Methods and systems for hierarchical collaborative perception and control of unmanned aerial vehicle (UAV) formations

Abstract

This invention belongs to the field of intelligent cooperative control technology for unmanned aerial vehicles (UAVs), and particularly relates to a method and system for hierarchical cooperative perception task control of UAV formations. The method includes: acquiring current environmental state data of the UAV formation, operational data of each UAV, and current constraint pressure data; generating global configuration commands based on the acquired data; acquiring a set of local commands for each UAV under the joint constraints of the global configuration commands, its own local observation data, and constraint pressure data; acquiring the deviation between the feedback information generated after executing the set of local commands and preset long-term operational constraint indicators, and updating the constraint pressure data; using the updated data as input for the next decision cycle, and returning to the step of generating global configuration commands. By incorporating constraint pressure data into the closed-loop hierarchical decision-making of global and local commands, the stability and adaptability of UAV formation cooperative perception task control in dynamic environments are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent collaborative control technology for unmanned aerial vehicles (UAVs), and in particular to a method and system for hierarchical collaborative perception and control of UAV formations. Background Technology

[0002] In existing technologies, intelligent perception and collaborative decision-making for UAV swarms typically employ techniques such as multi-UAV collaborative task models, reinforcement learning-based decision optimization, multi-agent task planning, and operations research optimization under intelligent game theory to achieve collaborative perception and task execution in complex environments. While these methods can improve the collaborative efficiency and environmental adaptability of UAV swarms to some extent, they still have the following shortcomings in practical applications: Most existing intelligent perception methods for UAV swarms focus on multi-agent parallel decision-making or general collaborative optimization, neglecting the hierarchical relationship between swarm-level global perception decision-making and individual-level local perception responses. This results in a lack of a unified collaborative modeling mechanism between upper-level task configuration and lower-level execution strategies. Furthermore, existing methods often emphasize instantaneous perception performance or local collaborative effects, making it difficult to simultaneously incorporate long-term operational constraints such as energy consumption, collision risk, communication load, and swarm deviation into the decision-making process. This can easily lead to short-term optimization effectiveness but insufficient long-term operational stability. Furthermore, in complex dynamic environments, affected by target maneuvers, external disturbances, and local observation noise, it is difficult to simultaneously balance policy convergence efficiency, system stability, and environmental adaptability by using fixed step size, fixed iteration depth, or static optimization methods, thereby affecting the overall completion effect of UAV formation intelligent perception tasks. Summary of the Invention

[0003] To overcome the deficiencies in the prior art, the first objective of this application is to provide a method for hierarchical collaborative perception task control of UAV formations; the second objective of this application is to provide a system that utilizes the method for hierarchical collaborative perception task control of UAV formations.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] Firstly, a method for controlling a UAV formation-level collaborative perception task includes:

[0006] Acquire current environmental status data of the drone formation, operational data of each drone, and current constraint pressure data;

[0007] Based on the environmental status data, the operational data of each UAV, and the constraint pressure data, a global configuration command is generated.

[0008] Obtain the set of local instructions for each UAV under the common constraints of the global configuration instructions, its own local observation data, and constraint pressure data;

[0009] Obtain feedback information generated after executing the set of local instructions, and update the constraint pressure data based on the deviation between the feedback information and at least one preset long-term operating constraint index.

[0010] The updated constraint pressure data is used as input for the next decision cycle, and the process returns to the step of generating global configuration instructions.

[0011] Optionally, the method further includes: obtaining task completion data from the previous decision cycle before executing the step of generating global configuration instructions; the generation of global configuration instructions is also based on the task completion data;

[0012] Based on the feedback information generated by executing the set of local instructions, the task completion data is updated, and the updated task completion data is used as the input for the next decision cycle.

[0013] Optionally, the generation of global configuration instructions includes:

[0014] Take a candidate instruction from the preset solution space as the current candidate instruction, and calculate the first optimization target value of the current candidate instruction based on the environmental state data, the operation data of each UAV, the local instruction set of the previous decision cycle, the task completion data and the constraint pressure data.

[0015] With the goal of increasing the first optimization target value, the current candidate instruction is iteratively adjusted in the solution space until a preset first termination condition is met. The candidate instruction that reaches the first termination condition is then used as the generated global configuration instruction.

[0016] Optionally, calculating the first optimization target value of the current candidate instruction includes:

[0017] Based on the environmental status data, the operational data of each UAV, the local instruction set of the previous decision cycle, and the task completion data, calculate the global task benefit item;

[0018] Calculate the constraint penalty term based on the constraint pressure data;

[0019] The first optimization target value is obtained by taking the weighted difference between the global task benefit term and the constraint penalty term.

[0020] Optionally, the iterative adjustment of the current candidate instruction within the solution space, with the direction of increasing the first optimization objective value, includes:

[0021] In this iteration, for each configuration parameter in the current candidate instructions, the partial derivative of the first optimization target value with respect to each configuration parameter is calculated to form a gradient vector; the configuration parameters include at least one of the following: perception resource allocation parameters, task weight setting parameters, formation reference configuration parameters, and perception region division parameters.

[0022] Based on the gradient vector and the preset global step size parameter, calculate the adjustment amount of each configuration parameter;

[0023] The current candidate instruction is updated according to each adjustment amount, and the feasibility of the updated candidate instruction is verified. If it exceeds the solution space, it is projected into the solution space and used as the current candidate instruction after this iteration adjustment; otherwise, the updated candidate instruction is directly used as the current candidate instruction after this iteration adjustment; the candidate instruction obtained after this iteration adjustment is used as the current candidate instruction for the next iteration.

[0024] Optionally, the steps for determining local configuration instructions include:

[0025] Each drone determines its own action constraint range based on the received global configuration instructions;

[0026] Each execution drone acquires candidate local configuration instructions for this iteration, and calculates the second optimization target value corresponding to the candidate local configuration instructions based on its own local observation data and constraint pressure data;

[0027] With the goal of increasing the second optimization target value, the current candidate local configuration instruction is iteratively adjusted within the constraints of its own actions until the preset second termination condition is met;

[0028] The candidate local configuration instruction that is met when the second termination condition is met is used as the generated local configuration instruction.

[0029] Optionally, the step of updating the task completion data includes:

[0030] Extract at least one of the following from the feedback information: target coverage area index, target recognition accuracy index, and multi-machine information fusion consistency index;

[0031] The extracted metrics are directly assigned to the corresponding fields in the task completion data to update the task completion data.

[0032] Optionally, the step of updating the constraint pressure data includes:

[0033] Identify the deviation of the feedback information from various long-term operating constraints, the types of which include at least one of energy consumption constraints, collision risk constraints, communication load constraints, and formation deviation constraints;

[0034] For each type of long-term operating constraint, the deviation is compared with a preset constraint threshold.

[0035] If the deviation exceeds the constraint threshold, the corresponding pressure value in the constraint pressure data is accumulated; otherwise, the pressure value is reduced, and the reduced pressure value is ensured to be greater than or equal to zero.

[0036] The constraint pressure data after accumulation or reduction is used as the updated constraint pressure data.

[0037] Optionally, before performing the step of returning to generate the global configuration instruction, the method further includes:

[0038] Obtain a first metric characterizing the drastic degree of change in the global configuration instructions;

[0039] Obtain a second metric that characterizes the uncertainty in the process of each executing UAV determining local configuration commands;

[0040] Based at least on one of the first metric, the second metric, and the constraint pressure data, adjust the global step size parameter used when generating the global configuration command for the next decision cycle, and / or adjust the iteration number threshold and local update step size used by each executing UAV when determining the local configuration command in the next decision cycle.

[0041] Optionally, adjusting the threshold for the number of iterations used by each UAV when determining the local configuration command in the next decision cycle includes:

[0042] Calculate the square root of the sum of squares of all pressure values ​​in the constraint pressure data, and use the result as the total constraint pressure measurement value;

[0043] Based on the total constraint pressure metric, determine the required solution accuracy for the next decision cycle;

[0044] Based on the required solution accuracy, the threshold for the number of iterations when determining local configuration instructions in the next decision cycle is increased.

[0045] Optionally, adjusting the local update step size used by each executing UAV when determining the local configuration command in the next decision cycle includes:

[0046] Obtain the noise intensity value represented by the second metric;

[0047] Based on the noise intensity value, and following the mapping relationship that the larger the noise intensity, the smaller the step size, the local update step size used when determining the local configuration instruction in the next decision cycle is reduced.

[0048] Optionally, the global step size parameter used when generating the global configuration command for the next decision cycle can be adjusted, including:

[0049] Obtain the total constraint pressure metric determined by the constraint pressure data, and the decision drift value determined by the first metric;

[0050] Based on the total constraint pressure metric and the decision drift degree, and following the linkage adjustment mapping relationship where a larger total constraint pressure metric corresponds to a smaller step size and a larger decision drift degree corresponds to a smaller step size, the global step size parameter for the next decision cycle is calculated.

[0051] Optionally, the method further includes:

[0052] Based on the required solution accuracy, the first metric, and the second metric, the cycle length of the next decision cycle is calculated; wherein the cycle length increases with the increase of the solution accuracy represented by the required solution accuracy, increases with the increase of the first metric, and increases with the increase of the second metric.

[0053] Based on the calculated cycle length, the time span of the next decision cycle is controlled, so that the step of returning to execute the generation of global configuration instructions is triggered when the time condition corresponding to the cycle length is met.

[0054] Secondly, a hierarchical collaborative perception task control system for unmanned aerial vehicle (UAV) formations includes:

[0055] The status and constraint acquisition module is used to acquire the current environmental status data of the UAV formation, the operation data of each UAV, and the constraint pressure data. The constraint pressure data is updated cumulatively based on the deviation of at least one type of long-term operation constraint index.

[0056] The global instruction generation module is used to generate global configuration instructions based on the environmental state data, the operation data of each UAV, and the constraint pressure data.

[0057] The local instruction acquisition module is used to issue the global configuration instruction to each execution UAV and acquire the local instruction set of the local configuration instruction determined by each UAV under the common constraints of the global configuration instruction, its own local observation data and constraint pressure data.

[0058] The constraint update module is used to obtain feedback information generated after executing the local instruction set, and update the constraint pressure data according to the deviation between the feedback information and the long-term operating constraint index.

[0059] The iterative control module is used to take the updated constraint pressure data as input for the next decision cycle and trigger the global instruction generation module.

[0060] Optionally, an adaptive adjustment module is also included for:

[0061] Calculate the total constraint pressure metric based on the constraint pressure data;

[0062] Obtain a first metric that characterizes the degree of drastic change in the global configuration instructions;

[0063] Obtain a second metric that characterizes the uncertainty in the process of determining local configuration commands for each UAV;

[0064] Based at least one of the first metric, the second metric, and the total constraint pressure metric, adjust the global step size parameter used when generating the global configuration command for the next decision cycle, and / or adjust the local update step size and iteration number threshold used by each UAV when determining the local configuration command for the next decision cycle.

[0065] Thirdly, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0066] Fourthly, a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the above-described method.

[0067] Due to the application of the above technical solution, the present invention has the following advantages compared with the prior art:

[0068] 1. This application generates global configuration instructions by a decision-making unit. Each executing UAV, guided by these global configuration instructions and combining its own local observation data and constraint pressure data, iteratively determines its own local configuration instructions, ultimately forming a set of local instructions. This hierarchical correspondence allows the formation-level mission intent and individual-level execution actions to operate collaboratively within a unified control framework. Compared to processing formation decisions and individual responses separately, this approach more closely reflects the actual coordination between upper-level configuration and lower-level execution in formation collaborative tasks.

[0069] 2. By acquiring and updating constraint pressure data, the deviations of long-term operational constraint indicators such as energy consumption constraints, collision risk constraints, communication load constraints, and formation deviation constraints from preset constraint thresholds are converted into quantifiable cumulative pressure indicators. During the generation of global configuration instructions and the iterative determination of local configuration instructions, the constraint pressure data serves as input in the calculation, ensuring that compliance with long-term constraints is considered in each decision-making cycle. When the degree of exceeding a certain type of long-term operational constraint indicator intensifies, the corresponding constraint pressure component increases, thereby exerting a stronger inhibitory effect on subsequent decisions and guiding the decision-making unit and each executing UAV to select configuration instructions with less disturbance to that type of constraint. This closed-loop approach, through the continuous accumulation and release of constraint pressure data, ensures that the cumulative deviation of various long-term operational constraint indicators during continuous operation can be stably controlled within preset thresholds without requiring additional constraint correction steps outside of the decision-making cycle.

[0070] 3. When generating global configuration instructions for the current cycle, the decision-making unit considers not only environmental status data and constraint pressure data, but also task completion data. This allows the allocation of sensing resources and the setting of task weights in this round to be adjusted based on the achieved task results. Using task completion data to guide current decisions enables resource allocation to be dynamically adjusted as the task progresses. Areas where task results have not met expectations can receive more resource investment in subsequent cycles.

[0071] 4. By acquiring a first metric representing the drastic change in global configuration instructions and a second metric representing the uncertainty in the determination of local configuration instructions, and combining the constraint pressure data to calculate the total constraint pressure metric, at least one of the global step size parameter, iteration number threshold, and local update step size for the next decision cycle is adjusted. When constraint pressure increases, instructions change drastically, or local uncertainty increases, the step size is reduced or the number of iterations is increased accordingly, making the solution process automatically more robust when environmental disturbances are large; when the system is in a relatively stable operating state, the solution pace can be accelerated. Through this adaptive adjustment method, problems such as response lag, local oscillation, or constraint accumulation that may occur under fixed step size or fixed number of iterations are reduced, thereby achieving dynamic matching between the solution pace of global configuration instruction generation and local configuration instruction determination and the system operating state, significantly improving the overall stability, coordination, and robustness of formation coordination tasks in complex dynamic environments.

[0072] To make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0074] Figure 1 This is a flowchart of the control method in an embodiment of this application;

[0075] Figure 2 This is a schematic diagram of the system in an embodiment of this application. Detailed Implementation

[0076] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Example 1: See Figure 1 As shown, this embodiment provides a method for controlling a UAV formation-level collaborative perception task, including:

[0078] Step S101: Obtain the current environmental status data of the drone formation, the operation data of each drone, the mission completion data, and the constraint pressure data.

[0079] The constraint pressure data is updated cumulatively based on the deviation of at least one type of preset long-term operating constraint index.

[0080] It should be noted that the environmental state data refers to the collection of various elements in the external environment and mission situation of the formation that can be measured by sensors or intelligence systems and converted into data form. The external environment refers to the natural conditions and electromagnetic environment in the space where the formation is located, such as meteorological conditions, terrain, and electromagnetic interference background. The mission situation refers to battlefield or operational scenario information directly related to the current mission execution, such as the activity status of enemy targets, the positional relationship between the friendly formation and the mission area, and the population or building distribution within the mission area. The environmental state data includes at least target distribution information, threat situation information, and communication link quality information. Specifically, the target distribution information describes the spatiotemporal state of objects to be identified or tracked in the reconnaissance area, and may include the location of each object, its speed range, and the confidence level in its presence. In an optional implementation, the target distribution information is represented by a target distribution heatmap obtained by airborne sensors and intelligence networks. The threat situation information is a graded expression of the potential danger level in different airspaces, and may include the threat level identifier corresponding to each airspace unit. In an optional implementation, the threat situation information is represented by a threat situation assessment matrix. The communication link quality information is a measure of the data link transmission capability between the various UAVs, and may specifically include the available bandwidth, current bit error rate, and signal propagation delay of each link. In an optional implementation, the communication link quality information is represented by a communication link quality report.

[0081] The operational data of each UAV is a collection of its real-time status information and the actions performed in the previous decision cycle. Each UAV's operational data includes at least the formation geometry, remaining energy status information, the current operating mode of each sensor, and the local configuration command record from the previous decision cycle. The formation geometry reflects the position and attitude of each UAV in space, specifically including three-dimensional coordinates, velocity vector, heading angle, pitch angle, and roll angle. This data can be output by the integrated navigation system onboard each UAV. The remaining energy status information reflects the current energy reserves available for consumption by each UAV, specifically including the remaining battery percentage or remaining fuel percentage, read by the energy management system. The operating modes of each sensor indicate whether the mission payload is currently in active detection, passive listening, or standby mode. The mission payload includes, for example, radar, electronic reconnaissance equipment, and electro-optical pods. The local configuration command record from the previous decision cycle refers to the local configuration command that the UAV finally determined and executed after iterative updates in the previous control cycle. Its specific form can be at least one of the following: local trajectory correction, beam pointing angle and dwell time, and sensor operating mode switching command. By incorporating the local configuration command into the operational data of each UAV, the process of generating the global configuration command can obtain the actual response status of each UAV at the previous moment. This information, along with the current environmental status data, task completion data, and constraint pressure data, forms the basis for decision-making, ensuring the integrity of information transmission during the closed-loop iteration process.

[0082] In an optional implementation, during the first decision cycle after the system's initial startup, since no historical cycles have been generated, the local configuration instruction record of the previous decision cycle is set to a preset initial value. This preset initial value can be a zero action, i.e., the default value or zero value of each adjustable action parameter; or it can be a pre-set default sensing configuration, such as setting the beam pointing angle to the pre-planned center direction of the search sector and setting the sensor operating mode to passive listening mode.

[0083] The task completion data is a quantitative evaluation of the formation's perception task completion status in the previous decision-making cycle, reflecting the actual task execution effect achieved by the formation in the previous cycle. In an optional implementation, the task completion data includes at least one of the following: the increase in target coverage area, the improvement in target recognition accuracy, and the gain in the situational consistency index after multi-aircraft information fusion. The target coverage area refers to the union area of ​​the effective detection range of all sensors in the formation within the task area; the target recognition accuracy refers to the proportion of correctly identified targets to the total number of detected targets; the situational consistency index after multi-aircraft information fusion measures the degree of agreement between the situational estimates of each UAV for the same target or the same area, and can be measured by the covariance or mutual information between the estimates. Using the task completion data of the previous cycle as input for the current decision-making cycle allows the leader to know the actual task performance of the formation in the previous stage when generating global configuration instructions, thereby making more targeted adjustments to the allocation of task resources in combination with the current environmental state and constraint pressure.

[0084] In an alternative implementation, during the first decision cycle after the system is first started, since no sensing action has been performed yet, the task completion data is set to a preset initial value, such as zero or a pre-set baseline performance value.

[0085] In this embodiment, the decision cycle refers to the time span of a complete control loop of this method, from acquiring state data, generating decisions, executing actions, to feedback updates and returning to prepare for the next round of decisions.

[0086] The constraint pressure data is a set of values ​​dynamically maintained by the system, used to characterize the cumulative exceedance degree of at least one type of long-term operating constraint. Each pressure value in the constraint pressure data is the current cumulative pressure indicator of the corresponding long-term operating constraint. For each type of long-term operating constraint, the system maintains a corresponding pressure value, which increases as the constraint exceeds its limit and decreases as the constraint is satisfied. In this way, abstract long-term stability requirements are transformed into numerical signals that can be acquired in real time and directly used in calculations.

[0087] In one optional implementation, for energy consumption constraints, an energy virtual queue is maintained. The current value of this energy virtual queue represents the cumulative amount by which the total energy consumption of the formation has exceeded a preset budget to date. This value serves as the current cumulative pressure indicator for the energy consumption constraint; a larger value indicates a higher energy pressure on the system. Similarly, a collision risk virtual queue can also be maintained, with its value serving as the current cumulative pressure indicator for the collision risk constraint. Its magnitude represents the cumulative degree to which the inter-machine distance falls below a safe threshold. Furthermore, a communication load virtual queue can be maintained, with its value serving as the current cumulative pressure indicator for the communication load constraint. Its magnitude represents the cumulative degree to which the communication load exceeds the link capacity.

[0088] In an optional implementation, the formation cooperative sensing process is described as a stochastic Stackelberg master-slave game model with long-term constraints, expressed as:

[0089] ;

[0090] in: Indicates time The system status includes information such as target distribution, threat situation, communication link quality, formation geometry, remaining energy, and environmental disturbances; This indicates the leader's actions, i.e., global configuration instructions, which can be specifically manifested as the allocation of sensing resources, the setting of task weights, the formation reference configuration, the division of sensing areas, or the scheduling parameters of the main sensor. Indicates the first The local following actions of the UAV, namely local configuration commands, can be specifically manifested as local trajectory correction, beam pointing, sensor operating mode, transmission power, view selection, or local sensing gain adjustment. A combination of local configuration instructions for all followers; Indicates a random state transition. These include random factors such as external disturbances, measurement noise, and target maneuvering. For the leader's reward function; For the first A follower's payoff function; This is a discount factor used to balance the relative importance of current returns and future returns.

[0091] Step S102: Based on the environmental status data, the operation data of each UAV, the constraint pressure data, and the task completion data, generate a global configuration command.

[0092] After acquiring the data described in step S101, the decision-making unit in the formation generates a global configuration command based on the environmental state data, the operational data of each UAV, the constraint pressure data, and the mission completion data. The decision-making unit in this step can be a mission planning computer located at a ground control station or command vehicle, or it can be a core UAV dynamically elected within the formation according to preset rules, which acts as the leader. The preset rules can be set based on conditions such as having the most fuel, the strongest computing power, etc.

[0093] In an optional implementation, this method is applied to a formation reconnaissance scenario in which a core UAV acts as the leader and multiple reconnaissance UAVs act as followers. In this scenario, the leader is responsible for generating search area division and sensor working mode configuration instructions based on mission intelligence. Each follower then autonomously adjusts its track and beam pointing within the designated area to improve the probability of target detection. At the same time, the system continuously monitors the overall energy consumption and safe distance of the formation and dynamically corrects subsequent decisions through feedback.

[0094] The global configuration command is a set of macroscopic decision variables used to guide the coordinated action of the entire formation. In one optional implementation, the global configuration command is represented as a structured set of configuration parameters, including: priority reconnaissance sectors for each UAV, search and identification task weights for each sector, and the desired formation configuration at a specific mission phase, such as a line formation or a wedge formation. In another optional implementation, the global configuration command is represented as a continuous control parameter vector, which consists of several continuously valued control parameters, such as the desired cruise speed, desired flight altitude, desired lateral spacing of each UAV in the formation, and the desired beam azimuth and transmit power of each airborne sensor.

[0095] When generating the global configuration command, the leader's decision-making basis includes environmental state data reflecting the mission situation, operational data of each UAV, mission completion data reflecting historical mission performance, and constraint pressure data reflecting the long-term stability pressure of the system. The mission completion data provides the leader with information on the formation's actual mission performance in the previous phase, enabling the leader to adjust the current mission resource allocation based on the achieved mission results. For example, if the mission completion data from the previous period indicates that the target recognition rate in certain areas remains low, the leader can allocate more perception resources or increase their mission weight in these areas when generating the global configuration command for this period. Thus, the generated global configuration command achieves a balance between mission gains and system stability.

[0096] In an optional implementation, the leader generates the global configuration instruction by solving a planning problem within a preset solution space, whereby the solution space defines the feasible range of values ​​for each configuration parameter in the global configuration instruction. Specifically, this includes the following sub-steps:

[0097] Step S1021: Select a candidate instruction from the preset solution space as the current candidate instruction. Based on the environmental state data, the operational data of each UAV, the local instruction set of the previous decision cycle, the task completion data, and the constraint pressure data, calculate the first optimization target value of the current candidate instruction. The local instruction set of the previous decision cycle is a set composed of the local configuration instructions that each UAV finally determined and executed in the previous decision cycle.

[0098] The solution space is a set of constraints defining the range of values ​​for each configuration parameter in the global configuration command. It delineates the feasible domain of all configuration commands that the formation can actually execute within the limits of physical capabilities, resource limits, and task rules. Any feasible solution to the global configuration command lies within this solution space. In this step, a candidate command is selected from the solution space as the starting point for iterative solution, i.e., the current candidate command. This initial candidate command can be a pre-set default configuration command or a feasible command generated through random sampling within the solution space. Using the current candidate command, the environmental state data, the operational data of each UAV, the local command set of the previous decision cycle, the task completion data, and the constraint pressure data as inputs, the first optimization objective value corresponding to the candidate command is calculated; the first optimization objective value is the difference between the weighted value of the global task benefit term and the weighted value of the constraint penalty term.

[0099] The global task benefit item reflects the overall task completion effect that the formation is expected to achieve under the current candidate command. The composition of this benefit item is related to the specific perception task type. In an optional implementation, if the core task of the formation is to cover, identify, and fuse perception of the target area, then the global task benefit item is obtained by weighted summation of at least one of the following: coverage quality benefit, target identification benefit, and multi-aircraft information fusion benefit. Specifically, coverage quality benefit measures the effective coverage area or coverage overlap rate of the formation's sensors over the task area; target identification benefit measures the accuracy of correctly classifying and identifying detected targets; and multi-aircraft information fusion benefit measures the degree to which the fusion processing of perception information from each UAV improves the consistency of overall situational awareness. The weights of each benefit can be adjusted according to the task stage or the commander's intentions; for example, coverage benefit is emphasized in the search phase, and identification benefit is emphasized in the confirmation phase. To reflect the corrective effect of historical execution on current decisions, the task completion data of the previous cycle is also introduced as a benchmark or correction factor when calculating each benefit, so that the benefit assessment not only depends on the estimated value of the current environment and candidate commands but is also calibrated by the actual task effect of the previous cycle.

[0100] The constraint penalty term transforms the current long-term operational stability pressure of the system into an explicit cost to candidate instructions. This penalty term is obtained by weighted summation of deviations from various long-term operational constraints using the constraint pressure data. The constraint pressure data is a dynamically maintained set of values, where each pressure value corresponds to the current cumulative pressure indication of a type of long-term operational constraint. The types of long-term operational constraints include at least one of energy consumption constraints, collision risk constraints, communication load constraints, and formation deviation constraints. For each type of constraint, its pressure value is updated based on the historical cumulative deviation of that constraint: if a certain type of constraint has frequently exceeded its limit historically, the corresponding pressure value gradually increases; if it is consistently met, the pressure value gradually decreases. The deviation refers to the instantaneous over-limit of that type of constraint expected to occur under the current candidate instruction, i.e., the portion of the expected actual consumption or risk level exceeding a preset benchmark value within the current decision-making cycle. Multiplying the pressure values ​​corresponding to each type of constraint by the deviation and summing them yields the constraint penalty term. When the pressure value of a certain type of long-term operating constraint is large, even if its current deviation is not large, the penalty corresponding to the constraint will be amplified due to the multiplicative effect, and a stronger inhibitory effect will be exerted on the candidate instructions that may further worsen the constraint in the optimization objective. For constraints with smaller pressure values, the penalty term is weakened accordingly, releasing more optimization space for improving task benefits.

[0101] The first optimization objective value is obtained by subtracting the weighted value of the global task benefit term from the weighted value of the constraint penalty term. This difference structure allows the first optimization objective value to comprehensively reflect the candidate instruction's overall score across both task performance and stability dimensions: if any candidate instruction only pursues task benefits while ignoring long-term constraints, the total score will be lowered due to the increased constraint penalty term; if it is overly conservative and sacrifices task benefits, the contribution of the global task benefit term will be insufficient. In this way, the first optimization objective value provides a unified quantitative basis for subsequent iterative searches.

[0102] Step S1022: With the goal of increasing the first optimization target value, iteratively adjust the current candidate instruction in the solution space until a preset first termination condition is met. The candidate instruction that reaches the first termination condition is then used as the generated global configuration instruction.

[0103] The solution space is a constraint on the values ​​of each configuration parameter in the global configuration instruction, used to ensure that the generated configuration instruction does not exceed the formation's physical capabilities, resource limits, or task preset boundaries. After each iteration adjustment, if the adjusted candidate instruction exceeds the solution space, it is corrected to the boundary of the solution space, so that subsequent iterations always proceed within the feasible domain.

[0104] The direction of increasing the first optimization target value refers to the adjustment direction that, starting from the current candidate instruction, increases the first optimization target value within the current neighborhood. In an optional implementation, this direction is determined by calculating the partial derivatives of the first optimization target value with respect to each configuration parameter in the current candidate instruction. The vector composed of the partial derivatives of each configuration parameter is the gradient information that guides the adjustment direction.

[0105] In one optional implementation, the preset first termination condition is set to the number of iterations reaching a preset upper limit. In another optional implementation, the preset first termination condition can also be triggered by judging the convergence state of the current candidate instruction. For example, a projected gradient map is defined to measure the deviation of the current global configuration instruction from the first-order stable point.

[0106] ;

[0107] in, For feasible region projection operators, This is the global step size parameter. Optimize the gradient of the queue-weighted task target with respect to the global configuration instructions. When When the value gradually decreases, it indicates that the global configuration instruction is approaching the stationary region in a feasible sense, meaning that the current instruction is close to a locally stable solution.

[0108] In an optional implementation, the iterative adjustment of the current candidate instructions within the solution space, with the direction of increasing the first optimization objective value, specifically includes the following steps:

[0109] Step S1022a: In this iteration, for each configuration parameter in the current candidate command, calculate the partial derivative of the first optimization target value with respect to each configuration parameter to form a gradient vector. The configuration parameters include at least one of the following: perception resource allocation parameters, task weight setting parameters, formation reference configuration parameters, and perception region division parameters. The perception resource allocation parameters specify the sensor working time or detection resource ratio allocated to each UAV; the task weight setting parameters specify the search priority or identification confidence requirement for each sector; the formation reference configuration parameters specify the geometric parameters of the desired formation, such as the desired position or spacing of each UAV; and the perception region division parameters specify the airspace boundary to be reconnoitered by each UAV. These configuration parameters together constitute the adjustable dimension of the current candidate command. When calculating the partial derivative, a small perturbation can be applied to each configuration parameter at its current value, and the change in the first optimization target value can be observed. The ratio of this change to the perturbation amount is used as an approximate value of the partial derivative corresponding to that configuration parameter. The gradient vector formed by combining the partial derivatives of all configuration parameters indicates the direction in which the first optimization objective value increases the fastest when adjusting each configuration parameter within the neighborhood of the current candidate instruction.

[0110] Step S1022b: Calculate the adjustment amount of each configuration parameter based on the gradient vector and the preset global step size parameter. In an optional implementation, the adjustment amount is obtained by multiplying the first-order partial derivative value of each configuration parameter in the gradient vector by the global step size parameter. The global step size parameter controls the magnitude of the adjustment in each iteration.

[0111] Step S1022c: Update each configuration parameter in the current candidate instruction according to the adjustment amounts to obtain a new candidate instruction. Specifically, the update method is as follows: perform an algebraic summation of the current value of each configuration parameter and its corresponding adjustment amount to obtain a new value for that configuration parameter. The updated values ​​of all configuration parameters together constitute the new candidate instruction.

[0112] Step S1022d involves performing a feasibility check on the updated candidate instructions to determine whether the new candidate instructions exceed the preset solution space. If the new candidate instructions exceed the solution space, they are projected into the solution space, and the projected configuration instructions are used as the current candidate instructions after this iteration. Otherwise, the updated candidate instructions are directly used as the current candidate instructions after this iteration. This projection operation ensures that the candidate instructions obtained after each round of iteration are always within the feasible region.

[0113] Through repeated execution of the above steps, the candidate instructions gradually evolve in the direction of increasing the first optimization objective value, gradually approaching a solution that maximizes the first optimization objective value within the feasible region.

[0114] The first termination condition is a criterion used to determine whether the iterative process has converged and no further adjustments are needed. In an optional implementation, the first termination condition includes at least one of the following: the improvement of the first optimization target value in the current iteration compared to the previous iteration is less than a preset convergence threshold; the magnitude of the iterative update direction is less than a preset gradient threshold, indicating that the current candidate instruction has approached a local optimum; the current iteration reaches a preset maximum iteration count limit to ensure the time controllability of the decision cycle. When any of the above termination conditions are met, the iterative process ends, and the current candidate instruction is output as the final global configuration instruction generated in this round.

[0115] By constructing and solving the planning problem described above, leaders, while pursuing high task rewards, will be adjusted by constraint pressure data, thereby achieving a balance between task effectiveness and long-term stability at the decision-making level.

[0116] In an optional implementation, the leader's decision-making process is further described as a two-level optimization problem with follower-optimal responses. Specifically, the leader-side objective is expressed as:

[0117] ;

[0118] in: ;

[0119] In the above formula, This represents the optimal leader action to be determined, i.e., the global configuration instructions to be generated in this round; This indicates the candidate global configuration directive. This represents the response actions of the followers, i.e., the set of local configuration instructions for each drone. This indicates that, given a global configuration directive and constraint pressure data Equilibrium response mapping of lower followers; This represents the constraint pressure data at the current moment; Represents the system state. Function This represents the overall mission quality gain of the formation, corresponding to the aforementioned global mission gain item; function Represents system-level cost; vector It is a constraint-over-limit vector; parameters This is a benefit-stability balance parameter used to balance task benefits and long-term constraint stability, adjusting the relative importance between perceived benefits and long-term constraint stability. Objective function The last item This refers to the constraint penalty term, which is derived from the constraint pressure data. Constraints exceeding limits Weighted composition; symbols This represents the expected operation in a random environment.

[0120] Under the above method, the constraint pressure data As an explicit pressure input, this causes the decision-making unit to consider the current system's stability pressure when generating instructions. For example, when the constraint pressure component corresponding to energy consumption is at a high level, a configuration instruction that improves mission coverage but requires high-speed maneuvering and high-power transmission from each aircraft, while increasing... The value is determined, but at the same time, a large deviation in energy consumption will lead to an increase in the constraint penalty, thus affecting the overall score. It will be subject to a significant penalty, thereby guiding the generation of a more energy-efficient and balanced global configuration instruction.

[0121] Step S103: Issue the global configuration command to each execution UAV, and obtain the local command set of local configuration commands determined by each UAV under the common constraints of the global configuration command, its own local observation data and constraint pressure data.

[0122] After the global configuration command is generated, it is distributed to each execution UAV within the formation. Each execution UAV, based on the received global configuration command, its own local observation data, and the constraint pressure data, iteratively updates its own local configuration command, ultimately resulting in a local command set composed of the local configuration commands of all UAVs. During this process, each UAV determines its own action constraint range according to the global configuration command and optimizes within that range.

[0123] Each UAV faces a local decision-making task: under the premise of complying with the macroscopic requirements of the global configuration command, and combining its own local information, and constrained by the stability pressure transmitted by the constraint pressure data, it determines a local configuration command. The local configuration command is a refinement and physical implementation of the global configuration command; that is, each executing UAV, based on the specified macroscopic task framework and its own local environment and current state, autonomously decides on specific flight actions and sensor operations. In an optional implementation, the local configuration command includes: within the assigned sector, the UAV fine-tunes its flight path to better track a moving target; or, dynamically adjusts its radar beam direction and dwell time in a certain area to optimize tracking accuracy; or, switches between passive listening and active detection modes.

[0124] The iterative update refers to the UAV gradually optimizing its local configuration commands within a local iterative loop, rather than calculating the final result all at once. The local configuration commands of each UAV are interconnected; the local configuration command of one UAV can alter the collaborative benefits and constraint penalties of other UAVs. Therefore, repeated adjustments and mutual adaptation are necessary to achieve an overall coordinated equilibrium. During the iteration process, the UAV evaluates the current candidate local configuration commands of its neighbors to achieve collaboration, while considering the relationship between its own actions and system stability pressures. Through multiple rounds of adjustments, it ultimately converges to a local configuration command that achieves a balance between local benefits, team collaboration, and stability constraints. The final local configuration commands determined by all UAVs constitute the local command set.

[0125] In one optional implementation, the process by which each UAV determines its local configuration instructions specifically includes the following sub-steps:

[0126] In step S1031, each executing UAV determines its own action constraint range according to the received global configuration instructions.

[0127] The self-motion constraint range refers to the limitation on the action values ​​of the UAV under the physical mobility and mission requirements. When performing local motion optimization, each UAV cannot adjust its motion parameters indefinitely; it must do so within the range of physical capabilities and the boundaries allowed by the mission. In an optional implementation, the self-motion constraint range includes at least one of the following: a maximum turning radius limit, determined by the UAV's maneuverability and flight envelope; upper and lower limits of flight speed, jointly determined by the UAV's power system performance and mission efficiency requirements; the azimuth and pitch angle range of the sensor beam, determined by the mechanical limits of the sensor servo mechanism; and an upper limit of transmission power, determined by the rated power and heat dissipation capacity of the airborne transmitter.

[0128] Step S1032: Each execution drone acquires the candidate local configuration instructions for this iteration and calculates the second optimization target value corresponding to the candidate local configuration instructions.

[0129] The second optimization objective value is composed of a weighted sum of local task benefit terms, collaborative benefit terms, and constraint penalty terms.

[0130] The local task benefit item reflects the task effect that the UAV can obtain with its current candidate local configuration command, and is related to its local target recognition rate, coverage gain or detection quality.

[0131] The collaborative benefit term reflects the complementarity or consistency between the current candidate local configuration commands of the UAV and those of other UAVs, and is used to guide the collaborative cooperation among the UAVs. For example, when two UAVs simultaneously cover the same target area from different angles, the quality of the fused information is higher than the simple sum of their individual detections; this additional gain is reflected in the collaborative benefit term. Conversely, if the sensor beams of two UAVs overlap and cover the same area on the same frequency band, mutual interference may occur, and the collaborative benefit term will be negative.

[0132] The constraint penalty term is obtained by weighted summing of the deviations of the UAV from various long-term operational constraints based on the constraint pressure data. It is used to transmit the current stability pressure of the system to the individual UAV's decision-making process. The construction method of this constraint penalty term is similar to that used when generating global configuration commands, the difference being that the UAV only evaluates the local contribution of its own actions to various constraints. When the current cumulative pressure indication of a certain type of long-term operational constraint is large, the corresponding constraint pressure component is high, and the proportion of this constraint penalty term in the second optimization objective value increases accordingly, thereby guiding the UAV to select local configuration commands that are more favorable to that type of constraint.

[0133] In an optional implementation, the second optimization objective further includes an entropy increase term. This entropy increase term is related to the information entropy of the UAV's current strategy and is used to quantify the uncertainty of its strategy selection. The reason for including the entropy increase term in the optimization objective is that if the UAV makes deterministic decisions solely with the goal of maximizing mission gains and minimizing penalties, the strategy may prematurely converge to a local optimum, losing its ability to explore the environment. The entropy increase term is a regularization term, encouraging the UAV to maintain a certain degree of strategy randomness. In an optional implementation, the information entropy is obtained by calculating the negative of the sum of the logarithms of the probabilities of each candidate action under the current strategy and the products of those probabilities.

[0134] In step S1033, each executing UAV, with the direction of increasing the second optimization target value, iteratively adjusts the current candidate local configuration command within its own action constraint range until the preset second termination condition is met.

[0135] In each round of iterative adjustment, each UAV determines a local iterative update direction that increases the second optimization target value, and adjusts the candidate local configuration command along this direction with a preset local update step size. The local iterative update direction can be determined by calculating the partial derivative of the second optimization target value with respect to each adjustable parameter in the current candidate local configuration command. The local update step size controls the magnitude of each round of adjustment. If the adjusted candidate local configuration command exceeds its own action constraint range, it is corrected to the boundary of the constraint range, and the correction method is similar to the projection operation when generating global configuration commands.

[0136] In an optional implementation, the first The first follower The inner layer update is represented as:

[0137] ;

[0138] in, This indicates that the local configuration command will be projected back to the first... Projection operator for the self-motion constraint range of a drone. For local update step size, For the first The noisy gradient estimation of the regularization benefit of a drone with respect to its local configuration commands. Indicates the first The combination of candidate local configuration instructions for all UAVs during round iteration. This represents the threshold number of iterations at the current moment.

[0139] During iterative updates, each executing drone also acquires candidate local configuration instructions from other executing drones in the current round to evaluate the synergistic benefits between its own actions and those of other drones. Over multiple iterations, the candidate local configuration instructions for each drone continuously change. When the action of one drone changes, the synergistic benefits for other drones also change, thus affecting their action updates. Through this continuous adjustment of interactions, the candidate local configuration instructions of all followers gradually approach an equilibrium state.

[0140] Step S1034: The candidate local configuration instruction when the second termination condition is met is used as the generated local configuration instruction.

[0141] The second termination condition is a criterion used to determine whether the local iteration process has converged and no further adjustments are needed. In an optional implementation, the second termination condition includes at least one of the following: the norm of the change in the action vector between the current iteration round and the previous round is less than a preset convergence threshold, indicating that the local configuration instructions of each executing UAV have basically stabilized and no longer change significantly; the current iteration round reaches a preset maximum local iteration count limit to ensure the time controllability of the local decision-making process and avoid consuming too much computation time in a single decision cycle. When any of the above second termination conditions are met, the iteration process ends, and the current candidate local configuration instruction is taken as the final local configuration instruction generated by the executing UAV in this decision cycle.

[0142] In one optional implementation, the mathematical description of each UAV local policy iteration is as follows. For a given global configuration instruction... , No. The local benefit of a drone is defined as:

[0143] ;

[0144] in, This represents the task gains related to local target recognition rate, coverage gain, and detection quality. This indicates the collaborative benefits related to neighboring machine collaboration, consistency awareness, or information complementarity. Indicates energy consumption cost; This indicates penalties related to collision risk, formation deviation, and link congestion. , , , The corresponding weights.

[0145] In an optional implementation, to prevent the drone strategy from becoming too aggressive or degenerating into a deterministic local strategy, an entropy regularization term is introduced, thereby... The regularized profit of a drone is defined as:

[0146] ;

[0147] in, To constrain pressure data; Indicates the first Deviation from long-term constraints; This indicates that the constraint applies to the first... The impact coefficient of revenue per drone; Entropy regularity strength; Current strategy for drones The information entropy is used to quantify the uncertainty of its strategy selection. It should be noted that the structure of this regularized reward function corresponds to the aforementioned second optimization objective value: where... Corresponding to local task benefits and collaborative benefits, Corresponding constraint penalty item, This corresponds to the entropy increase term.

[0148] In the given global configuration directive and constraint pressure data Subsequently, a parametric game is formed among all the executing drones, and its equilibrium response is denoted as... Each component satisfies:

[0149] ;

[0150] This formula represents how each executing drone, assuming the local configuration instructions of other executing drones remain unchanged, chooses the local configuration instruction that maximizes its own regularization gain. The system reaches equilibrium when no executing drone can further increase its regularization gain by unilaterally changing its local configuration instructions. The equilibrium response at this point is... This refers to the set of instructions for each executing UAV to achieve the optimal local configuration under the current global configuration instructions and constraint pressure data.

[0151] In an optional implementation, the equilibrium condition is characterized using variational inequalities. A pseudo-gradient mapping is defined. That is, the vector mapping formed by the negative gradients of the relevant benefit components in the second optimization objective value of each executing UAV with respect to its own local configuration command, then the equilibrium satisfy ,in This represents the Cartesian product of all actions performed by the drone. The intuitive meaning of this inequality is that at the equilibrium point, any change in local configuration commands that deviates from the equilibrium will lead to a decrease in regularization gains. have Strong monotonicity and - Lipschitz continuity ensures a unique equilibrium among the executing drones, and the equilibrium mapping exhibits Lipschitz continuity with respect to the global configuration commands. Lipschitz continuity means that small changes in the global configuration commands will not cause drastic jumps in the equilibrium response of the executing drones, which is crucial for ensuring the stability of the master-slave game solution process. Each executing drone approximates this equilibrium through the aforementioned multiple iterations.

[0152] Step S104: Obtain feedback information generated after executing the set of local instructions, and update the constraint pressure data according to the deviation between the feedback information and at least one preset long-term operating constraint index.

[0153] After the formation executes the aforementioned set of local commands, it generates a series of feedback information through sensors and communication links. This feedback information is an objective record of the formation's actual operational performance, including raw data detected by each sensor, actual communication records between each UAV, and actual movement trajectories and energy consumption data of each UAV.

[0154] Updating the constraint pressure data specifically includes: identifying the deviation of the feedback information from various long-term operational constraint indicators, wherein the types of long-term operational constraint indicators include at least one of energy consumption constraints, collision risk constraints, communication load constraints, and formation deviation constraints. The deviation refers to the portion of the actual consumption or actual risk level of this type of constraint exceeding a preset long-term allowable upper limit within the current period. In an optional implementation, the deviation of the energy consumption constraint can be obtained by subtracting the periodic average energy budget from the actual energy consumed in the current period; the deviation of the collision risk constraint can be obtained by subtracting the minimum distance between any two machines recorded in the current period from a preset safety distance; the deviation of the communication load constraint can be obtained by subtracting the link capacity upper limit from the actual data transmitted in the current period; and the deviation of the formation deviation constraint can be obtained by subtracting a preset allowable deviation upper limit from the average Euclidean distance between the actual position of each machine and the desired position in the reference configuration.

[0155] For each type of long-term operational constraint, the deviation is compared with a preset constraint threshold. If the deviation exceeds the constraint threshold, the corresponding pressure value in the constraint pressure data is accumulated; otherwise, the pressure value is reduced, ensuring that the reduced pressure value is greater than or equal to zero. The constraint pressure data after accumulation or reduction is used as the updated constraint pressure data.

[0156] In this embodiment, the method further includes updating task completion data. Specifically, at least one of the following is extracted from the feedback information: target coverage area index, target recognition accuracy index, and multi-machine information fusion consistency index. The extracted index is then directly assigned to the corresponding field in the task completion data to update the task completion data. The updated task completion data reflects the actual task performance achieved by the formation within the current decision cycle. This data will be acquired in step S101 of the next decision cycle and used as the input basis for generating global configuration instructions in step S102.

[0157] In an optional implementation, the update process for the constraint pressure data is based on the Lyapunov stability principle. The Lyapunov function is defined as:

[0158] ;

[0159] in, This represents the constraint pressure data at the current moment. For the first The current cumulative pressure indication of long-running constraints. This represents the total number of long-term constraints. The Lyapunov function is used to measure the overall energy of the system's current constraint stress state: the larger the function value, the more severe the current constraint exceedance of the system. A one-step conditional drift is further defined as:

[0160] ;

[0161] The one-step conditional drift represents the current state. and current constraint pressure data Under certain conditions, the expected change of the Lyapunov function after one decision cycle is given. If this drift is positive and continues to increase, it means that the system has a tendency to evolve towards instability.

[0162] To simultaneously optimize perception performance and ensure that long-term constraints do not diverge, a drift-penalty joint criterion is adopted:

[0163] ;

[0164] in, For the return-stability balance parameter, To improve the overall mission quality and benefits of the formation. This comes at a system-level cost. The criterion will constrain the stability objective (determined by drift). The perceived performance target (reflected by perceived net benefit) is combined with the perceived performance target in the same expression, through balancing parameters. Adjust the relative weights of the two.

[0165] By expanding the square of the virtual queue (i.e., the constraint pressure data), the upper bound of the drift-penalty joint criterion has the following structure:

[0166] ;

[0167] in The constant term is a result of the boundedness of the constraint increment, and its existence is guaranteed by the boundedness of the instantaneous overlimit. For the first Deviation from long-term operational constraints. For the first The preset constraint threshold for long-term operation constraints.

[0168] Therefore, minimizing this drift-penalty upper bound within each decision cycle is equivalent to maximizing the following queue-weighted task optimization objective:

[0169] ;

[0170] This objective consists of two parts: the first part represents the parameters... The amplified global task net benefit is the difference between the global task benefit and the system-level cost multiplied by the balancing parameter; the second term represents the constraint penalty weighted by the constraint pressure data, i.e., the value of each virtual queue. With the corresponding instantaneous over-limit Multiply and sum. The purpose of this objective is that the system does not mechanically switch between pure performance maximization and pure constraint minimization, but rather dynamically adjusts the benefit objective at each step through queue pressure.

[0171] In one optional implementation, the specific formula for updating the constraint pressure data is defined as follows:

[0172] ;

[0173] in, For the first The constraint pressure component of the class constraint at the next time step. The constraint pressure component at the current moment. This represents the deviation detected during this period. For the first The preset constraint threshold for a certain type of constraint. The meaning of this formula is consistent with the aforementioned update logic: if the deviation of a certain type of constraint within the current period... Exceeding the preset constraint threshold If the pressure exceeds the limit, the corresponding pressure value increases by the difference between the two values; if it does not exceed the limit, the pressure value is gradually reduced, ensuring that the updated pressure value is never less than zero. The updated constraint pressure data reflects the cumulative pressure indication of various long-term operating constraints up to the current moment.

[0174] In an optional implementation, based on the queue-weighted task optimization objective, the leader can update the global configuration instructions according to the following update formula when generating subsequent global configuration instructions:

[0175] ;

[0176] in, A projection operator that projects the updated global configuration instructions back into the solution space. This is the global step size parameter. A noisy gradient estimate of the queue-weighted task optimization objective with respect to the global configuration instruction is provided. This update causes the global configuration instruction to evolve progressively along the direction of improving the queue-weighted task optimization objective, and this update process occurs in step S102 of the next decision cycle.

[0177] The updated constraint pressure data is used as input for the next decision cycle, and the process returns to step S102 to generate the global configuration instruction for the next time step, thereby forming a closed-loop control from task decision to stability arbitration.

[0178] Step S105: Use the updated constraint pressure data as input for the next decision cycle, and return to the step of generating global configuration instructions.

[0179] This step forms a closed loop of task planning, decision-making, execution, feedback, and correction. Updated constraint pressure data is carried over to the next control cycle as a form of pressure memory, and updated task completion data is also carried over as a form of performance memory. This pressure memory enables the system to learn from historical operations: if a constraint exceeded its limits in the previous cycle, its cumulative pressure indicator increases, thus more strongly suppressing decisions that would further worsen the constraint in the next cycle; if, after adjustments over multiple cycles, the cumulative pressure of a constraint gradually decreases, the system's restrictions on that constraint will be relaxed accordingly, releasing more optimization space for improving task performance. This performance memory allows the decision-making unit to guide subsequent resource allocation based on the actual task results already achieved. As the system continues to operate, this closed-loop approach ensures that the cumulative amount of all long-term constraints remains stable within a preset threshold in a time-averaged sense, while maintaining the overall task performance of the formation.

[0180] In an optional implementation, the global task benefit item in S102 is calculated as follows. The formation-level global task performance function is defined as:

[0181] ;

[0182] in, This indicates coverage of quality benefits. Indicates the benefit of target identification. Indicates the benefits of multi-machine information fusion. , , The corresponding weights are used to adjust the relative importance among the three types of perception sub-targets. This formula indicates that the formation-level benefit is not a single indicator, but a structured integration of the coverage, identification, and fusion of three core perception targets, with the contribution of each sub-target to the overall benefit controlled by its corresponding weight.

[0183] The system-level cost function is defined as:

[0184] ;

[0185] in, Indicates the cost of energy consumption. Indicates the cost of collision or threat risk. Indicates the cost of formation deviation. Indicates the cost of communication load. , , , The corresponding weights are used. The system-level cost function unifies the operating costs of different dimensions under the same scale, making it easier to make comprehensive trade-offs in the optimization objective.

[0186] In an optional implementation, before the step of returning to execute the global configuration instruction in step S105, i.e., before the end of the current decision cycle and the start of the next decision cycle, the method further includes an adaptive adjustment step. The system adaptively adjusts the solution parameters for the next cycle based on the operating state within the current cycle, dynamically adjusting the solution rhythm of the closed-loop control to enhance the robustness and real-time performance of the system in complex dynamic environments. Specifically, this includes:

[0187] Step S1061: Obtain a first metric characterizing the degree of drastic change in the global configuration instructions.

[0188] The first metric is used to quantify the magnitude of the leader's adjustments to global configuration instructions between two consecutive decision cycles. In an optional implementation, the first metric is obtained by calculating the norm of the difference between the global configuration instruction vectors generated by the leader in the two consecutive decision cycles, denoted as . The calculation formula is as follows:

[0189] ;

[0190] in, Mapping the balanced response of drones The Lipschitz constant is used to map the changes in global configuration commands to an estimated upper bound of the changes in the UAV's equilibrium response; The global configuration directives generated in this cycle; This refers to the global configuration command from the previous cycle; This represents the Euclidean norm of a vector, which is the square root of the sum of the squares of the vector's components. The larger the value of this first metric, the more drastic the changes in mission configuration, and the more time and finer the iterations required for the drone to adapt to the new higher-level decisions.

[0191] Step S1062: Obtain a second metric that characterizes the uncertainty in the process of each executing UAV determining local configuration instructions.

[0192] The second metric is used to quantify the noise level or information uncertainty faced by the UAV during local policy iteration. In an optional implementation, the second metric is obtained by statistically analyzing the variance or noise intensity of the policy gradient estimation for each UAV during the iteration process, denoted as... The larger the value of this second metric, the noisier the local decision-making environment and the less reliable the gradient estimation. In this case, if an excessively large local update step size is used, it can easily lead to policy oscillations and non-convergence.

[0193] Step S1063: Based at least one of the first metric, the second metric, and the constraint pressure data, adjust the global step size parameter used when generating the global configuration instruction for the next decision cycle, and / or adjust the iteration number threshold and local update step size used by each executing UAV when determining the local configuration instruction in the next decision cycle.

[0194] The total pressure state represented by the constraint pressure data can be characterized by a total constraint pressure metric. In an optional implementation, the total constraint pressure metric is defined as... Its calculation formula is This is the square root of the sum of the squares of all constraint pressure components. This metric comprehensively reflects the total current stability pressure of the system. The larger the value, the more severe the cumulative exceedance of various long-term constraints, and the higher the stability pressure faced by the system.

[0195] The total pressure state represented by the constraint pressure data is defined by a Lyapunov function, i.e. The value of this function comprehensively reflects the total current stability pressure of the system. The larger the value, the more serious the cumulative exceedance of various long-term constraints, and the higher the stability pressure faced by the system.

[0196] In one optional implementation, the required solution accuracy for the next decision cycle is determined based on the total constraint pressure metric. Calculate according to the following formula:

[0197] ;

[0198] in, These are the preset basic accuracy parameters. The preset queue sensitivity coefficient, Let be the total constraint pressure metric. This formula indicates that when When the value increases, i.e. when the system is under high constraint pressure, the required solution accuracy value increases. As the numerical value decreases, the required precision increases accordingly.

[0199] Based on the required solution accuracy, the threshold for the number of iterations when determining local configuration instructions in the next decision cycle is increased. Calculate according to the following formula:

[0200] ;

[0201] in, The preset iteration depth coefficient, Let be the required solution accuracy value, and be the symbol. This indicates rounding up. This formula shows that the higher the required precision, the more precise the desired result. The smaller the threshold, the fewer iterations are required. The more iterations, the more likely it is to happen. The rounding up operation ensures that the number of iterations is always an integer.

[0202] In an optional implementation, adjusting the local update step size used by each UAV when determining local configuration instructions in the next decision cycle specifically includes: calculating the local update step size according to the second metric using the following formula. :

[0203] ;

[0204] in, The preset base step size parameter, The preset noise sensitivity coefficient, This is the second metric. This formula shows that the larger the local noise, i.e., the larger the value of the second metric, the larger the calculated local update step size. The smaller the size, the more robust the noise-resistant update.

[0205] In an optional implementation, adjusting the global step size parameter used when generating the global configuration instruction for the next decision cycle specifically includes: calculating the global step size parameter according to the following formula based on the total constraint pressure metric and the first metric. :

[0206] ;

[0207] in, The preset base step size parameter, and The preset weighting coefficients, This is the total constraint pressure measurement value. The first metric is defined by this formula. This formula indicates that when the system stability pressure is high or the global configuration instructions change drastically, the global step size parameter will automatically decrease, making the decision-making process smoother and more conservative, and avoiding large-scale strategy adjustments when the state is unstable or the information changes drastically.

[0208] In an optional implementation, the method further includes: calculating the cycle length of the next decision cycle based on the required solution accuracy, the first metric, and the second metric. Calculate according to the following formula:

[0209] ;

[0210] in, and The preset period length coefficient, This is the required solution accuracy value. For the second metric, This is the primary metric. The formula indicates that the higher the required solution accuracy, the stronger the local noise, and the more drastic the changes in global configuration instructions, the longer the required decision cycle length. This allows for more computation time for optimization at both upper and lower levels, ensuring decision quality. The adaptive adjustment of the cycle length, combined with the adjustment of the step size and the number of iterations, constitutes a coordinated control of the solution rhythm.

[0211] Furthermore, the tracking error of each UAV for instantaneous equalization satisfies the following boundary:

[0212] ;

[0213] in, , , and These are constants related to the system characteristics, depending on the geometric properties of the UAV's local optimization problem and the monotonicity of the pseudo-gradient mapping, the Lipschitz constant, etc. The first term... It is the iterative convergence term, reflecting the number of iterations. The impact on convergence accuracy: the more iterations, the faster this term decays; the second term... The first term is the noise term, reflecting the contribution of local learning noise to the equalization tracking error. The stronger the noise, the higher the lower limit of the tracking error. The third term... This is the drift term, reflecting the impact of equilibrium drift caused by changes in global configuration commands on tracking error. The more drastic the changes in the leader's commands, the more difficult it is for the followers to accurately track the new equilibrium position. This formula indicates that only simultaneous adjustment... , , and Only in this way can we balance tracking accuracy and overall system stability within a limited time. The adaptively adjusted parameters are then applied to the leader decision-making and follower iteration processes in subsequent cycles, thus forming a closed-loop control with an adaptive solution rhythm.

[0214] This embodiment also provides a hierarchical collaborative perception task control system for UAV formations, including:

[0215] The status and constraint acquisition module is used to acquire the current environmental status data of the UAV formation, the operation data of each UAV, and the constraint pressure data. The constraint pressure data is updated cumulatively based on the deviation of at least one type of long-term operation constraint index.

[0216] The global instruction generation module is used to generate global configuration instructions based on the environmental state data, the operation data of each UAV, and the constraint pressure data.

[0217] The local instruction acquisition module is used to issue the global configuration instruction to each execution UAV and acquire the local instruction set of the local configuration instruction determined by each UAV under the common constraints of the global configuration instruction, its own local observation data and constraint pressure data.

[0218] The constraint update module is used to obtain feedback information generated after executing the local instruction set, and update the constraint pressure data according to the deviation between the feedback information and the long-term operating constraint index.

[0219] The iterative control module is used to take the updated constraint pressure data as input for the next decision cycle and trigger the global instruction generation module.

[0220] Optionally, an adaptive adjustment module is also included for:

[0221] Calculate the total constraint pressure metric based on the constraint pressure data;

[0222] Obtain a first metric that characterizes the degree of drastic change in the global configuration instructions;

[0223] Obtain a second metric that characterizes the uncertainty in the process of determining local configuration commands for each UAV;

[0224] Based at least one of the first metric, the second metric, and the total constraint pressure metric, adjust the global step size parameter used when generating the global configuration command for the next decision cycle, and / or adjust the local update step size and iteration number threshold used by each UAV when determining the local configuration command for the next decision cycle.

[0225] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the method described in any of the above method embodiments.

[0226] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the above method embodiments.

[0227] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for hierarchical collaborative perception task control of unmanned aerial vehicle (UAV) formations, characterized in that, include: Acquire current environmental status data of the drone formation, operational data of each drone, and current constraint pressure data; Based on the environmental status data, the operational data of each UAV, and the constraint pressure data, a global configuration command is generated. Obtain the set of local instructions for each UAV under the common constraints of the global configuration instructions, its own local observation data, and constraint pressure data; Obtain feedback information generated after executing the set of local instructions, and update the constraint pressure data based on the deviation between the feedback information and at least one preset long-term operating constraint index. The updated constraint pressure data is used as input for the next decision cycle, and the process returns to the step of generating global configuration instructions.

2. The method according to claim 1, characterized in that, The method further includes: obtaining task completion data from the previous decision cycle before executing the step of generating global configuration instructions; the generation of global configuration instructions is also based on the task completion data; Based on the feedback information generated by executing the set of local instructions, the task completion data is updated, and the updated task completion data is used as the input for the next decision cycle.

3. The method according to claim 2, characterized in that, The generation of global configuration instructions includes: Take a candidate instruction from the preset solution space as the current candidate instruction, and calculate the first optimization target value of the current candidate instruction based on the environmental state data, the operation data of each UAV, the local instruction set of the previous decision cycle, the task completion data and the constraint pressure data. With the goal of increasing the first optimization target value, the current candidate instruction is iteratively adjusted in the solution space until a preset first termination condition is met. The candidate instruction that reaches the first termination condition is then used as the generated global configuration instruction.

4. The method according to claim 3, characterized in that, The calculation of the first optimization target value of the current candidate instruction includes: Based on the environmental status data, the operational data of each UAV, the local instruction set of the previous decision cycle, and the task completion data, calculate the global task benefit item; Calculate the constraint penalty term based on the constraint pressure data; The first optimization target value is obtained by taking the weighted difference between the global task benefit term and the constraint penalty term.

5. The method according to claim 3, characterized in that, The step of iteratively adjusting the current candidate instructions within the solution space, with the goal of increasing the first optimization objective value, includes: In this iteration, for each configuration parameter in the current candidate instructions, the partial derivative of the first optimization target value with respect to each configuration parameter is calculated to form a gradient vector; the configuration parameters include at least one of the following: perception resource allocation parameters, task weight setting parameters, formation reference configuration parameters, and perception region division parameters. Based on the gradient vector and the preset global step size parameter, calculate the adjustment amount of each configuration parameter; The current candidate instruction is updated according to each adjustment amount, and the feasibility of the updated candidate instruction is verified. If it exceeds the solution space, it is projected into the solution space and used as the current candidate instruction after this iteration adjustment; otherwise, the updated candidate instruction is directly used as the current candidate instruction after this iteration adjustment; the candidate instruction obtained after this iteration adjustment is used as the current candidate instruction for the next iteration.

6. The method according to claim 1, characterized in that, The steps to determine local configuration directives include: Each drone determines its own action constraint range based on the received global configuration instructions; Each execution drone acquires candidate local configuration instructions for this iteration, and calculates the second optimization target value corresponding to the candidate local configuration instructions based on its own local observation data and constraint pressure data; With the goal of increasing the second optimization target value, the current candidate local configuration instruction is iteratively adjusted within the constraints of its own actions until the preset second termination condition is met; The candidate local configuration instruction that is met when the second termination condition is met is used as the generated local configuration instruction.

7. The method according to claim 2, characterized in that, The step of updating the task completion data includes: Extract at least one of the following from the feedback information: target coverage area index, target recognition accuracy index, and multi-machine information fusion consistency index; The extracted metrics are directly assigned to the corresponding fields in the task completion data to update the task completion data.

8. The method according to claim 1, characterized in that, The step of updating the constraint pressure data includes: Identify the deviation of the feedback information from various long-term operating constraints, the types of which include at least one of energy consumption constraints, collision risk constraints, communication load constraints, and formation deviation constraints; For each type of long-term operating constraint, the deviation is compared with a preset constraint threshold. If the deviation exceeds the constraint threshold, the corresponding pressure value in the constraint pressure data is accumulated; otherwise, the pressure value is reduced, and the reduced pressure value is ensured to be greater than or equal to zero. The constraint pressure data after accumulation or reduction is used as the updated constraint pressure data.

9. The method according to claim 1, characterized in that, Before performing the step of returning to generate the global configuration instruction, the method further includes: Obtain a first metric characterizing the drastic degree of change in the global configuration instructions; Obtain a second metric that characterizes the uncertainty in the process of each executing UAV determining local configuration commands; Based at least on one of the first metric, the second metric, and the constraint pressure data, adjust the global step size parameter used when generating the global configuration command for the next decision cycle, and / or adjust the iteration number threshold and local update step size used by each executing UAV when determining the local configuration command in the next decision cycle.

10. The method according to claim 9, characterized in that, The threshold for adjusting the number of iterations used by each UAV in determining local configuration instructions in the next decision cycle includes: Calculate the square root of the sum of squares of all pressure values ​​in the constraint pressure data, and use the result as the total constraint pressure measurement value; Based on the total constraint pressure metric, determine the required solution accuracy for the next decision cycle; Based on the required solution accuracy, the threshold for the number of iterations when determining local configuration instructions in the next decision cycle is increased.

11. The method according to claim 9, characterized in that, The adjustment of the local update step size used by each executing UAV when determining local configuration instructions in the next decision cycle includes: Obtain the noise intensity value represented by the second metric; Based on the noise intensity value, and following the mapping relationship that the larger the noise intensity, the smaller the step size, the local update step size used when determining the local configuration instruction in the next decision cycle is reduced.

12. The method according to claim 9, characterized in that, The global step size parameters used when adjusting the global configuration instructions for the next decision cycle include: Obtain the total constraint pressure metric determined by the constraint pressure data, and the decision drift value determined by the first metric; Based on the total constraint pressure metric and the decision drift degree, and following the linkage adjustment mapping relationship where a larger total constraint pressure metric corresponds to a smaller step size and a larger decision drift degree corresponds to a smaller step size, the global step size parameter for the next decision cycle is calculated.

13. The method according to claim 10, characterized in that, The method further includes: Based on the required solution accuracy, the first metric, and the second metric, the cycle length of the next decision cycle is calculated; wherein the cycle length increases with the increase of the solution accuracy represented by the required solution accuracy, increases with the increase of the first metric, and increases with the increase of the second metric. Based on the calculated cycle length, the time span of the next decision cycle is controlled, so that the step of returning to execute the generation of global configuration instructions is triggered when the time condition corresponding to the cycle length is met.

14. A hierarchical collaborative perception and control system for unmanned aerial vehicle (UAV) formations, characterized in that, include: The status and constraint acquisition module is used to acquire the current environmental status data of the UAV formation, the operation data of each UAV, and the constraint pressure data. The constraint pressure data is updated cumulatively based on the deviation of at least one type of long-term operation constraint index. The global instruction generation module is used to generate global configuration instructions based on the environmental state data, the operation data of each UAV, and the constraint pressure data. The local instruction acquisition module is used to issue the global configuration instruction to each execution UAV and acquire the local instruction set of the local configuration instruction determined by each UAV under the common constraints of the global configuration instruction, its own local observation data and constraint pressure data. The constraint update module is used to obtain feedback information generated after executing the local instruction set, and update the constraint pressure data according to the deviation between the feedback information and the long-term operating constraint index. The iterative control module is used to take the updated constraint pressure data as input for the next decision cycle and trigger the global instruction generation module.

15. The system according to claim 14, characterized in that, It also includes an adaptive adjustment module for: Calculate the total constraint pressure metric based on the constraint pressure data; Obtain a first metric that characterizes the degree of drastic change in the global configuration instructions; Obtain a second metric that characterizes the uncertainty in the process of determining local configuration commands for each UAV; Based at least one of the first metric, the second metric, and the total constraint pressure metric, adjust the global step size parameter used when generating the global configuration command for the next decision cycle, and / or adjust the local update step size and iteration number threshold used by each UAV when determining the local configuration command for the next decision cycle.

Images