Method and system for sealing and controlling unmanned aerial vehicle (UAV) cluster area in dynamic game scene

Abstract

The invention discloses an unmanned aerial vehicle cluster area sealing control method and system in a dynamic game scene. The method comprises the following steps: in a region exploration stage, realizing coverage and detection search of a task region based on a collaborative search algorithm of multi-agent partition perception and path planning; in the collaborative interception stage, based on the established target motion prediction model and the self-adaptive allocation mechanism of the enemy target, dynamic allocation is realized, the optimal interception path of the unmanned ship is planned, and meanwhile, energy charging and secondary flying of the unmanned ship are carried out in time; in the confrontation game stage, based on a decision tree model, in combination with multi-dimensional battlefield evaluation data, decision execution and evaluation are carried out, attack, avoidance or other tactical instructions are generated, and an optimal instruction and global target distribution are waited and cyclically made; and after an attack instruction is sent out, performing a safe predefined tactical attack in combination with tactical action execution logic. According to the method, the problems of poor unmanned cluster cooperative perception, difficulty in resource scheduling and untimely decision response in a complex dynamic environment are effectively solved, and the interception efficiency and comprehensive combat capability of regional sealing and control are improved.

Description

Technical Field

[0001] This invention belongs to the field of multi-agent collaborative control and artificial intelligence decision-making technology, specifically relating to a method and system for regional control of unmanned aerial vehicle (UAV) swarms in a dynamic game scenario. Background Technology

[0002] With the rapid development of microelectromechanical systems (MEMS), communication technology, and artificial intelligence, unmanned systems have gradually evolved from early single remote-controlled devices into intelligent swarm systems with perception and collaboration capabilities. Among these, the common air-sea swarms composed of drones and unmanned surface vessels (USVs) are widely used in key areas such as border control and regional containment due to their small size, maneuverability, and perception capabilities. Through information sharing and coordination among intelligent agents, unmanned swarms can overcome the limitations of field of view and range for single-task execution, addressing complex challenges arising from dynamic games and adversarial situations.

[0003] Area control is a common application scenario for unmanned swarm collaborative operations, with specific tasks including target search, tracking, and interception. In maritime control scenarios, defending vessels often need to deal with strategically penetrating intruders under relevant resource constraints. Intruding targets are usually relatively unstable (relative to the defenders), and their movement (such as evasive maneuvers) changes with the battlefield situation, requiring the defenders to have precise perception and decision-making capabilities.

[0004] Cooperative tracking is the theoretical foundation for solving regional control problems, mainly exploring how multiple agents can minimize capture time or maximize efficiency through cooperative strategies. The literature (see Qu, Xingru, et al. An Overview of Recent Advances in Pursuit–Evasion Games with Unmanned Surface Vehicles[J]. Journal of Marine Science and Engineering, 2025, 13(3): 458.) summarizes cooperative tracking game theory for unmanned surface vessels, pointing out that in complex marine environments, analytical methods based on the HJI equation and data-driven methods based on multi-agent reinforcement learning (MARL) are current research hotspots. These methods can provide theoretically optimal cooperative strategies and perform well in handling obstacle avoidance and kinematic constraints. However, existing analytical methods are complex to model when dealing with complex clusters; reinforcement learning methods often require pre-setting omnidirectional perception conditions for agents, ignoring unavoidable limitations in practical scenarios such as the actual sensor field of view.

[0005] Resource allocation is a key technology for ensuring sustained unmanned swarm operations, aiming to guarantee mission efficiency under limited resources (such as power or ammunition). The literature (see Doostmohammadian, Mohammadreza, et al. Survey of distributed algorithms for resource allocation over multi-agent systems[J]. Annual Reviews in Control, 2025, 59: 100983.) systematically reviews distributed resource allocation algorithms based on game theory and consensus control, mainly discussing the algorithm's adaptability under conditions of limited communication topology and multiple complex agents. These algorithms, through iterative optimization, converge the system to Nash equilibrium, solving the distributed task and resource matching problem. However, in long-term regional containment missions, existing allocation methods often focus on maximizing short-term gains, lacking a global understanding and analysis of the scenario, which can easily lead to poor long-term collaborative effects.

[0006] Leader-follower games and hierarchical decision models are also widely used to solve multi-agent decision problems. The literature (see Zhang, Bin, et al. Stackelberg decision transformer for asynchronous action coordination in multi-agent systems[J]. arXiv preprintarXiv:2305.07856, 2023.) proposes a decision transformer framework based on the Stackelberg game concept, embedding the game structure into sequence modeling. This method utilizes the Transformer architecture model to capture long-term dependencies and improve the efficiency of cooperation among agents. However, such deep neural network-based models exhibit significant "black box" characteristics, lacking interpretability for the decisions necessary in military and security mission scenarios. Furthermore, these models heavily rely on massive amounts of simulated adversarial data for training, making it difficult to guarantee a balanced dataset and sufficient uncontaminated samples. Consequently, their generalization ability and reliability are weak in small sample sizes or specific scenarios.

[0007] In summary, existing unmanned swarm cooperative control technologies still face technical challenges in addressing stable tracking, energy consumption balancing, and game-theoretic decision-making in small-sample scenarios under certain practical constraints. Therefore, there is an urgent need for a control-based game-theoretic decision-making method that integrates globally optimized search strategies, conditionally constrained search and tracking, and real-time scenario-based decision control. Summary of the Invention

[0008] Purpose of the invention: The purpose of this invention is to provide a method and system for regional control of unmanned aerial vehicle (UAV) swarms in dynamic game scenarios, which can effectively solve the technical problems existing in the current technology, such as uneven energy consumption load leading to poor long-term control effect, sensor field of view constraints leading to tracking difficulties, and the lack of interpretability and generalization of traditional black box models.

[0009] Technical solution: The present invention provides a method for regional control of unmanned aerial vehicle (UAV) swarms in a dynamic game scenario, comprising the following steps:

[0010] A basic task area model is established, and based on an intelligent swarm search optimization algorithm that combines perception and path coordination, the swarm path variance is minimized, and the search paths of each individual UAV swarm are planned to achieve comprehensive search of the task area.

[0011] Based on UAV detection data, the next state position of the black target is predicted using a Brownian motion-driven model and a Kalman filter.

[0012] An adaptive target allocation mechanism is used to allocate targets in different scenarios; at the same time, the interception path of the white unmanned surface vessel is planned based on a geometric constraint model.

[0013] The UAV is controlled to perform a second launch based on a rectangular coverage detection model; after approaching the target, it switches to a cycloidal escort motion model to maintain continuous monitoring of the target.

[0014] A data assessment model for battlefield situation is constructed to calculate the advantage index of angle, speed, and distance in real time. Based on this, a decision tree model is constructed to map features into tactical decision instructions for attack or evasion. Finally, based on the tactical action execution logic, tactical attacks or other actions are carried out.

[0015] Furthermore, the specific steps for establishing the task area model and path planning are as follows:

[0016] Modeling the adversarial scenario: Define the task space to include the starting area and intermediate combat area of ​​the defender and attacker respectively, and set a defense line to separate the combat task areas; the tactical intention of the attacker is to break through the defense line. If the target successfully crosses the red line, it is judged as a failure of the blockade and is penalized.

[0017] Collaborative Search and Path Optimization: The objective function for minimizing the variance of the search path of the UAV swarm is expressed as:

[0018]

[0019] in, This represents the variance of the flight distance of all drones. For the number of drones, For the first The corresponding distance for deploying a drone The objective function is the average distance traveled by all drones. The equal-perimeter segmented search path is based on this objective function, which divides the task area equally according to the number of drones, so that each drone has a smaller overlapping area and the perimeter of the search path is approximately the same.

[0020] Furthermore, the state equation for predicting the next state position of the black target using the Brownian motion-driven model and Kalman filter is expressed as follows:

[0021]

[0022]

[0023] in, for The position vector at time , for The velocity vector at time t, For time intervals, To follow a Gaussian distribution The system uses a Kalman filter to correct the predicted state based on the current observations and calculates the Y-direction coordinate with the highest probability of black appearing for subsequent target tracking and allocation.

[0024] Furthermore, the specific logic of the adaptive target allocation mechanism includes:

[0025] When the number of black targets is less than that of white unmanned surface vessels (USVs), black targets farther from the center of the mission area are prioritized for allocation. When the number of black targets is equal to that of white USVs, a simple matching method based on descending Y-coordinates is used. When the number of black targets is greater than that of white USVs, the K-means clustering algorithm is used, with the predicted Y-axis information of the black targets as features, to divide them into clusters equal in number to the number of white USVs, and each white USV is assigned to the center of its corresponding cluster.

[0026] Furthermore, the geometrically optimal constraint planning interception route is based on speed obstacles and the Apollonius circle principle in mathematics to determine the interception target point. It satisfies the following generalized interception equation:

[0027]

[0028] in, The initial coordinates of the white unmanned surface vessel. Here are the coordinates of the black target. For the White team's interception speed, The escape speed of the black player. , The preset safe attack distance; when the speeds of both sides are equal and the safe distance is ignored, the equation simplifies to the equation of a linear perpendicular bisector, which is:

[0029]

[0030] Furthermore, the secondary launch and escort maneuvers performed by the drone specifically include:

[0031] Rectangular Coverage Detection: Set a detection activation threshold. When the distance between the white unmanned surface vessel and the interception point is less than the threshold, control the unmanned surface vessel to execute a rectangular search strategy. The specific path planning is as follows: first, fly along the axis perpendicular to the predicted heading of the black side, then fly along the axis parallel to the predicted heading, and finally execute a reverse closed flight.

[0032] Helical flight tracking motion based on field-of-view constraints: setting the approximation threshold as... When the drone is at a distance from the black target When switching to the field-of-view constrained escort mode, the drone must simultaneously meet the conditions of maintaining relative distance and sensor field-of-view constraints. (in In the direction of velocity, Direction of sight (where the drone's field of view is the angle of view); under this constraint, the drone's flight path... For the variable curvature spiral approximation model:

[0033]

[0034] in The target position for Black; Let be the radius of rotation, when the field of view angle At that time, the trajectory is a spiral curve that gradually converges towards the target until it converges or reaches a certain distance to execute other optional accompaniment logic.

[0035] Furthermore, the basic kinematic dominance index in the battlefield situation assessment model is calculated as follows, including:

[0036] Angle Advantage Index The calculation formula is:

[0037]

[0038] in The difference in the angle between the velocity vectors of both sides and the line connecting the target;

[0039] Speed ​​Advantage Index Set speed ratio range The calculation formula is:

[0040]

[0041] Distance Advantage Index Set the optimal strike zone The calculation formula is:

[0042]

[0043] in This is the distance attenuation coefficient.

[0044] Furthermore, the indicators in the battlefield situation assessment model are calculated as follows:

[0045] Our health :

[0046]

[0047] in The remaining combat time of our drones, in actual and simulated scenarios, has a linear or non-linear relationship with the drone's electrical energy. The number of times the attack has been performed is locked. Maximum number of attack lock-on capabilities;

[0048] Lock ability level Overall weighted advantage value The calculation formula is:

[0049]

[0050] in It is a soft normalization function. To normalize the remaining time factor, To normalize the enemy-friend distance factor;

[0051] Threat Level Index :

[0052]

[0053] in To detect the number of enemies within the field of view, The ratio of the average distance between the nearest and second-nearest enemy groups. This is the difference factor for the number of times both sides lock onto the target during an attack;

[0054] Drone Status Index :

[0055]

[0056] in These are the normalized drone flight time, distance to the returning mothership, and distance to the target, respectively. These are the corresponding weighting coefficients.

[0057] Furthermore, the decision tree model and tactical execution logic include:

[0058] Decision tree training: A classification and regression tree algorithm is used to generate feature vectors based on a defined standard tactical scenario. The model is trained using the training sample set with decision labels;

[0059] Dynamic tactical execution: This is reflected in the tactical execution phase of the simulation system, which continuously evaluates decisions and judges and changes the action logic based on the decision labels obtained at different times, and ultimately converges to safe offensive or safe evasive actions.

[0060] The present invention discloses a dynamic game-driven cross-domain unmanned aerial vehicle (UAV) swarm area control system, comprising:

[0061] Simulation Interactive Environment Module: Used to construct an environment with physical and rule constraints. This module contains definable task parameters and loads configurations including attack lock-on conditions, agent dynamics constraints, and security boundaries; this module calculates the motion state of UAVs and unmanned surface vessels in real time, handles collision detection logic, and outputs environmental state observation information;

[0062] Heterogeneous Intelligent Agent Modeling Module: Used to implement UAV and unmanned surface vessel objects. This module includes:

[0063] The dynamics unit calculates the pose of the agent in the next simulation step based on preset velocity and acceleration limits;

[0064] The sensor unit simulates the detection range and noise characteristics based on the detection radius and field of view parameters;

[0065] The resource management unit monitors and updates the battery level and remaining attack lock count of each agent in real time based on the preset battery life and payload quantity.

[0066] Algorithm Decision Control Module: As the control core of the system, it interacts with the simulation environment through a communication interface and includes: a collaborative planning unit that runs coverage search and geometric interception algorithms to generate navigation waypoints;

[0067] The situation assessment unit performs Brownian motion prediction and situation data calculation.

[0068] The game reasoning unit loads the trained decision tree model and outputs tactical instructions to each agent based on the real-time situation.

[0069] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are as follows: (1) Based on the search strategy of minimizing the path variance, the energy consumption balance of the cluster in the regional exploration task is realized, avoiding the problem of individual energy being exhausted prematurely in long-term tasks, which leads to the offline of the agent and the lack of regional control, thereby improving the continuous combat capability of the cluster, the stable state of each agent and the complete coverage of the task area; (2) Combining the prediction model driven by Brownian motion with an adaptive target allocation mechanism, the success rate of predicting the correct position is increased, and the low task efficiency caused by arbitrary allocation strategy is avoided, thereby improving the interception success rate; (3) A variable curvature spiral escort tracking motion model based on field of view constraint is proposed, which solves the problem of easy target loss during tracking when using narrow field of view sensors, and realizes continuous monitoring and tracking of maneuvering targets; (4) A white-box decision tree game model based on situation data is constructed, which uses situation assessment indicators to replace the traditional black-box deep learning model, has strong interpretability and transparency, and overcomes the problem of small sample data being difficult to converge in military simulation scenarios, thereby enhancing the real-time performance and robustness of decision-making. In summary, this invention achieves a complete mission action process through path variance optimization search planning, interception allocation, constraint-based escort, and related game theory decision-making. This invention effectively solves the problems of poor collaborative perception, resource scheduling, and delayed decision-making response in unmanned swarms under complex dynamic environments, improving the mission execution efficiency of regional control tasks and possessing broad application and practical value. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the overall process of the UAV / Ship swarm area containment method and system in the dynamic game scenario of the present invention;

[0071] Figure 2 This is a schematic diagram of task area modeling and one type of equal-perimeter cooperative search path planning for UAVs in an embodiment of the present invention;

[0072] Figure 3 This is a schematic diagram of the target allocation logic of the white unmanned surface vessel using an adaptive mechanism in an embodiment of the present invention;

[0073] Figure 4 This is a schematic diagram illustrating the principle of interception target point determination based on generalized geometric constraints in an embodiment of the present invention;

[0074] Figure 5 These are schematic diagrams of two types of UAV spiral flight paths based on field-of-view constraints in embodiments of the present invention;

[0075] Figure 6 This is a schematic diagram of the rectangular coverage detection path for the secondary launch of the UAV in an embodiment of the present invention;

[0076] Figure 7This is a statistical table of the containment and countermeasure results of the method in the embodiment of the present invention under a simulation test environment. Detailed Implementation

[0077] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following will provide a more detailed description of these aspects in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments described are for illustrative purposes only and are not intended to limit the scope of the invention.

[0078] Example: Figure 1 As shown in the embodiments of this invention, the method and system for controlling a cluster of unmanned aerial vehicles (UAVs) in a dynamic game scenario first establishes an adversarial model of the task area and uses a swarm search algorithm based on perception and path coordination to plan segmented paths and achieve coverage search. Then, based on data detected by the UAVs, the motion state of the black target is predicted using a Brownian motion model and a Kalman filter. Next, an adaptive target allocation mechanism is used, and a geometric model is used to plan the UAV interception path. During the approach process, the UAVs are controlled to perform rectangular detection and spiral flight under constraints to maintain continuous monitoring. Finally, a data evaluation model of the battlefield situation is modeled, and offensive and defensive tactical commands are generated using a decision tree. Finally, relevant simulations demonstrate the improved interception efficiency and robustness of the method of this invention.

[0079] Specifically, the method mainly includes the following steps:

[0080] Step 1: Establish a task area model and employ an intelligent swarm search optimization algorithm based on partition awareness and path coordination. With minimizing the swarm path variance as the optimization objective, plan the equal-perimeter segmented search path for the UAV swarm. Specifically:

[0081] Constructing an offensive and defensive scenario model: First, a digital airspace and sea area model for task execution is established using a simulation interactive environment. The task space is defined as a polygonal region enclosed by a sequence of vertices. For example... Figure 2 As shown, in this embodiment, the mission boundary is formed by eight vertices (A1-A8) in sequence, forming a narrow control zone that is about 500km long from north to south and about 300km wide from east to west.

[0082] In terms of tactical situational division, the system vertically divides the mission area into the defensive deployment zone, the central engagement zone, and the offensive invasion zone. Specifically, the defensive unmanned swarm is initially deployed near the A7-A8 line on the left side of the area, serving as a refueling and launch site; while the potential attacking force's range is defined as the trapezoidal area on the right side enclosed by A2-A3-A4-A5. To quantify the containment effectiveness, the system sets a key defensive red line at the outermost edge of the defensive side (i.e., the A1-A6 line). The Black (attacking) force's tactical intention is to cross the central sea area and break through this red line, while the White (defending) force's core mission is to establish a multi-layered interception defense line based on this red line.

[0083] This modeling based on the "deployment area-combat zone-defense red line" is well-suited to actual needs such as maritime anti-smuggling, border control, and maritime blockade. By setting key defense red lines and in-depth combat zones, it simulates a typical "defense in depth" situation: the defender cannot rely solely on terminal interception, but must make full use of the space in the middle area to buy time for forward reconnaissance and attack targeting.

[0084] Construct a collaborative coverage path planning model based on minimizing path variance:

[0085] First, the system establishes a top-level optimization objective centered on minimizing cluster path variance. Regardless of the specific overlay pattern used, the algorithm must satisfy the following constraints:

[0086]

[0087] in, For the number of drones, For the first The planned total distance for the drone. The average distance traveled is denoted by . The physical meaning of this objective function is to constrain the energy consumption of each node to be consistent, preventing the persistence of the containment network from being compromised due to excessively long travel distances at some nodes.

[0088] To obtain an accurate solution Furthermore, this invention performs a mathematical decomposition of the single-aircraft range. Defined as the total cost of starting from the deployment area, executing the coverage task, and returning:

[0089]

[0090] in This indicates that the drone travels between the initial deployment point and the mission sub-area. The distance between them for maneuver; Indicates in sub-region The effective probe distance required to perform a full-coverage search within the area.

[0091] If we choose to cover the triangle with a simple shape, we can solve for the trajectory paths of the five UAVs in this example, such as... Figure 2 As shown in (a), however, for specific task scenarios, this can lead to large areas of empty coverage, resulting in missed detections (such as...). Figure 2 (a) For area control tasks with high coverage requirements, this invention preferably employs the Boustrophedon reciprocating scanning strategy. In this mode, the coverage path length is highly correlated with the sub-region area, and its calculation model is extended as follows:

[0092]

[0093] To allocate to the first The area of ​​the sub-region of the drone; This refers to the effective detection swath of the UAV sensor; To account for the correction factor of the overlap rate, it is generally taken as 1; The additional maneuvering cost incurred by turning back and forth.

[0094] Based on the above formula, the algorithm transforms the complex path planning problem into a dynamic region partitioning problem. The model is iteratively adjusted and modified. The size, thereby dynamically adjusting Ultimately offset by The difference in total distance caused by (different distances) makes it possible to satisfy The convergence conditions are determined, and the initial search paths for each UAV are obtained.

[0095] Step 2: Introduce a prediction model combining a sliding window mechanism for location prediction:

[0096] In this embodiment, the system does not use fixed kinematic parameters for prediction. Instead, it adjusts the process noise parameters of the Brownian motion model by analyzing the statistical characteristics of the target's motion during the observation period. The specific implementation process is as follows:

[0097] When the drone continued to detect the black target At that time, the system records the historical heading angle sequence within the observation time window. Quantify the severity of the target's maneuvering and calculate the sample variance of the heading angle. :

[0098]

[0099] in, This represents the average heading angle within the observation window. The higher the value, the more aggressively the target is maneuvering to avoid collisions. A value close to 0 indicates that the target can penetrate the defenses in a relatively straight line.

[0100] Based on the above statistical characteristics, a discrete-time state-space model is established. The core of this model lies in utilizing... The noise terms in the basic equations are processed.

[0101] Velocity update equation:

[0102]

[0103] in, The noise is a process noise that follows a Gaussian distribution. To accurately reflect the target's maneuver uncertainty, the system uses the following adaptive formula for calculation. :

[0104]

[0105] in, As the reference noise matrix, This is the motor amplification factor. When... When the value is small (stable). The predicted trajectory converges relatively quickly, forming a slender ellipse; when When it is larger (maneuverable). Significantly increased, making the formula in This generates greater random diffusion, and the prediction results will cover a wider potential area.

[0106] Position update equation:

[0107]

[0108] This formula uses the updated velocity vector to deduce the position. The initial value... The direction is determined according to the following logic: if (High mobility), then forced to The direction was corrected to the strategic main course pointing towards the defense red line to prevent the prediction range from being greatly exaggerated due to high-frequency maneuvers. Therefore, the prediction was processed directly according to the straight-line flight information.

[0109] Based on the above model, during the blind spot period after the drone loses contact (i.e. The system lacks real-time observation and correction data. The system will perform blind zone recursive prediction using Kalman filtering. It uses the posterior estimate corrected from the last observation as initial conditions and iteratively executes the time update step.

[0110] Recursive prediction: At each simulation step, the system performs iterative calculations based on the following state extrapolation equation and covariance extrapolation equation:

[0111]

[0112] in:

[0113] for The predicted prior state value at time t. This is the state transition matrix; this equation reflects the target's movement along the velocity vector. Inertia indicating direction.

[0114] for The prior error covariance matrix at time t; this formula reflects the diffusion of positional uncertainty. Wherein For the preceding steps based on heading variance The calculated adaptive process noise matrix.

[0115] The above process is iterated repeatedly until the interception phase is reached. The system outputs a predicted state vector. Finally, the system extracts the vector from the vector. Axis components (i.e., longitudinal coordinates) The location where the interception line is most likely to be breached will be used for subsequent dynamic task allocation.

[0116] Step 3: An adaptive target allocation mechanism based on quantity comparison is used to allocate targets in different scenarios; the interception path of the white unmanned surface vessel is planned based on a geometrically optimal constraint model, specifically as follows:

[0117] This step aims to solve the target allocation and interception path determination problems in multi-vessel cooperative blockade. The system uses the black side's predicted state from the previous steps. and quantity , combined Figure 3 The logical framework shown first performs an adaptive target allocation based on quantity comparison:

[0118] like Figure 3 As shown, the system first calculates the number of black detections. Number of unmanned surface vessels available to the White side The ratio is dynamically switched and allocated based on the three typical situations of asymmetric confrontation:

[0119] a) when At this time, the White side has relatively abundant resources. In addition to ensuring that each side's target is locked by at least one White side vessel, the remaining White side resources will be prioritized for allocation to the most threatening targets (usually the targets furthest from the centerline that are trying to break through from the flanks).

[0120] Therefore, a threat / priority weighting function is defined. :

[0121]

[0122] in, For the first The predicted ordinate of each black target. As the center line of the region, Its penetration speed. Weight The larger the value, the more interception resources are allocated.

[0123] b) When At this point, the Y-axis sorting matching strategy is executed, sorting the black and white sets respectively according to their Y-axis coordinate values ​​from largest to smallest:

[0124]

[0125] Let the white side A boat went to intercept the black team's first One goal (i.e.) This mapping minimizes spatial intersections of interception paths, reduces collision risks, and ensures that each unmanned surface vessel is assigned the nearest enemy unmanned surface vessel.

[0126] c) When At that time, K-means clustering was used for allocation. The system implemented partitioned blocking for the black square cluster.

[0127] The algorithm uses the predicted ordinates of all black targets. As the input sample, the number of unmanned surface vessels of the White side is used. Number of cluster centers Construct the following clustering optimization objective function:

[0128]

[0129] in, For the first Cluster set, Let be the geometric center of this cluster. After iterative convergence, the th... A white unmanned surface vessel will be assigned to intercept the first The central location of each cluster This strategy ensures that even when there is a mismatch in troop strength, the White side can still intercept enemy unmanned vessels attempting to break out within the sub-area.

[0130] After identifying the target to be intercepted, the system needs to calculate the optimal interception point. This invention implements interception path planning based on geometric constraints, proposing an interception model based on speed barriers and the Apollonius circle principle, wherein:

[0131] The generalized interception equation is derived as follows:

[0132] Let the initial position of the white unmanned surface vessel be... Interception speed is The current position of the black target is The speed of the rush is .

[0133] The condition for a successful interception is set as follows: the white team reaches the target point. Time Equals the time it takes for Black to reach that point. Furthermore, considering the weapon range or safety distance margin in actual combat, a safe attack distance needs to be introduced. (That is, the white side must reach the target distance) Only then can it be considered a safe interception.

[0134] Wherein: the interception relationship equation is established as:

[0135]

[0136] Let the speed ratio Squaring both sides of the above equation and rearranging, we obtain the coordinates of the interception point. The generalized interception circle equation:

[0137]

[0138] After expanding and combining like terms, we get the standard form:

[0139]

[0140] This equation describes an Apollonius circle (when...) The intersection of this circle and the predicted motion ray of the black square (based on the course predicted in step 2) is the interception point with the least time consumption. .

[0141] The special case is simplified to: same speed / opposite direction interception ( )

[0142] In actual naval warfare, when the speeds of both sides are close ( And when in a head-on interception situation, the coefficient of the above quadratic term When the value becomes 0, the equation degenerates into a linear equation.

[0143] If we further ignore the safety distance (assuming...) Then the equation simplifies to the perpendicular bisector form in the claims:

[0144]

[0145] At this point, the optimal interception point is the intersection of the black side's flight path and the perpendicular bisector of the line connecting black and white. Figure 4 The interception mode exhibited.

[0146] The actual execution logic involves the system solving the above equations in real time. If the calculated... Located within the mission area boundary, the Belarusian unmanned surface vessel will use this point as its navigation target and advance in a straight line, taking the geometrically shortest time path; if If the boundary is exceeded, the intersection point of the boundary is taken as the suboptimal interception point.

[0147] Step 4: Control the UAV to perform a second launch based on the rectangular coverage model; after approaching the target, switch to a spiral flight tracking motion model based on field-of-view constraints to maintain monitoring, specifically:

[0148] The system re-establishes situational awareness by having the UAV take off a second time and uses a flight escort strategy to maintain continuous monitoring of the target in order to solve the problem of position loss caused by target maneuvering at the end of the interception.

[0149] The rectangular coverage detection model is generated based on the prediction error elliptical coverage, including:

[0150] At the interception endpoint, due to the random maneuvering of the black target within the blind zone, its true position follows a two-dimensional Gaussian distribution that diffuses over time. To maximize the interception probability of a second launch, an adaptive rectangular coverage based on the error covariance matrix is ​​constructed.

[0151] The system monitors the location of the Belarusian unmanned surface vessel in real time. With preset interception guidance point The Euclidean distance is such that a second launch is triggered when the following conditions are met:

[0152]

[0153] in This is the minimum distance threshold required for a second launch.

[0154] rectangular scanning area of ​​the drone This is to cover the confidence interval of the target location. The system uses the final prediction covariance matrix output by the Kalman filter. Calculate the major and minor axes of the prediction error ellipse. (Geometric parameters of the scan rectangle are also considered.) Determined by the following formula:

[0155]

[0156] in: These represent the standard deviations of the predicted location in the horizontal and vertical directions, respectively. This is the confidence coefficient (usually taken as 3 to cover a 99.7% probability interval). This refers to the expansion of the boundary resulting from the target's maximum maneuverability during the blind zone period. For example... Figure 6 It is a specific matrix detection mode that has been constructed.

[0157] This formula ensures that the drone can cover the largest possible area of ​​target presence with minimal range loss.

[0158] When the drone is launched a second time and approaches the target to within the distance threshold, the system switches to escort mode. This is for fixed-wing drones with speeds significantly greater than the target's. And the sensor's field of view ( Given the limited physical conditions, a spiral motion model was constructed.

[0159] Wherein: the equation of state for relative motion is:

[0160] Establish with the black side as the target A moving coordinate system with the origin at which the UAV is located; the position of the UAV in this coordinate system. Defined by the following polar coordinate parametric equations:

[0161]

[0162] in:

[0163] This represents the real-time displacement of the Black target. The instantaneous flight angular velocity of the UAV is dynamically adjusted as the radius decreases to maintain a constant tangential velocity; The initial phase angle when entering the escort trajectory is determined by the geometric connection between the entry point and the target.

[0164] Traditional circular escort flight requires the drone's velocity vector Always perpendicular to the line-of-sight vector (i.e., tangential flight), such as Figure 5 As shown in (a). However, this is limited by the limited field of view of the airborne sensors. (For example The drone must meet the following field-of-view maintenance constraints:

[0165]

[0166] in The velocity direction angle, This is the angle of the line of sight.

[0167] when At that time, the drone cannot maintain pure tangential flight. A radial velocity component pointing towards the target must be generated. To correct course deviation, such as Figure 5 As shown in (b). To ensure that the UAV always locks onto the target at the optimal observation angle (i.e., the center of the field of view) during the approach, this radial component results in a gyration radius. It no longer remains constant, but instead obeys the following differential relation:

[0168]

[0169] in For approximation ratio, which is also the radial velocity component, and Geometric constraints must be satisfied To ensure the front angle Constant. Integrating the above equation again, we obtain the time evolution law of the Archimedes spiral trajectory:

[0170]

[0171] This formula indicates the relative distance between the drone and the black target. It decays linearly over time. Based on this, in order to maintain the cruise speed of the fixed-wing UAV without stalling... tangential velocity component With radial velocity component The law of conservation of energy must be satisfied. From this, the real-time angular velocity of the drone's escort flight can be derived. The dynamic adjustment formula:

[0172]

[0173] This formula reveals the dynamic characteristics of the system: as the radius of gyration increases... The linear decrease in kinetic energy, in order to consume a constant kinetic energy, reduces the drone's angular velocity during flight. It increases dramatically in a hyperbolic manner.

[0174] (Note: The velocity solution in the actual coordinate system requires vector synthesis to follow the parallel velocity of the enemy unmanned surface vessel.)

[0175] The termination condition for the escort flight is set to reach the minimum safe hovering radius of the drone. Then from the initial cutting distance Physical time required to converge to the limiting equilibrium state for:

[0176]

[0177] During tactical execution, the system calculates the remaining convergence time in real time. When this happens, it means that the drone is about to converge to the target point and needs to adopt other flight strategies or reset the tracking. For example, in our case, we used the method of flying over the target point a certain distance and then turning around to track the position in a straight line until the unmanned surface vessel makes a reasonable attack decision command and is about to launch an attack lock or the drone's energy consumption forces it to return.

[0178] Step 5: Construct a battlefield situation data assessment model and calculate relevant indices in real time; based on the decision tree model, map features to tactical decision commands; finally, based on the tactical action execution logic, conduct tactical attacks on targets or execute other actions, specifically as follows:

[0179] First, the basic geometric relationship between the white unmanned surface vessel and the current candidate black target is calculated and quantified into a normalized advantage score.

[0180] Among them, the angle advantage index for:

[0181]

[0182] This represents the angle deviation between the white player's velocity vector and the vector connecting both sides. The angle deviation is then expressed as... Mapped to And cut into .when hour, This indicates that the White side is in a better forward interception position; when the angle is too large, the advantage decreases, indicating the actual directional physical requirements of the weapon launch or lock-on attack.

[0183] Among them, the speed advantage index for:

[0184]

[0185] These are the velocity moduli of our side and the enemy side, respectively. The parameter 'a' is set to consider the continuity of the function. In the example, we set and This is the critical threshold.

[0186] Among them, the distance advantage index for:

[0187]

[0188] Constructed from a variant of the Sigmoid function. [30, 40] km: Set as the optimal strike range, with a score of 1. This corresponds to the optimal effective range of airborne / shipborne weapon systems; < 30 km: Forced to zero, reflecting the high risk of entering the enemy's close-in weapon system range, etc.; > 40 km: S-shaped decay, reflecting the non-linear decrease in hit probability or lock-on success rate as distance increases.

[0189] The system integrates the above indicators with resource status to construct a four-dimensional core feature vector. :

[0190] Feature A: The health status of the white-clad person is:

[0191]

[0192] (Remaining battery life) and (Attack lock count consumed) are two factors that limit combat capability. If either one is exhausted, overall health will decline. Reset to zero immediately. For total battery life, This represents the total number of attacks.

[0193] Feature B: Lock-on capability level is:

[0194]

[0195] The first part is a weighted sum of the basic kinematic advantages and a soft normalization function. Smoothing;

[0196] The weighted coefficients represent the contribution weights of each advantage factor to the lock-on success rate. In this embodiment, the weight combination system is obtained by conducting Monte Carlo local adversarial experiments within a roughly defined range of values ​​to maximize the lock-on rate. The latter part introduces... Item, of which .in The remaining endurance of the unmanned surface vessel. The minimum attack lock duration required in practice. Let be the time normalization constant. This indicates that the remaining time is insufficient to complete a full lock, characteristic Set to zero; , The current closest distance between enemy and friendly forces. To predict the average approximation velocity, This is the baseline interception time constant.

[0197] Feature C: Threat Level Index is:

[0198]

[0199] This indicator considers more than just the total number of enemies within the area's field of view. It also introduced the distance ratio. :

[0200] in This represents the average distance between the nearest Black unmanned surface vessel and each of our unmanned surface vessels. This represents the average distance to the nearest Black unmanned surface vessel. A smaller ratio indicates a lower level of pressure and a smaller threat. The correction reflects the asymmetry of the locking relationship between the two sides. If our side is locked on more times than the enemy, the threat index increases, which is used to quantify the asymmetry in the combat capabilities of both sides. In real-world scenarios, parameters also need to be estimated based on the specific mission. For discrete hierarchical functions, the above continuous calculated values ​​are mapped to... Five discrete threat levels. The mapping thresholds are also derived from the risk distribution statistics of historical simulation data. For example, a state with a posterior probability of being defeated greater than 80% is classified as level 5.

[0201] Feature D: The drone state index is:

[0202]

[0203] The indicators comprehensively evaluate the drone's battery life ( ), pullback safety ( ) and mission distance ( ), used to incorporate weights of reconnaissance resources into attack decisions.

[0204] in: This reflects the remaining flight time of the drone; This reflects whether the current distance between the drone and the mothership is within the safe radius calculated using the available remaining energy. Inside; This reflects whether the distance between the drone and the nearest target is within the optimal observation range; To correspond to the weights, the optimal values ​​should be obtained in practical application scenarios, or assigned accordingly. The value is relatively high because insufficient energy is the primary reason for the lack of reconnaissance capabilities.

[0205] Based on the above features, the system uses a decision tree algorithm to construct a decision classifier. Specifically:

[0206] During inference and execution, each simulation step-size decision tree is based on the real-time input feature vector. Output discrete instruction labels. In the example system, to simplify the action concept, two action instructions are set for the final interception phase of the agent, including:

[0207] Attack command: Triggers multi-level locking logic. The White unmanned surface vessel prioritizes locking onto the nearest Black target; if that target is already frozen, it automatically switches to locking onto the next nearest target, achieving automatic allocation and replacement.

[0208] Evade command: Triggers the evacuation logic. Calculates the average velocity vector of the local enemy group. Control the white unmanned surface vessel along Maneuver in different directions to create distance and break contact, reducing the possibility of being surrounded and locked on by the enemy.

[0209] The specific training data used a standard tactical scenario (low threat and high health). Offensive, high threat level, and limited resources (Evacuation), generate a seed set, and then expand the boundary samples through simulation and related consultation.

[0210] The training data is shown in the table below:

[0211]

[0212] The system of the present invention includes:

[0213] The simulation interaction environment module is used to realize a virtual adversarial environment with physical and rule constraints. It is responsible for providing test scenarios and can be configured with task constraints including attack lock-in judgment conditions (such as lock-in time of 300s), agent dynamics constraints (maximum speed, turning radius) and safety boundaries (such as A1-A8 coordinates). The logic calculation is used to calculate the motion state update of UAVs and unmanned surface vessels in real time, handle collision detection (such as collision judgment when the distance is less than 100m) logic, and output environmental state observation information data to the algorithm end.

[0214] The intelligent agent modeling module is used to implement UAVs and unmanned surface vessels in a simulation environment, simulating the physical characteristics of intelligent agents, and includes:

[0215] The dynamics unit presets speed and acceleration limits, calculates and restricts the pose update of the agent in the next simulation step, and simulates the minimum turning radius constraint of the fixed-wing UAV, etc.; the sensor unit simulates the effective detection range and noise characteristics of radar or optoelectronic payloads under actual conditions based on the detection radius and field of view parameters; the resource management unit is responsible for real-time monitoring and updating of the status data of each agent, including the remaining energy of the UAV, the remaining attack lock count of the unmanned surface vessel, and the current health status, etc.

[0216] The algorithm decision control module is used to implement algorithm control and interact with the simulation environment. It interacts with the simulation environment through a communication interface and specifically includes:

[0217] The collaborative planning unit runs coverage search and generalized geometric interception algorithms to calculate, generate, and set navigation waypoints for each agent; the situation assessment unit runs a Brownian motion Kalman prediction model to calculate scene situation data and output the predicted target position in real time. With feature vectors Game reasoning unit: Loads a pre-trained decision tree model, outputs tactical instructions based on real-time input situational data, and distributes them to each agent to execute specific locking or evacuation actions.

[0218] Figure 2 (a) and (b) illustrate the task area modeling and UAV perimeter-based collaborative search path planning schematic diagram and task area map in an embodiment of the present invention. The octagonal area in the diagram represents the task boundary for air and sea control, and the green dashed line represents a possible collaborative flight trajectory for five UAVs in the instantiated scenario, with the geometric perimeters of the corresponding paths tending to be consistent. Similarly, different path optimizations based on area can be set according to different scenario conditions. This planning method achieves coverage of the task area and a relatively balanced load on each agent.

[0219] Figure 3 This diagram illustrates the adaptive target allocation logic for the white unmanned surface vessel (USV) in an embodiment of the present invention. The diagram shows how the algorithm makes different decisions based on the number of black and white targets. For example, when "black > white," a clustering method is used to cluster the black targets according to their spatial location, and then the white USV is directed to intercept them along the cluster center.

[0220] Figure 4 This diagram illustrates the principle of target point determination based on a geometric constraint model in an embodiment of the present invention, simplified to the case where both sides have the same speed and no security attack distance threshold is set. This is the initial position for the white player. For Black's position, The calculated optimal interception point is determined by constructing an interception equation. Instead of simply pursuing head-on, the White unmanned surface vessel (USV) consistently navigates towards the geometrically optimal interception point.

[0221] Figure 5 (a) illustrates a schematic diagram of a UAV spiral flight trajectory based on field-of-view constraints but with relatively weak constraints, according to an embodiment of the present invention. The curve in the diagram is a non-convergent spiral (taking a wide field of view as an example). Figure 5 (b) shows the convergent flight spiral in a relative coordinate system with a smaller field of view and stronger constraints.

[0222] Figure 6This is a schematic diagram of the rectangular coverage detection path for the secondary launch of the UAV in an embodiment of the present invention. The figure shows the rectangular flight trajectory of the UAV. This scanning strategy attempts to correct for late-stage positional biases caused by missing observations, providing target coordinate updates for the model.

[0223] Figure 7 This is a statistical table showing the blocking and countermeasure results of the method in this embodiment of the invention under a simulation test environment. The table records relevant indicators such as the number of breakthroughs, the number of successful lock-ons, and the number of collisions under 15 different test scenarios. The data shows that even under conditions with a large number of enemy forces (such as serial numbers 12-15), this method can still maintain a high lock-on success rate, and the number of collisions is almost zero. These results indicate that the algorithm proposed in this invention can achieve high interception efficiency and robustness while ensuring security.

[0224] As can be seen from the simulation results and schematic diagrams above, the UAV swarm area control method and system designed in this invention for dynamic game scenarios has good effect on tracking and intercepting maneuvering targets under complex dynamic environment and asymmetric confrontation conditions. In particular, the search strategy achieves energy consumption balance, the escort model under field of view constraints solves the problem of tracking loss, and the decision tree model improves the transparency of decision-making.

[0225] This invention provides a highly efficient, flexible, and interpretable two-dimensional collaborative blockade method for air and sea operations, which demonstrates satisfactory results in scenarios such as maritime law enforcement area blockade. The above are merely preferred embodiments of this invention. It should be noted that the above embodiments do not constitute a limitation of this invention. Various changes and modifications made by those skilled in the art without departing from the technical concept of this invention all fall within the protection scope of this invention.

Claims

1. A method for regional control of unmanned aerial vehicle (UAV) swarms in a dynamic game scenario, characterized in that, Comprise the following steps: A task area model is established, an intelligent search algorithm based on partition perception and path cooperation is used, and a cluster path is planned to minimize the variance of the cluster path to achieve coverage search of the task area by each sub-boat of the cluster; Based on the data detected by the unmanned aerial vehicle, a target prediction model driven by Brown motion is established, and the target position is updated through a Kalman filter to obtain the posteriori estimation of the next state of the black side; Based on an adaptive target allocation mechanism, the maximum posteriori longitudinal coordinate is used as an input parameter for target allocation to achieve target allocation logic in different scenarios; The white side (interception side) unmanned boat interception path is planned based on a geometric constraint model; The opportunity for the unmanned aerial vehicle to perform secondary release is determined based on a rectangular coverage detection model; and a cycloid companion flight motion model is performed to maintain a continuous monitoring state after approaching the target; A data evaluation model of the battlefield situation is modeled, and an angle, speed and distance advantage index is calculated in real time; a decision tree model is constructed based on the above data structure, including health degree A, locking ability level B, threat level index C, and unmanned aerial vehicle state index D; finally, the features {A, B, C, D} are input into the trained decision tree, and the actual tactical decision instruction of attack or retreat is output; finally, the multi-enemy target is controlled and closed according to the tactical action execution logic.

2. The method of claim 1, wherein, The specific steps for establishing the task area model and path planning are: An opposing situation scenario model is constructed: the task defines the task space to include the starting area of the defense side and the attack side, the intermediate combat area, and a defense line separating the combat task areas; the goal of the attack side is to break through the defense line, and if the target successfully crosses, the task scenario is judged as a containment failure and is counted as a penalty; Collaborative search and path optimization: the optimization objective is to minimize the variance of the search path of the unmanned aerial vehicle cluster, and the objective function is represented as: wherein, represents the variance of all UAV flight paths, is the number of UAVs, is the first path corresponding to the UAV, is the average of all UAV paths; the equicircumference segmentation search path based on this objective function is based on the target function, the task area is equally divided according to the number of UAVs, so that each UAV corresponds to a less overlapping area, and the search path circumference is approximately uniform.

3. The method of claim 2, wherein, The next state position of the black target is predicted using a Brown motion driven model and a Kalman filter, and the state equation is represented as: wherein, is the position vector at time t, is the velocity vector at time t, is the velocity vector at time t, is the time interval, is the process noise subject to Gaussian distribution The model corrects the above predicted state by using Kalman filter and calculates the Y direction coordinate with the maximum posterior distribution probability of black square to perform subsequent target tracking assignment.

4. The method of claim 3, wherein, The specific logic of an adaptive target allocation mechanism includes: when the number of black targets is less than the number of white unmanned boats, a priority strategy is adopted to allocate the black targets that are farther from the center of the task area first; when the number of black targets is equal to the number of white unmanned boats, the model performs corresponding matching in descending order of Y coordinate; when the number of black targets is greater than the number of white unmanned boats, the K-means clustering method is combined, the maximum posteriori coordinate information of the Y direction of the black target is calculated as a feature, and the black unmanned boat group is divided into clusters equal to the number of white unmanned boats, and the white unmanned boats are correspondingly allocated to the cluster centers.

5. The method of claim 4, wherein, The geometric constraint planning interception route is determined based on a speed obstacle and a mathematical Apollonius circle principle to determine an interception target point, which satisfies the following generalized interception equation: satisfies the following generalized interception equation: wherein, is the initial coordinate of the white side unmanned surface vehicle, is the target coordinate of the black side, is the intercept speed of the white side, is the escape speed of the black side, is the preset safe attack distance; when the speed of the black and white sides is equal and the safe distance is ignored, the equation can be simplified as a linear vertical bisector equation, that is: 。 6. The method of claim 5, wherein, The secondary release and companion flight motion performed by the unmanned aerial vehicle specifically include: Rectangular coverage search: set a search initiation threshold When the white drone is less than the threshold from the intercept point, control the drone to implement a rectangular search strategy; the specific path planning is: first fly along the axis perpendicular to the predicted heading of the black side to intercept, then fly along the axis parallel to the predicted heading, and finally perform a closed return flight; Helix companion tracking motion based on field of view constraint: set a proximity threshold value as When the UAV is at a distance of from the black target, switch to the field of view constraint companion mode of the helix model; in this mode, the UAV needs to meet the conditions of maintaining a certain relative distance and the field of view angle constraint of the sensor , wherein is the velocity direction, is the line of sight direction, is the field of view angle of the UAV; under this constraint, the companion tracking trajectory of the UAV is a variable-curvature helix proximity model: wherein is the black square target position; is the gyration radius, when the field of view angle When the trajectory is a spiral curve that gradually converges to the target until it converges or reaches a certain distance and then executes other companion flight logic.

7. The method of claim 6, wherein, The basic kinematics index calculation in the data evaluation model of the battlefield situation is as follows, including: angle advantage index : wherein is the difference between the angle of the velocity vector of both parties and the connecting line to the target; Speed advantage index : Set speed ratio interval The calculation formula is: Distance advantage index : Setting optimal hitting interval , the calculation formula is: wherein is the distance attenuation coefficient.

8. The method of claim 7, wherein, The complex index calculation in the data evaluation model of the battlefield situation is as follows, including: Our health : wherein T is the remaining combat time of the UAV, which is linear or nonlinear with the energy of the UAV in reality or simulation, is the number of attack lockups performed, is the maximum number of attack lockup capabilities; Locking capability level : integrated weighted advantage value , the calculation formula is: wherein is a soft normalization function, is a normalized remaining time factor, is a normalized enemy friendly distance factor; Threat level index : wherein for detecting the number of enemy forces in the field of view, for the ratio of the average distance of the nearest and second nearest enemy groups, for the difference value factor of the number of attack locks Drone state index : wherein are normalized endurance time, distance to the mother ship and distance to the target, respectively, are corresponding weight coefficients.

9. The method of claim 8, wherein, The decision tree model and tactical execution logic include: Decision tree training: using decision tree algorithm, defined tactical scenarios (including low threat priority attack, high threat condition attack, very high threat avoidance, etc. rules) to generate a set of training samples containing feature vectors and decision labels, wherein the feature vectors of the samples are obtained in the manner described in the above claims, and the sample labels are the corresponding tactical categories, including attack and retreat, thereby training the decision tree; Tactics execution: embodied in the tactics execution section of the simulation system, real-time information is obtained and the feature vector is calculated The decision evaluation is continuously carried out, the decision label of each moment is obtained at different moments, the judgment and change of the action logic are realized, and finally the safe attack behavior action or the safe avoidance behavior action is converged.

10. A UAV boat cluster area containment system for dynamic game play scenarios for implementing the method of any one of claims 1 to 9, characterized by, The system comprises: Simulation interaction environment module: used to build a virtual environment containing physical and rule constraints, which contains a task parameter interface for loading configurations including attack lock judgment conditions, agent dynamics restrictions such as maximum angular velocity, and actual task scene boundaries; this module is responsible for real-time calculation of the motion state of the UAV and the unmanned ship under the agreement, collision detection, and real-time output of environment state observation information; Agent modeling module: used to instantiate the UAV and unmanned ship objects, which contains: Dynamics unit: according to the preset speed and acceleration limit, calculate the pose of the agent in the next simulation step; sensor unit: according to the detection radius and field of view angle parameters, simulate the radar detection range and noise characteristics; Resource management unit: according to the preset endurance time and load quantity, monitor and update the power and remaining lock times of each agent; Algorithm decision control module: as the control core of the system, it interacts with the simulation environment through the inter-process communication interface, which specifically contains: Cooperative planning unit: run coverage search and geometric interception algorithm, generate preset navigation waypoints; Situation assessment unit: run Brownian motion prediction and Kalman filter observation update, output posterior state estimation and maximum posterior Y coordinate of black target; Game reasoning unit: load the trained decision tree model, calculate the feature vector according to the real-time situation Output instructions to each agent.

Images