existing "reactive" control.

[0289] (4) Comparison of the effects of proactive redistribution and reactive redistribution ( Figures 5D1-5D2 )

[0290] Figures 5D1-5D2 A quantitative comparison was made between the proactive redistribution method of the present invention and the reactive redistribution method in the prior art. Proactive redistribution is the method of the present invention—in Time-based (based on predicted risk exceeding a threshold) redistribution is triggered in advance; reactive redistribution is the existing technology method—redistribution is passively triggered after the drone's energy is below 15% or after a conflict actually occurs.

[0291] Figure 5D1 The task completion times for each of the eight drones under the two methods are shown in the bar chart (blue represents the present invention, and red represents the prior art). Figure 5D2 The cumulative completion time curve is sorted by completion time (blue represents the present invention, and red represents the prior art).

[0292] from Figures 5D1-5D2 As can be seen, at the individual drone level, the task completion time of 7 out of 8 drones under proactive redistribution was better than that under reactive redistribution, with an average reduction of 42 seconds. UAV-3 and UAV-6 showed the most significant improvements (reduced by 68 seconds and 72 seconds respectively), because these two drones were in a low-energy state before redistribution, and proactive redistribution transferred their load to drones with sufficient energy (UAV-2 and UAV-8) in advance. At the cluster level, the cumulative completion time curve shows that the total task completion time of this invention was 486 seconds, while reactive redistribution was 681 seconds, representing a 28.6% reduction in total completion time.

[0293] The above results directly verify the significant advantages of the proactive dynamic redistribution method (multi-objective optimization + NSGA-II + robust verification) of this invention compared to existing reactive redistribution methods. The core conclusion is that this invention achieves a nearly 30% improvement in task efficiency through "advance prediction and advance adjustment," while reactive methods suffer significant efficiency losses due to decision-making lag caused by "post-event response."

[0294] (5) Multi-objective optimization Pareto front and NSGA-II convergence

[0295] Figures 5E(a)-5E(d) The optimization effects of the multi-objective optimization model (Formulas 6.5.1-1 to 6.5.1-5) and the NSGA-II solution algorithm (steps one to six in Example 5) of this invention are demonstrated.

[0296] Figure 5E(a) shows the task completion time. For the horizontal axis, energy consumption balance The vertical axis represents the Pareto front. Red dots represent Pareto optimal solutions, and gray dots represent non-dominated solutions. The red dashed line connecting the Pareto optimal solutions forms the Pareto front. The clear distribution of the Pareto front indicates that the NSGA-II algorithm effectively searches the trade-off boundary of the multi-objective optimization problem.

[0297] Figure 5E(b) shows the communication load. With redistribution disturbance The solution distribution is shown, with the ideal point marked by a star "★". and (Simultaneously taking the theoretical point of minimum). The solution distribution is concentrated near the ideal point, indicating that the NSGA-II algorithm is... - It also has good search performance in the target space.

[0298] Figure 5E(c) shows four objective functions ( Completion time Energy balance, Communication load, A radar chart of the redistribution perturbation was used to compare representative solutions with three different optimization preferences: "Minimum" "Preference (blue, focusing on completion time)," "Minimum" "Preference" (green, focusing on energy consumption balance) and "Trade-off Solution" (red, the optimal solution for all four objectives). The radar chart visually illustrates the trade-offs between objectives under different optimization preferences.

[0299] Figure 5E(d) shows the NSGA-II algorithm during 50 iterations. and The convergence curve. It converged rapidly from approximately 160 to approximately 35. The convergence rate decreased from approximately 85 to approximately 25, and both values ​​stabilized after approximately 30 generations, verifying the convergence speed and stability of the NSGA-II algorithm in solving the multi-objective optimization problem of this invention.

[0300] The above results verify the effectiveness and engineering feasibility of the multi-objective optimization model and NSGA-II solution algorithm of this application.

[0301] Simulation conclusions

[0302] Based on the above five sets of simulation experiments, the following conclusions can be drawn:

[0303] (1) The spatiotemporal evolution prediction method (physical model + GAT + LSTM hybrid prediction) of the present invention can accurately predict the spatiotemporal evolution of the UAV cluster in the next 20 seconds, with an average position error of 8.7m, which is 34.1% lower than that of the pure physical model, thus verifying the effectiveness of the hybrid prediction model of the present application.

[0304] (2) The comprehensive risk assessment mechanism of this invention (formulas 6.4.1-2, 6.4.2-2, 6.4.3-2, 6.4.4-1) can identify in advance the risks of future spatial conflicts, mission delays, and energy depletion. The proactive reallocation was successfully triggered, verifying the effectiveness of the risk assessment mechanism of this invention.

[0305] (3) Compared with the existing reactive redistribution method, the proactive redistribution method of the present invention reduces the total task completion time by 28.6%, which verifies the significant advantages of the redistribution method of the present invention.

[0306] (4) The multi-objective optimization model and NSGA-II solution algorithm of the present invention successfully searched the Pareto optimal solution set, and the algorithm converged after about 30 generations, which verified the effectiveness of the multi-objective optimization model and the related optimization methods of the NSGA-II solution algorithm.

[0307] In summary, the three-layer architecture of "spatiotemporal evolution prediction + proactive dynamic redistribution + closed-loop adaptive optimization" proposed in this invention has demonstrated good prediction accuracy, risk identification capability and redistribution optimization effect in the simulation environment, verifying the technical feasibility and effectiveness of this invention.

[0308] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An autonomous collaborative control system for unmanned aerial vehicle (UAV) swarms, characterized in that, include: The module includes a data acquisition module (101), a data preprocessing module (102), a spatiotemporal evolution prediction module (103), a bottleneck and risk assessment module (104), a proactive reallocation decision module (105), an execution control module (106), and a feedback and online learning module (107). The data acquisition module (101) is used to collect multi-dimensional status data of each UAV in the cluster in real time, including the first... A drone Position vector at time , No. A drone velocity vector at time t , No. A drone Remaining energy at any given moment , No. A drone The current amount of work completed at any given moment , No. The drone and the first Between drones Communication link quality at any moment Environmental perception data; The data preprocessing module (102) is used to perform filtering, noise reduction, spatiotemporal registration and normalization on the collected data; The spatiotemporal evolution prediction module (103) is used to predict the spatiotemporal evolution of the cluster at multiple future time steps based on preprocessed data and through a hybrid architecture of physical prediction model and data-driven prediction model. The bottleneck and risk assessment module (104) is used to prospectively assess the risks at future moments based on the prediction results of the spatiotemporal evolution situation. The risks include spatial conflict risk, mission progress risk and energy depletion risk. The proactive reallocation decision module (105) is used to generate a reallocation scheme before the actual occurrence of the risk when the risk assessment result meets the triggering conditions; The execution control module (106) is used to execute the redistribution scheme and update the task queues and tracks of each UAV; The feedback and online learning module (107) is used to collect actual execution effect data, calculate prediction error and redistribution effect index, and update the parameters of the spatiotemporal evolution prediction module (103) and the proactive redistribution decision module (105) accordingly.

2. The unmanned aerial vehicle (UAV) swarm autonomous collaborative control system according to claim 1, characterized in that, The spatiotemporal evolution prediction module (103) includes: The physical prediction model unit is used to calculate the predicted position and velocity for the next K time steps based on the kinematic equations and motion constraints of the UAV. The data-driven prediction model unit is used to learn cluster interaction patterns from historical time series data based on graph attention network GAT and long short-term memory network LSTM, and predict the state sequence of the next K time steps, where K is the number of prediction steps; The prediction fusion unit is used to adaptively weight and fuse the prediction results of the physical prediction model and the data-driven prediction model to obtain the final prediction result. The confidence interval calculation unit is used to calculate the confidence interval of the final prediction result based on the distribution of historical prediction errors.

3. The unmanned aerial vehicle (UAV) swarm autonomous collaborative control system according to claim 2, characterized in that, In the data-driven prediction model unit: The GAT layer models drone swarms as a graph. ,in For a collection of drone nodes, A set of communication links; node exist The feature vector at time t is ; The GAT layer calculates the attention weights between neighboring nodes using a multi-head attention mechanism, and then aggregates the features of the neighboring nodes in a weighted manner to output the aggregated features. , Represents a node exist The aggregation characteristics of time; LSTM layers with historical time windows Aggregated feature sequences within Given the input, predict the state sequence for the next K time steps.

4. The unmanned aerial vehicle (UAV) swarm autonomous collaborative control system according to claim 3, characterized in that, The bottleneck and risk assessment module (104) includes: The spatial conflict risk assessment unit is used to calculate the predicted distance between any two UAVs based on the predicted location, and to calculate the spatial conflict risk index based on a Gaussian function. The task schedule risk assessment unit is used to calculate the task schedule risk index based on the deviation between the predicted task completion schedule and the target schedule. The energy depletion risk assessment unit is used to calculate the energy depletion risk index based on the deviation between the predicted remaining energy and the safety threshold. The comprehensive risk calculation unit is used to weight and sum the space conflict risk index, mission schedule risk index and energy depletion risk index to obtain a comprehensive risk index. The reallocation trigger judgment unit is used to trigger proactive dynamic reallocation when there is a step k such that the comprehensive risk index exceeds a preset risk threshold. .

5. The unmanned aerial vehicle (UAV) swarm autonomous collaborative control system according to claim 1, characterized in that, The proactive reallocation decision module (105) includes: The multi-objective optimization unit is used to establish a multi-objective optimization model with the optimization objectives of minimizing task completion time, maximizing energy consumption balance, balancing communication load, and minimizing redistribution disturbance. The constraint processing unit is used to apply task coverage constraints, UAV capability constraints, energy constraints, and communication constraints to the multi-objective optimization model. The solution unit is used to solve the multi-objective optimization model using the non-dominated sorting genetic algorithm NSGA-II to obtain the Pareto optimal solution set; The robust processing unit is used to verify, by combining the predicted confidence interval with the Pareto optimal solution set and using the scenario sampling method, whether the scheme satisfies the safety constraints in all possible scenarios, and select the comprehensive optimal solution as the final redistribution scheme.

6. A method for autonomous cooperative control of unmanned aerial vehicle (UAV) swarms, characterized in that, Includes the following steps: S1. Real-time collection of status data and environmental data of each UAV, including the first... A drone Position vector at time , No. A drone velocity vector at time t , No. A drone Remaining energy at any given moment , No. A drone The current amount of work completed at any given moment , No. The drone and the first Between drones Communication link quality at any moment Environmental perception data; S2. Perform filtering, noise reduction, spatiotemporal registration, and normalization on the collected data; S3. Input the preprocessed data into the hybrid prediction model to predict the spatiotemporal evolution of the cluster over the next K time steps, where K is the number of prediction steps. S4. Based on the prediction results of the spatiotemporal evolution, calculate the spatial conflict risk, mission progress risk, and energy depletion risk at each future moment, and calculate the comprehensive risk index. S5. If there exists a step k such that the comprehensive risk index exceeds the preset risk threshold, then proactive reallocation is triggered, and the process proceeds to S6; otherwise, the process proceeds to S8. ; S6. Based on the multi-objective optimization model and the NSGA-II algorithm, a redistribution scheme is generated, and the final scheme is determined after robust constraint verification. S7. Distribute the redistribution scheme to each UAV and update the task queue and flight path of each UAV. S8. Each UAV executes control commands according to its current task assignment. S9. Collect actual execution effect data and calculate prediction error and redistribution effect indicators; S10. Update the parameters of the hybrid prediction model and the multi-objective optimization model based on the actual execution effect data; S11, return to S1, and enter the next control cycle.

7. The method for autonomous cooperative control of unmanned aerial vehicle (UAV) swarms according to claim 6, characterized in that, In step S3: The hybrid prediction model includes a physical prediction model and a data-driven prediction model. The physical prediction model makes predictions based on the kinematic equations of the UAV, and the data-driven prediction model makes predictions based on GAT+LSTM. No. A drone in Final prediction result of position at time Fusion as follows: ， in The output of the physical prediction model A drone in Predicted location at time The first output of the data-driven prediction model A drone in Predicted location at time To predict the step size, ;in and For the weighting coefficients, satisfying ,and and It adaptively adjusts with the number of steps k.

8. The method for autonomous cooperative control of unmanned aerial vehicle (UAV) swarms according to claim 7, characterized in that, In step S4: Space Conflict Risk Index Calculate using the following formula: ， Among them, the predicted first The drone and the first A drone in Interval of time , Parameters that are sensitive to conflict; Task progress risk index Calculate using the following formula: ， Among them, the predicted first The task is in Progress of completion at any given moment , To execute the first A collection of drones for a single mission. For the predicted first A drone in The amount of work completed at any given time. For the task Total workload For the task exist The target progress at any given time; Energy depletion risk index Calculate using the following formula: ， Among them, the predicted first A drone in Remaining energy at any given moment , For the predicted first The drone in Energy consumption per time step For the first Energy safety threshold for drones; Comprehensive risk indicators Calculate using the following formula: ， in For the weighting coefficients, satisfying .

9. The method for autonomous cooperative control of unmanned aerial vehicle (UAV) swarms according to claim 6, characterized in that, The multi-objective optimization model mentioned in step S6 is: , For task completion time, As an energy consumption balance indicator, For communication load, For redistribution disturbances; in This is the redistribution scheme matrix. Indicates the first Are drones assigned to missions? , For the first The drone completed its first The predicted time required for each task. For the first drones in the redistribution scheme Predicted energy consumption This represents the average energy consumption. In the redistribution scheme Downlink Communication load, This represents the current allocation status, where N is the total number of drones. This represents the total number of tasks.

10. The method for autonomous cooperative control of an unmanned aerial vehicle (UAV) swarm according to claim 6, characterized in that, The robust constraint verification in step S6 includes: Considering the uncertainty of prediction, the redistribution scheme must ensure that the safety constraints are met in all possible scenarios within the prediction confidence interval. The safety constraints include at least: the distance between any two drones is greater than or equal to the preset minimum safety distance; and the remaining energy of each drone is greater than or equal to the energy that needs to be reserved. By sampling from the confidence interval The solution is required to satisfy the security constraints in all sampled scenarios.

Images