Method for scheduling autonomous cooperative reconnaissance task of unmanned aerial vehicle cluster

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) control and mission planning technology, specifically disclosing a method for scheduling autonomous collaborative reconnaissance missions of UAV swarms. The method involves real-time collection and fusion of UAV flight status, environmental perception, and mission situation data to calculate mission effectiveness characteristic values, environmental risk indices, and swarm collaboration requirement coefficients. An adaptive control parameter is dynamically generated based on these coefficients using a meta-learning policy network. Each UAV then uses these parameters to perform distributed intent inference, forming a probability distribution of its belief in the intentions of other UAVs. This distribution is used to construct a trajectory optimization problem incorporating probabilistic collaborative collision avoidance constraints. The optimal collaborative trajectory control command is obtained through distributed solution. The command is executed, and collaborative effect data is collected, with online iterative updates to the policy network and belief distribution. This invention achieves adaptive adjustment of control parameters to dynamic missions, improving the collaborative reconnaissance efficiency and overall autonomy of the swarm in complex and uncertain environments.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control and mission planning technology, specifically to a method for scheduling autonomous and collaborative reconnaissance missions of UAV swarms. Background Technology

[0002] With the rapid development of drone technology, utilizing multiple drones to form swarms for collaborative reconnaissance missions has become an important development direction for improving area awareness and enhancing mission robustness. The core challenge of drone swarm collaborative reconnaissance lies in how to autonomously and in real-time assign appropriate reconnaissance tasks and flight paths to each drone in dynamic, uncertain, and complex environments, while ensuring that the entire swarm can efficiently execute its missions while avoiding collisions and maintaining a stable collaborative relationship.

[0003] The existing technology has the following shortcomings: In dynamic and uncertain complex reconnaissance environments, traditional UAV swarm collaboration methods, due to their use of fixed-parameter optimization frameworks and deterministic collaboration constraints, cannot simultaneously achieve adaptive adjustment of control strategies to the real-time mission situation and quantitative fusion of uncertainties regarding the intentions of friendly aircraft. As a result, it is difficult for the swarm's autonomous collaborative decision-making to achieve a fundamental balance between mission efficiency and flight safety. Summary of the Invention

[0004] The purpose of this invention is to provide a method for scheduling autonomous and collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms, in order to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: The method for scheduling autonomous collaborative reconnaissance missions by unmanned aerial vehicle (UAV) swarms includes the following steps: S1: Real-time acquisition of flight status data of each UAV in the cluster, environmental perception data obtained by airborne sensors, and mission situation data obtained by parsing mission instructions; S2: The flight status data, environmental perception data and mission situation data are fused and processed to calculate the mission effectiveness characteristic value, real-time environmental risk index and cluster collaboration requirement coefficient of each UAV. S3: Based on task performance feature values, environmental risk index and cluster collaboration demand coefficient, a set of adaptive control parameters are dynamically adjusted and generated through meta-learning strategy network. The adaptive control parameters include prediction time domain length adjustment coefficient, multi-objective cost function weight coefficient and constraint relaxation priority sequence. S4: Each UAV uses adaptive control parameters to perform distributed intent inference and form a belief probability distribution of the intentions of other UAVs; based on the belief probability distribution, a trajectory optimization problem containing probabilistic cooperative collision avoidance constraints is constructed, and the optimal cooperative trajectory control command of each UAV is obtained through distributed solution. S5: Execute the optimal cooperative trajectory control command to complete the cooperative reconnaissance action; collect the cluster cooperative effect data after the command execution, use the cluster cooperative effect data to fine-tune the meta-learning policy network online, and iteratively update the belief probability distribution.

[0006] As a further aspect of the present invention: the calculation process for the task performance feature value is as follows: Extract energy state parameters and maneuverability parameters from flight status data; Extract target value parameters and task time limit parameters from the task situation data; Normalize the energy state parameters, maneuverability parameters, target value parameters, and mission time limit parameters; Based on the terrain shading factor in the environmental perception data, the weight coefficients of each parameter after normalization are dynamically adjusted. The task performance feature value is output by weighting and fusing the results based on the adjusted weighting coefficients.

[0007] As a further aspect of the present invention: the calculation process of the real-time environmental risk index is as follows: Calculate the relative motion relationship between each threat source and the drone to obtain the threat proximity. By combining meteorological conditions and terrain permeability from environmental perception data, the visibility and impact range of threat sources are calculated. Based on the threat proximity and impact range, a local risk map is generated using risk overlay rules; The local risk map is integrated and normalized to output a real-time environmental risk index.

[0008] As a further aspect of the present invention: the calculation process of the cluster collaboration demand coefficient is as follows: Determine the collaborative task mode based on task status data; Obtain the mission performance characteristics and real-time environmental risk index of each UAV; Based on the collaborative task model, evaluate the overall cluster performance and risk balance. Based on the overall cluster performance and risk balance, calculate the expected synergistic gains of different formation configurations; Based on the current cluster communication link quality, select the gain value corresponding to the optimal formation configuration and output the cluster coordination requirement coefficient.

[0009] As a further aspect of the present invention: the dynamic adjustment and generation of a set of adaptive control parameters through a meta-learning policy network specifically includes: The task performance feature value, environmental risk index, and cluster collaboration requirement coefficient are correlated with the stored historical collaboration decision records in multiple dimensions to form the current task context feature vector. The current task context feature vector is input into the meta-learning policy unit. Through the multi-layer attention matching and weight allocation structure inside the meta-learning policy unit, the initial prediction time domain length adjustment coefficient, multi-objective cost function weight coefficient, and constraint relaxation priority sequence are generated. Based on the execution feedback of the current cluster's real-time actions and the estimated achievement of the task objectives, the initial adaptive control parameters are incrementally optimized and corrected online, and the final adaptive control parameters used for trajectory optimization are output.

[0010] As a further aspect of the present invention: the formation of the belief probability distribution regarding the intentions of other drones specifically includes: Each UAV generates a set of candidate trajectory segments for multiple future moments and the corresponding intent priority of the UAV based on its own adaptive control parameters and real-time status. Receive limited historical trajectory fragments of nearby drones and some publicly released intent priorities of drones through the communication link; Based on the received fragment and priority information, the belief probability of all neighboring UAVs taking each candidate trajectory within the planning time domain is calculated and updated through a Bayesian inference process that associates state transition with intent.

[0011] As a further aspect of the present invention: the step of obtaining the optimal cooperative trajectory control command for each UAV through distributed solution specifically includes: Each UAV constructs a local optimization problem based on its own generated set of candidate trajectory segments and belief probability distribution. The main optimization objective is to minimize its own trajectory cost, and the core constraint is to keep the probability of conflict with the trajectory of neighboring UAVs below a set threshold. The system exchanges current preferred trajectories and corresponding constraint satisfaction with neighboring drones through multiple rounds of communication, and dynamically adjusts its own candidate trajectory selection based on the exchanged information. When the conflict probabilities among the preferred trajectories reported by all drones meet the constraints, it is determined that a distributed consensus has been reached, and the control command corresponding to the trajectory selected by the drone at this time is taken as the optimal cooperative trajectory control command.

[0012] As a further aspect of the present invention: the completion of the coordinated reconnaissance action specifically includes: Each UAV executes the optimal cooperative trajectory control command and uses its onboard sensors to scan and collect information about the designated area in real time during flight. When a drone detects target features that meet preset conditions during scanning, it immediately triggers local action adjustments. Based on the target's geographical features and task priority, it calculates and updates the drone's own task performance feature value and environmental risk index in real time. By broadcasting updated information to nearby drones via communication links, the drones are guided to dynamically adjust their reconnaissance paths and sensor orientations based on the updated mission effectiveness characteristics and environmental risk index, thereby forming a collaborative reconnaissance operation among multiple drones that dynamically converges around the target characteristics.

[0013] As a further aspect of the present invention: the online fine-tuning of the meta-learning policy network using cluster collaboration effect data and the iterative updating of the belief probability distribution specifically includes: After the collaborative reconnaissance operation is completed, collect multi-dimensional collaborative effect data, including mission area coverage, target confirmation time, overall cluster energy consumption, and command execution deviation. By correlating and comparing the multidimensional collaborative effect data with the task context feature vector before the current action adjustment, the strategy evaluation signal is calculated. Using the policy evaluation signal, the internal weight allocation structure of the meta-learning policy unit is slightly modified through the gradient backpropagation process. The actual trajectories executed by each UAV are matched with the trajectories predicted in the previous belief probability distribution. The confidence increment is calculated for the corresponding belief probability distribution entries based on the matching degree. The belief probability distributions of the action intentions of all neighboring UAVs are then updated using a weighted iterative method.

[0014] The beneficial effects of this invention are: (1) This invention introduces a parameter dynamic generation mechanism driven by a meta-learning strategy (step S3), enabling the cluster to adjust the core parameters of trajectory optimization (such as prediction time domain and cost weight) online based on situational information such as task performance feature values ​​and environmental risk indexes obtained through real-time fusion computing. More importantly, through intention prediction based on distributed Bayesian inference, each UAV can proactively infer the possible actions of its neighbors, rather than passively responding, thereby avoiding conflicts in advance and achieving proactive collaboration at the planning level. This closed loop of "perception-inference-adaptive optimization" enables the cluster to act like an organic whole, quickly and flexibly reorganizing its formation and task allocation when facing complex scenarios such as suddenly discovered signs of life in valley search and rescue or high-speed escape of targets in urban tracking, thereby improving the success rate of tasks and the efficiency of resource utilization.

[0015] (2) This invention introduces probabilistic cooperative collision avoidance constraints into the distributed trajectory optimization problem. Instead of assuming that the trajectories of their neighbors are fixed, each UAV calculates the overall probability of its own trajectory colliding with the possible trajectories of all its neighbors based on the probability distribution of their beliefs about the neighbors' intentions, and strictly constrains this probability below a safety threshold. This method essentially transforms the uncertain intentions of neighbors into a quantifiable risk constraint, allowing UAVs to more boldly plan optimized trajectories conducive to cooperative reconnaissance (such as clustered observation and cross-coverage) while ensuring an extremely low collision probability. This fundamentally eliminates the collision risk caused by communication delays or prediction errors, and unleashes the potential of cooperative actions, thus achieving a balance between safety and efficiency in scenarios with stringent requirements for both security and effectiveness, such as wide-area coverage border patrols and covert penetration in combat environments. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the method described in this invention. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1 As shown, this invention is a method for scheduling autonomous collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms, comprising the following steps: S1: Real-time acquisition of flight status data of each UAV in the cluster, environmental perception data obtained by airborne sensors, and mission situation data obtained by parsing mission instructions; S2: The flight status data, environmental perception data and mission situation data are fused and processed to calculate the mission effectiveness characteristic value, real-time environmental risk index and cluster collaboration requirement coefficient of each UAV. S3: Based on task performance feature values, environmental risk index and cluster collaboration demand coefficient, a set of adaptive control parameters are dynamically adjusted and generated through meta-learning strategy network. The adaptive control parameters include prediction time domain length adjustment coefficient, multi-objective cost function weight coefficient and constraint relaxation priority sequence. S4: Each UAV uses adaptive control parameters to perform distributed intent inference and form a belief probability distribution of the intentions of other UAVs; based on the belief probability distribution, a trajectory optimization problem containing probabilistic cooperative collision avoidance constraints is constructed, and the optimal cooperative trajectory control command of each UAV is obtained through distributed solution. S5: Execute the optimal cooperative trajectory control command to complete the cooperative reconnaissance action; collect the cluster cooperative effect data after the command execution, use the cluster cooperative effect data to fine-tune the meta-learning policy network online, and iteratively update the belief probability distribution.

[0020] In S1, flight status data of each UAV in the cluster, environmental perception data acquired by airborne sensors, and mission situation data obtained by parsing mission instructions are collected in real time. The acquisition of flight status data relies on the navigation and status awareness modules mounted on the UAV itself. Specifically, by integrating the outputs of the global navigation satellite system receiver, inertial measurement unit, and airspeed indicator, the UAV's geospatial position, three-dimensional velocity, flight attitude angles, and heading angles are acquired and updated in real time. Simultaneously, the flight control computer reads and calculates the UAV's real-time remaining available energy, current flight time, and health status information of each actuator from the power system and power management unit.

[0021] Environmental perception data acquisition is accomplished collaboratively by multiple heterogeneous sensors carried by the UAV. LiDAR or millimeter-wave radar actively scans the area in front of and below the UAV, generating point cloud data containing terrain elevation and obstacle outlines. Visible light or infrared camera modules image designated areas, and onboard processing units extract key features from the images in real time. Additionally, meteorological sensors collect local wind speed, wind direction, and temperature data. After timestamp alignment and coordinate system unification, this raw perception data forms multimodal perception information describing the real-time physical environment surrounding the UAV.

[0022] The acquisition of mission situation data stems from the parsing and fusion of received mission instructions from higher authorities. The UAV continuously receives global mission instructions and updates broadcast from command nodes or lead aircraft via data link. The onboard mission management module parses these instructions, extracting the geographical range of the target area to be reconnoitered in the current phase, the priority weights of each sub-target, the overall time window constraints for mission execution, and the preset communication relay points or no-fly zone boundary information. This module then preliminarily correlates the parsed structured information with the UAV's own flight status and environmental perception data to form a preliminary understanding of the local mission situation.

[0023] In S2, flight status data, environmental perception data, and mission situation data are fused and processed to calculate the mission effectiveness characteristic value, real-time environmental risk index, and cluster collaboration requirement coefficient for each UAV. Specifically, this includes: The calculation process for mission effectiveness characteristic values ​​is as follows: First, two basic parameters are extracted from flight status data: energy state parameter and maneuverability parameter. The energy state parameter is obtained by dividing the current remaining battery power by the full battery capacity, and its value is between 0 and 1. The maneuverability parameter is obtained through a composite calculation, which divides the current airspeed by the maximum safe flight speed to obtain one ratio, and divides the current available overload margin by the maximum design overload to obtain another ratio. The smaller of these two ratios is taken as the maneuverability parameter, and its value is also between 0 and 1. Second, the target value parameter and mission time limit parameter are extracted from mission situation data. The target value parameter is directly mapped according to the priority assignment of the reconnaissance subtargets to be reconnoitered by the UAV in the mission instructions. High priority is mapped to a value of 1, medium priority to a value of 0.6, and low priority to a value of 0.3. The mission time limit parameter is obtained by calculating the percentage of remaining time from the current time point to the latest time specified in the mission. For example, it is 1 when there is sufficient remaining time, and 0 when there is zero remaining time. Next, the four parameters are normalized. Since they are already in the range of 0 to 1, this step mainly ensures their scaling consistency. Then, a terrain occlusion factor from the environmental perception data is introduced for dynamic weight adjustment. The terrain occlusion factor, calculated by analyzing LiDAR point cloud data, represents the average degree of obstruction of signals and line of sight by the terrain surrounding the UAV's current location, with a value between 0 (completely open) and 1 (completely obstructed). Based on this factor, the weights of the maneuverability parameter are adjusted: for every 0.1 increase in the occlusion factor, the weight coefficient of the maneuverability parameter decreases by 5% of its initial value. Finally, the adjusted weight coefficients of each parameter are normalized to a sum of 1. This sum is then used to weight and sum the values ​​of the energy state parameter, maneuverability parameter, target value parameter, and mission time limit parameter. The result is the mission effectiveness characteristic value of the UAV.

[0024] The calculation process of the real-time environmental risk index is as follows: First, from the static obstacles and dynamic threat sources (such as other aircraft and known air defense radar illumination areas) identified in the environmental perception data, for each threat source, its relative motion relationship with the UAV is calculated to obtain the threat proximity degree. The threat proximity degree is obtained through a two-dimensional calculation: the first dimension is the distance proximity degree, which is obtained by dividing the current distance between the threat source and the UAV by a preset safety alarm distance threshold. If this value exceeds 1, it is taken as 1. The second dimension is the velocity proximity degree, which is obtained by calculating the magnitude of the component of the relative velocity vector of the threat source pointing towards the UAV and dividing it by the maximum speed of the UAV. This value is also limited to between 0 and 1. The final threat proximity degree is the average of the values ​​of these two dimensions. Second, combined with the real-time meteorological conditions (such as wind speed and visibility) and terrain transparency (the line-of-sight obstruction rate calculated by the digital elevation model) in the environmental perception data, the visibility and impact range correction factor of the threat source are calculated. For example, when the visibility is lower than the preset threshold, the visibility of the dynamic threat source decreases, and its impact range correction factor increases accordingly. Then, based on the threat proximity and modified impact range of each threat source, a local risk map centered on the drone is generated using risk overlay rules. The risk overlay rules stipulate that at each geographic grid point on the risk map, the risk value is the sum of the risk values ​​caused by all threat sources at that point. The risk value caused by a single threat source at a point is equal to the threat proximity of that source divided by the distance from that point to the geometric center of the threat source (which must be greater than a minimum distance constant). Next, the generated local risk map is numerically integrated within the fan-shaped region for drone planning, i.e., the risk values ​​of all grid points within that region are accumulated. Finally, this integrated result is divided by the total number of grids in the fan-shaped region for regional averaging, and then the average value is divided by a preset maximum risk reference value for normalization, outputting a real-time environmental risk index between 0 and 1.

[0025] The calculation process for the cluster collaboration requirement coefficient is as follows: First, based on the clearly defined task type and instructions in the task situation data, the collaborative task mode is determined. Task modes are predefined into several categories, such as "wide-area coverage search," "multi-target tracking," and "penetration reconnaissance." Second, the task performance characteristic values ​​and real-time environmental risk indices of all UAVs within the cluster, calculated through the aforementioned process, are obtained. Next, based on the selected collaborative task mode, the overall cluster performance and risk balance are evaluated. Overall performance is obtained by calculating the arithmetic mean of the task performance characteristic values ​​of all UAVs; risk balance is obtained by calculating the standard deviation of the real-time environmental risk indices of all UAVs, with a smaller standard deviation indicating a more balanced risk distribution. Then, based on the calculated overall performance and risk balance, a pre-defined collaborative gain mapping table is queried. This mapping table defines several recommended formation configurations (such as "diamond," "linear," and "ring") and their corresponding expected collaborative gain values ​​for each collaborative task mode, different overall performance ranges, and risk balance ranges. The expected collaborative gain value is a value greater than 0; a higher value indicates a better expected collaborative effect of the configuration under the current situation. Finally, based on the current cluster communication link quality obtained through communication link status awareness (comprehensively evaluated by average signal-to-noise ratio and packet loss rate), configurations that meet the communication robustness requirements are selected from the formation configurations recommended by the mapping table, and the configuration with the highest expected cooperative gain value is selected. The gain value corresponding to this configuration is directly output as the cluster cooperative requirement coefficient.

[0026] In S3, based on task performance feature values, environmental risk index, and cluster collaboration requirement coefficient, a set of adaptive control parameters is dynamically adjusted and generated through a meta-learning policy network. These adaptive control parameters include a prediction time-domain length adjustment coefficient, a multi-objective cost function weight coefficient, and a constraint relaxation priority sequence, specifically including: The formation process of the current task context feature vector is as follows: First, the task performance feature value of the UAV calculated at the current moment, its real-time environmental risk index, and the cluster collaboration requirement coefficient are combined into a three-dimensional basic vector. Second, a search and matching process is performed from the historical collaborative decision record database stored locally by the UAV. Each historical record in this database contains the aforementioned three-dimensional basic vector stored under similar task stages in the past, as well as the combination of adaptive control parameters that were actually adopted and verified to be effective in that historical context. The matching process is achieved by calculating the Euclidean distance between the current three-dimensional basic vector and the vector in each historical record, specifically by calculating the square root of the sum of the squares of the differences in the corresponding dimensions. The five historical records with the smallest Euclidean distance are selected, and their corresponding historical adaptive control parameters are extracted. Finally, the current three-dimensional basic vector and these five sets of historical parameters together constitute an extended feature set. After flattening, this set forms the current task context feature vector.

[0027] The initial adaptive control parameter mapping generation process is as follows: The current task context feature vector formed in the previous step is input into a pre-trained parameter mapping structure. This structure contains an attention matching layer and a weight allocation layer. The attention matching layer first calculates the similarity weights between the current feature vector and the context vectors corresponding to the five sets of historical parameters. The similarity weights are calculated by taking the square of the cosine similarity between the current vector and each historical context vector, and then applying a normalization function to make the sum of the five weights equal to 1. The weight allocation layer then uses the calculated five similarity weights to perform weighted fusion on the prediction time-domain length adjustment coefficient, the multi-objective cost function weight coefficient, and the constraint relaxation priority sequence in the five sets of historical parameters. For example, the initial value of the prediction time-domain length adjustment coefficient is equal to the sum of the coefficient values ​​in the five historical records multiplied by their corresponding similarity weights. Through this mapping process, a preliminary set of adaptive control parameters suitable for the current context is generated.

[0028] The online optimization and correction process for adaptive control parameters is as follows: After generating initial parameters and applying them to short-term trajectory control, an incremental optimization loop is initiated. This loop collects real-time feedback data after the cluster executes actions within a preset time window, including the actual coverage rate change rate of the task area and the average deviation from the planned trajectory. Simultaneously, based on the current environmental perception and task situation data, the achievement rate of local task objectives for the next time segment is predicted, such as predicting the proportion of unknown areas that can be covered within the next thirty seconds under current parameters. Next, the difference between the feedback data and the expected target is calculated: the actual coverage rate change rate is compared with a preset expected rate threshold, and the average trajectory deviation is compared with an allowable deviation threshold. If the actual rate is lower than 90% of the expected threshold, "insufficient coverage efficiency" is determined; if the average deviation is greater than 120% of the allowable threshold, "insufficient tracking performance" is determined. These qualitative judgments and the magnitude of the difference together constitute a strategy evaluation signal. This signal is used to fine-tune the initial parameters: if "coverage efficiency is insufficient," the weight coefficient of the term positively correlated with coverage in the cost function is increased by a fixed step size (e.g., 0.05); if "tracking performance is insufficient," the prediction time domain length adjustment coefficient is decreased by a fixed step size. This fine-tuning process continues until the evaluation signal indicates that the difference between the actual effect and the expected target converges to an acceptable range. At this point, the finely tuned parameters output are the final adaptive control parameters used for the next stage of trajectory optimization.

[0029] In S4, each UAV uses adaptive control parameters to perform distributed intent inference, forming a belief probability distribution of the intentions of other UAVs. Based on this belief probability distribution, a trajectory optimization problem with probabilistic cooperative collision avoidance constraints is constructed, and the optimal cooperative trajectory control command for each UAV is obtained through distributed solution, specifically including: The process of forming a probability distribution of beliefs about the actions of other drones is as follows: First, each drone generates a set of candidate trajectory segments for a planned future time domain based on its own adaptive control parameters and real-time flight status. This set is generated by: using the drone's current position and velocity as the initial state, and based on its dynamic constraints, by sampling different control input sequences, forward simulation is used to deduce multiple possible forward paths. Each path corresponds to a state sequence over several future time periods, i.e., a trajectory segment. Simultaneously, an intent priority is calculated for each trajectory segment. This priority is obtained through a composite calculation, comprehensively considering the value of the sub-target the segment points to, estimated energy consumption, and the degree of matching with the drone's current mission performance characteristics. Second, the drone receives limited historical trajectory segments periodically broadcast by its direct communication neighbors (i.e., neighboring drones) through its established communication link. This broadcast information includes the trajectory coordinate sequence segments actually executed by neighboring drones within a short time window in the past, as well as simplified priority labels (e.g., "heading to area A" or "tracking target B") regarding their currently publicly declared main action intentions for the next stage. Finally, based on the received historical trajectory fragments and intent priority labels of the neighbor, a Bayesian inference process is initiated to update the probability distribution of the belief in the neighbor's future actions. The core of this inference is calculating the posterior probability that, given the observed specific historical trajectory fragments and intent labels, the neighbor will execute a trajectory fragment from its own candidate set. The key to this calculation lies in assessing the likelihood of a correlation between state transitions and intents. Specifically, for neighboring drones… (Use subscripts) Each candidate trajectory segment (represented) ( (for fragment indexes), their belief probabilities The update calculation follows the formula: ; In this formula: This indicates the neighbor's drone before the update. Execution trajectory segment The probability of prior beliefs; This represents the updated posterior belief probability; Indicates the observed neighboring drones Fragments of historical trajectory; Indicates the observed neighboring drones The intent priority tag for the release; It is a likelihood function, representing the likelihood of a neighboring drone. The true intention is the execution trajectory. Under these conditions, its historical fragments were observed to be And the published tag is The likelihood value is calculated using a predefined evaluation function: the function value is proportional to the historical data. With candidate trajectory fragments The initial part's spatial morphological matching degree, multiplied by the intention label. With trajectory The semantic consistency coefficient of the implied intent (the consistency coefficient is between 0 and 1). Indicates the use of neighboring drones All candidate trajectory segments.

[0030] The denominator is all possible candidate trajectory segments. The summation of the corresponding numerators ensures that the sum of all posterior probabilities is 1.

[0031] By iteratively performing the above Bayesian update on each neighboring drone, the belief probability distribution of each candidate trajectory that all neighboring drones may take within the planning time domain is finally obtained.

[0032] The process of obtaining the optimal cooperative trajectory control commands for each UAV through distributed solution is as follows: First, each UAV constructs a local trajectory optimization problem based on its own generated candidate trajectory fragment set and the belief probability distribution updated in the previous step. The goal of this optimization problem is to select a trajectory from its own candidate set that minimizes the estimated cost of flying along that trajectory. The cost calculation comprehensively considers flight energy consumption, time cost to reach the sub-target, and trajectory smoothness. The core constraint is that the overall probability of the selected UAV's trajectory spatially conflicting with the trajectory that any neighboring UAV might take (i.e., the distance is less than a preset safety radius) must be lower than a set safety threshold, such as 0.1%. This conflict probability is calculated based on the belief probability distribution: for a candidate trajectory of itself... Its neighboring drones Conflict probability Calculated by traversing the neighbors. All candidate trajectory segments ,like and If the spatial distance between two points at the same time is less than the safe radius, it is considered a conflict event, and the neighbor trajectory belief probability corresponding to the event is calculated. This is added to the conflict probability. The total conflict probability is the maximum value of the conflict probabilities with all neighbors, that is: The optimization problem is to satisfy: Under the constraints of cost, find ways to reduce costs smallest Secondly, since this optimization relies on beliefs about the intentions of neighbors (which may change due to the drone's own choices), distributed negotiation through multiple rounds of communication is required. Each drone selects the lowest-cost feasible trajectory from its conflict-constraining feasible trajectories as its "preferred trajectory" for the current round, and broadcasts its corresponding estimated cost and calculated maximum conflict probability (i.e., constraint satisfaction) to its neighbors. Upon receiving the preferred trajectory information from its neighbors, each drone treats it as new strong evidence of the neighbor's intention, immediately uses the aforementioned Bayesian update process to significantly increase its belief probability for that specific neighbor trajectory, and re-solves its own local optimization problem based on the updated belief to generate a new preferred trajectory. This process is iterative. Finally, a consensus criterion is set: when all drones' reported preferred trajectories remain unchanged in two consecutive rounds of communication, and the conflict probability between each drone's own preferred trajectory calculated based on its current belief and all neighbor preferred trajectories is below a safety threshold, the cluster is considered to have reached a distributed consensus. At this point, each UAV will output the sequence of control commands corresponding to the preferred trajectory selected by itself under consensus (i.e., the speed, heading, and other control quantities required to achieve the trajectory) as the optimal cooperative trajectory control command for the current planning cycle to the flight control execution unit.

[0033] In S5, the optimal cooperative trajectory control command is executed to complete the cooperative reconnaissance action; data on the cluster cooperative effect after command execution is collected, and the cluster cooperative effect data is used to fine-tune the meta-learning policy network online, and the belief probability distribution is iteratively updated, specifically including: The process of completing the coordinated reconnaissance operation is as follows: Each UAV receives and executes the optimal coordinated trajectory control command sequence derived from the previous step. During flight along the planned trajectory, the UAV utilizes its onboard multispectral imaging sensor or synthetic aperture radar to perform strip or point-to-point scanning of the geographical area specified in the mission plan, achieving real-time image and signal information collection. When any UAV, through its onboard processing algorithm, identifies a target matching a preset feature template (e.g., thermal signature or radio signature of a vehicle of a specific size) in the real-time scan data, a local action adjustment event is triggered. Upon triggering, the UAV immediately suspends the execution of subsequent non-urgent commands in the original trajectory. First, based on the precise geographic coordinates, type attributes, and preset mission rule base of the newly discovered target, it recalculates its own mission effectiveness characteristic value and environmental risk index. Specifically, the target's geographic coordinates are used to reassess its association value with each predetermined sub-target, thereby updating the target value parameters; simultaneously, the terrain surrounding the target is reassessed to update the terrain masking factor. Subsequently, the UAV broadcasts its updated mission effectiveness characteristic value, environmental risk index, and basic target information via its data link with high priority. Upon receiving the broadcast information, the nearby UAVs immediately integrate it into their real-time situational data and, based on the updated global information, initiate a simplified local trajectory replanning. The goal of the replanning is to ensure that, while meeting basic collision avoidance and endurance constraints, the UAVs quickly align their sensor coverage with or approach the area where the newly discovered target is located. This allows them to spontaneously form a dynamic, clustered reconnaissance formation at the swarm level, centered on the target and acquiring information from multiple directions and angles.

[0034] The process of online fine-tuning and iterative updates using swarm collaboration effect data is as follows: After a collaborative reconnaissance operation cycle (e.g., from target discovery to initial identification and localization), the swarm initiates data collection and evaluation. The collected multi-dimensional collaboration effect data includes: the cumulative effective coverage of the mission area, calculated as the ratio of the union of all UAV scan areas to the total mission area; the target confirmation time from the first detection of the target to its type confirmation; the total energy consumption of the swarm during the operation cycle, obtained by summing the power consumption of each UAV and the time; and the average positional deviation between the actual flight trajectory of each UAV and the ideal trajectory planned in the previous cycle, i.e., the command execution deviation. Next, these effect data are correlated and compared with the mission context feature vector recorded before triggering this collaborative operation (i.e., the instant before the target is discovered). The comparison method is to calculate a difference vector: for example, subtracting the actual mission area coverage from the pre-operation estimated coverage to obtain the coverage difference value; subtracting the actual target confirmation time from the preset ideal time threshold to obtain the time difference value; and subtracting the actual energy consumption from the baseline energy consumption to obtain the energy consumption difference value. These discrepancies are normalized and weighted to generate a scalar policy evaluation signal. A positive signal indicates that the overall performance is better than expected, while a negative signal indicates that it is worse than expected. This policy evaluation signal is then used to make minor adjustments to the internal parameters of the meta-learning policy unit. The adjustment process is based on the gradient descent principle: the policy evaluation signal is used as an approximate gradient direction of the loss function, and the adjustable parameters within the unit (i.e., the strength of each connection in the weighting structure) are adjusted by a small step size along the direction of loss reduction. This step size is typically set to a fixed small value, such as 0.001. Simultaneously, the belief probability distribution is iteratively updated. The complete trajectory of each UAV's actual flight in that period is compared with the candidate trajectories with non-zero belief probabilities predicted for all neighbors in the previous period's intent inference step. The matching degree is measured by calculating the average spatial distance between the actual trajectory and each candidate trajectory at the same time point. For candidate trajectories with high matching degree (i.e., average distance less than a set threshold, such as 5 meters), their corresponding belief probability entries will receive a positive confidence increment, which is inversely proportional to the matching degree; for entries with low matching degree, a negative increment will be received. Finally, using these calculated confidence increments, a weighted update is performed on the belief probability distribution of all neighboring drone action intentions, that is, the probability value of each candidate trajectory is added to its corresponding increment, and then normalized to ensure that the sum of all probabilities is 1, thus completing the iteration of belief probability.

[0035] The working principle of this invention is as follows: First, real-time data on the flight status, environmental perception, and mission situation of each UAV are collected. Then, this data is fused and processed to calculate the mission performance characteristic value, real-time environmental risk index, and cluster collaboration requirement coefficient for each UAV. Next, based on these three key indicators, an adaptive control parameter is dynamically generated through a meta-learning policy network, including a prediction time-domain length adjustment coefficient, a multi-objective cost function weight coefficient, and a constraint relaxation priority sequence. Then, each UAV uses this parameter to perform distributed intent inference, forming a probability distribution of beliefs about the actions of other UAVs, and constructs a trajectory optimization problem including probabilistic cooperative collision avoidance constraints. The optimal cooperative trajectory control command is obtained through distributed negotiation. Finally, the command is executed to complete the reconnaissance action, and based on the cluster collaboration effect data collected after execution, the meta-learning policy network is fine-tuned online, while the probability distribution of beliefs is iteratively updated, thereby achieving autonomous, adaptive, and highly collaborative reconnaissance mission scheduling in complex dynamic environments.

[0036] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for scheduling autonomous and collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms, characterized in that: Includes the following steps: S1: Real-time acquisition of flight status data of each UAV in the cluster, environmental perception data obtained by airborne sensors, and mission situation data obtained by parsing mission instructions; S2: The flight status data, environmental perception data and mission situation data are fused and processed to calculate the mission effectiveness characteristic value, real-time environmental risk index and cluster collaboration requirement coefficient of each UAV. S3: Based on task performance feature values, environmental risk index and cluster collaboration demand coefficient, a set of adaptive control parameters are dynamically adjusted and generated through meta-learning strategy network. The adaptive control parameters include prediction time domain length adjustment coefficient, multi-objective cost function weight coefficient and constraint relaxation priority sequence. S4: Each UAV uses adaptive control parameters to perform distributed intent inference and form a belief probability distribution of the intentions of other UAVs; based on the belief probability distribution, a trajectory optimization problem containing probabilistic cooperative collision avoidance constraints is constructed, and the optimal cooperative trajectory control command of each UAV is obtained through distributed solution. S5: Execute the optimal cooperative trajectory control command to complete the cooperative reconnaissance action; collect the cluster cooperative effect data after the command execution, use the cluster cooperative effect data to fine-tune the meta-learning policy network online, and iteratively update the belief probability distribution.

2. The method for scheduling autonomous collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The calculation process for the task performance characteristic value is as follows: Extract energy state parameters and maneuverability parameters from flight status data; Extract target value parameters and task time limit parameters from the task situation data; Normalize the energy state parameters, maneuverability parameters, target value parameters, and mission time limit parameters; Based on the terrain shading factor in the environmental perception data, the weight coefficients of each parameter after normalization are dynamically adjusted. The task performance feature value is output by weighting and fusing the results based on the adjusted weighting coefficients.

3. The method for scheduling autonomous collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The calculation process for the real-time environmental risk index is as follows: Calculate the relative motion relationship between each threat source and the drone to obtain the threat proximity. By combining meteorological conditions and terrain permeability from environmental perception data, the visibility and impact range of threat sources are calculated. Based on the threat proximity and impact range, a local risk map is generated using risk overlay rules; The local risk map is integrated and normalized to output a real-time environmental risk index.

4. The method for scheduling autonomous collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The calculation process for the cluster collaboration demand coefficient is as follows: Determine the collaborative task mode based on task status data; Obtain the mission performance characteristics and real-time environmental risk index of each UAV; Based on the collaborative task model, evaluate the overall cluster performance and risk balance. Based on the overall cluster performance and risk balance, calculate the expected synergistic gains of different formation configurations; Based on the current cluster communication link quality, select the gain value corresponding to the optimal formation configuration and output the cluster coordination requirement coefficient.

5. The method for scheduling autonomous collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The process of dynamically adjusting and generating a set of adaptive control parameters through a meta-learning policy network specifically includes: The task performance feature value, environmental risk index, and cluster collaboration requirement coefficient are correlated with the stored historical collaboration decision records in multiple dimensions to form the current task context feature vector. The current task context feature vector is input into the meta-learning policy unit. Through the multi-layer attention matching and weight allocation structure inside the meta-learning policy unit, the initial prediction time domain length adjustment coefficient, multi-objective cost function weight coefficient, and constraint relaxation priority sequence are generated. Based on the execution feedback of the current cluster's real-time actions and the estimated achievement of the task objectives, the initial adaptive control parameters are incrementally optimized and corrected online, and the final adaptive control parameters used for trajectory optimization are output.

6. The method for scheduling autonomous collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The probability distribution for forming beliefs about the intentions of other drones specifically includes: Each UAV generates a set of candidate trajectory segments for multiple future moments and the corresponding intent priority of the UAV based on its own adaptive control parameters and real-time status. Receive limited historical trajectory fragments of nearby drones and some publicly released intent priorities of drones through the communication link; Based on the received fragment and priority information, the belief probability of all neighboring UAVs taking each candidate trajectory within the planning time domain is calculated and updated through a Bayesian inference process that associates state transition with intent.

7. The method for scheduling autonomous collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The process of obtaining the optimal cooperative trajectory control commands for each UAV through distributed solution specifically includes: Each UAV constructs a local optimization problem based on its own generated set of candidate trajectory segments and belief probability distribution. The main optimization objective is to minimize its own trajectory cost, and the core constraint is to keep the probability of conflict with the trajectory of neighboring UAVs below a set threshold. The system exchanges current preferred trajectories and corresponding constraint satisfaction with neighboring drones through multiple rounds of communication, and dynamically adjusts its own candidate trajectory selection based on the exchanged information. When the conflict probabilities among the preferred trajectories reported by all drones meet the constraints, it is determined that a distributed consensus has been reached, and the control command corresponding to the trajectory selected by the drone at this time is taken as the optimal cooperative trajectory control command.

8. The method for scheduling autonomous collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The completion of the coordinated reconnaissance operation specifically includes: Each UAV executes the optimal cooperative trajectory control command and uses its onboard sensors to scan and collect information about the designated area in real time during flight. When a drone detects target features that meet preset conditions during scanning, it immediately triggers local action adjustments. Based on the target's geographical features and task priority, it calculates and updates the drone's own task performance feature value and environmental risk index in real time. By broadcasting updated information to nearby drones via communication links, the drones are guided to dynamically adjust their reconnaissance paths and sensor orientations based on the updated mission effectiveness characteristics and environmental risk index, thereby forming a collaborative reconnaissance operation among multiple drones that dynamically converges around the target characteristics.

9. The method for scheduling autonomous collaborative reconnaissance missions of unmanned aerial vehicle (UAV) swarms according to claim 1, characterized in that, The process of using cluster collaboration effect data to fine-tune the meta-learning policy network online and iteratively updating the belief probability distribution specifically includes: After the collaborative reconnaissance operation is completed, collect multi-dimensional collaborative effect data, including mission area coverage, target confirmation time, overall cluster energy consumption, and command execution deviation. By correlating and comparing the multidimensional collaborative effect data with the task context feature vector before the current action adjustment, the strategy evaluation signal is calculated. Using the policy evaluation signal, the internal weight allocation structure of the meta-learning policy unit is slightly modified through the gradient backpropagation process. The actual trajectories executed by each UAV are matched with the trajectories predicted in the previous belief probability distribution. The confidence increment is calculated for the corresponding belief probability distribution entries based on the matching degree. The belief probability distributions of the action intentions of all neighboring UAVs are then updated using a weighted iterative method.

Images