A multi-unmanned aerial vehicle cooperative air combat strategic decision-making method based on large language model distillation

Abstract

The application discloses a multi-unmanned aerial vehicle cooperative air battle strategy decision-making method based on large language model distillation, and belongs to the technical field of unmanned aerial vehicle cooperative control. The method filters and normalizes key features of a battlefield situation, converts the key features into structured semantic text, generates an analyzable strategic intention based on a prompt word constraint large language model, maps the intention to a lower tactical cooperation layer task constraint, judges and updates the strategic task in advance based on a fixed cycle and an event triggering mechanism, adopts teacher-student distillation to train a lightweight student model and preferentially deploys the lightweight student model online. The application has the advantages that high-level strategic task allocation makes up for local tactical deficiencies, structured expression and constraint generation improves judgment stability and interpretability, the intention is effectively connected to the tactical layer, distillation deployment reduces reasoning delay and improves real-time performance.

Description

Technical Field

[0001] This invention relates to the field of collaborative control technology for unmanned aerial vehicles (UAVs), specifically to a method for allocating high-level strategic tasks among multiple UAVs based on large language model distillation. Background Technology

[0002] In recent years, with the development of beyond-visual-range air combat, multi-platform collaboration, and unmanned combat systems, UAV air combat decision-making has gradually shifted from single-platform maneuver control to multi-platform collaborative decision-making. A single UAV, limited by its detection range, weapon payload, and maneuverability, struggles to independently complete continuous search, attack, and evasion missions in complex combat environments. Multiple UAVs, working collaboratively, can complement each other in target detection, fire allocation, threat avoidance, and mission succession, thereby improving overall combat effectiveness.

[0003] Multi-UAV collaborative air combat decision-making typically involves multiple levels, including low-level maneuver control, tactical coordination, and high-level mission planning. The low-level focus is primarily on individual UAV heading, speed, and altitude control; the tactical level handles target allocation, skill switching, and local coordination; and the high-level level assesses the overall situation to determine the appropriate mission direction for the team, such as attack suppression, defensive evasion, search and recapture, or situational readjustment. Given the highly dynamic, highly contested, and information-impaired nature of air combat, effectively linking high-level strategic judgment with low-level tactical execution is a crucial issue in multi-UAV collaborative air combat decision-making.

[0004] Currently, multi-UAV collaborative decision-making methods mainly include rule-based methods, game-theoretic optimization methods, and multi-agent reinforcement learning methods. Rule-based methods rely on human experience and have limited ability to adapt to complex situations; game-theoretic optimization methods have high solution complexity in multi-platform and multi-objective scenarios; while multi-agent reinforcement learning methods can learn collaborative strategies, their outputs are mostly action numbers, task indices, or control parameters, lacking a clear team strategic intent. Therefore, existing methods still have shortcomings in high-level strategic task allocation and interpretable collaborative decision-making.

[0005] CN112947581B A Multi-Agent Reinforcement Learning-Based Multi-UAV Cooperative Air Combat Maneuver Decision-Making Method

[0006] This invention discloses a multi-UAV cooperative air combat maneuver decision-making method based on multi-agent reinforcement learning. It achieves multi-UAV cooperative maneuver decision-making by constructing an air combat situation, state space, action space, and reward mechanism. This scheme primarily addresses tactical-level cooperative maneuvering and target allocation issues; it does not construct a large language model for strategic intent generation based on the overall battlefield situation, nor does it address the mapping of strategic intent to the lower-level tactical coordination layer.

[0007] Reinforcement Learning Method for Multi-UAV Cooperative Adversarial Decision Making in CN113128698B

[0008] This invention discloses a reinforcement learning method for multi-UAV cooperative adversarial decision-making. It processes state data through transformation, completion, and normalization, and outputs the multi-UAV cooperative adversarial decision-making results based on the Actor-Critic algorithm. This scheme belongs to a typical numerical reinforcement learning decision-making method and can generate cooperative adversarial strategies, but it lacks semantic situational representation at the strategic level and interpretable task intent output.

[0009] CN117032300A Reinforcement Learning-Based Multi-UAV Multi-Target Cooperative Air Combat Maneuver Decision-Making Method

[0010] This invention discloses a multi-UAV cooperative air combat maneuver decision-making method based on reinforcement learning. It utilizes the DDPG algorithm to learn UAV swarm maneuver strategies for multi-UAV group confrontation and synchronous attacks. This approach primarily focuses on learning UAV swarm maneuver strategies and does not perform structured semantic modeling of high-level situational information such as radar lock-on, missile warning, attack windows, and boundary risks.

[0011] In existing technologies, most inventions focus on the design of multi-UAV cooperative maneuvers, target allocation, and local adversarial decision-making methods, without further developing into high-level strategic task allocation schemes applicable to multi-UAV beyond-visual-range cooperative air combat. Furthermore, existing inventions primarily rely on numerical state inputs and reinforcement learning strategy outputs, mainly addressing tactical-level issues such as target and task allocation, lacking a semantic understanding of the overall battlefield situation and phased task judgments, making it difficult to clearly define the high-level strategic intent the team should currently pursue. Secondly, while some methods can obtain cooperative strategies through multi-agent reinforcement learning, their decision-making process is mainly implicitly represented by neural networks, with outputs typically being action numbers or task indices, lacking clear decision-making basis and interpretability. Especially in beyond-visual-range air combat scenarios where multiple factors change simultaneously, how to transform complex situations into understandable and executable high-level strategic tasks and stably transmit them to the lower tactical coordination layer remains a significant challenge for existing technologies. Moreover, directly introducing large language models for high-level task planning results in open-ended natural language outputs and latency in online inference, making it difficult to directly meet the closed-loop requirements of real-time air combat. The aforementioned shortcomings limit the global mission planning capabilities, interpretability, and engineering deployment value of existing multi-UAV cooperative decision-making methods in complex air combat environments. Summary of the Invention

[0012] To address the aforementioned problems, this invention discloses a multi-UAV cooperative air combat strategic decision-making method based on large language model distillation.

[0013] The specific technical solution is as follows:

[0014] I. Decision-making methods and technical solutions

[0015] This invention constructs a hierarchical decision-making closed loop from global situational awareness to lower-level tactical execution. The method includes:

[0016] 1. Structured Situation Representation and Feature Selection

[0017] To address the challenge of directly applying high-dimensional numerical situational awareness data to high-level strategic judgment, the system acquires raw state and warning information from both sides of the battlefield environment. It then filters out key features relevant to high-level strategic judgment (including at least the number of surviving friendly drones, the number of surviving enemy targets, the total number of remaining friendly missiles, the number of terminal guidance warnings, the number of enemy targets captured by the team, and the minimum boundary security margin of the friendly platform). These key features are then normalized, and attack window scores and withdrawal urgency are calculated to form a state vector in a unified feature space. Subsequently, the numerical features are converted into structured semantic text with fixed fields and clear semantics, providing understandable input for a large language model.

[0018] 2. Constrained Large Language Model Strategic Intent Generation

[0019] To address the issue of open-ended output in the natural language output of large language models, the system constructs input prompts based on pre-defined prompt word templates. These prompts pre-include the task background, field meanings, optional strategic task categories, output format requirements, and security constraints. The optional strategic task categories are strictly limited to four types: concentrated fire annihilation, dispersed suppression, disengagement and reset, and conversion of advantages. Through these constraints, the large language model outputs only one of the pre-defined categories and returns information in a specified structured format (such as JSON) containing the strategic task category, the reason for selection, and the suggested target. This transforms the open-ended output into a parsable and verifiable structured strategic intent.

[0020] 3. Analytical mapping from strategic intent to tactical coordination layer

[0021] To effectively connect high-level strategic judgment with low-level tactical execution, the system analyzes the generated strategic intent and transforms it into task constraints or decision biases that can be directly executed by the lower-level tactical coordination layer. Specifically: if the objective is concentrated fire to annihilate, the system increases the priority of attack skills and guides priority allocation to locked targets; if the objective is dispersed suppression, the system reduces the focus weight of single targets and encourages multi-target pressure; if the objective is to escape and reset, the system increases the priority of evasion skills and generates retreat points or safe directions; if the objective is to convert advantages, the system increases the attack priority and relaxes the single-target allocation limit, encouraging the pursuit of remaining enemy aircraft.

[0022] 4. A replanning mechanism combining fixed cycles and event triggers.

[0023] To address the highly dynamic nature of the air combat environment, the system continuously monitors battlefield emergencies in addition to updating strategic missions at fixed intervals. When events such as initial radar acquisition, terminal guidance warning, aircraft loss, target loss, sudden changes in radar lock-on relationships, insufficient ammunition, or increased boundary risks occur, strategic replanning is immediately triggered, and the system re-enters the strategic mission allocation process to avoid using outdated strategies.

[0024] 5. Teacher-student distillation and lightweight deployment mechanism

[0025] To meet the real-time closed-loop requirements, the system employs a teacher-student distillation approach. A large language model is used as the teacher model to generate strategic task samples, which are then used to train a lightweight student model. During online deployment, the system prioritizes calling the strategic tasks output by the student model. When the highest category confidence score of the student model's output exceeds a set threshold, the task is directly adopted. When the confidence score is insufficient or the current situation exceeds the training coverage, the system reverts to calling the large language model or rules for auxiliary judgment, and stores the new samples in the dataset for subsequent updates to the student model. This approach retains high-level decision-making capabilities while reducing inference latency.

[0026] II. Decision System Technical Solution

[0027] Corresponding to the above method, this invention also provides a multi-UAV collaborative air combat strategic decision-making system based on large language model distillation. This system includes a situational information acquisition submodule, a strategic situational processing submodule, a structured semantic encoding submodule, a cue word constraint submodule, a large language model strategy generation submodule, a strategic intent parsing and mapping submodule, a replanning triggering submodule, a distillation training submodule, and a lightweight deployment submodule. These submodules work collaboratively in the order of "original situational input, key feature screening, structured text generation, strategic intent output, result parsing and mapping, post-execution monitoring and replanning, sample recording and distillation training" to execute the steps of the aforementioned decision-making method.

[0028] III. Storage Media Technology Solutions

[0029] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described multi-UAV cooperative air combat strategic decision-making method based on large language model distillation, thereby providing a carrier for the software deployment and engineering application of the solution.

[0030] The advantages of this invention are:

[0031] 1. A high-level strategic mission allocation method for multi-UAV cooperative air combat is provided, which can generate team-level missions such as attack suppression, defense evasion, search and recapture, and situation reorganization based on the overall battlefield situation, thus making up for the shortcomings of existing methods that focus on local tactical decision-making.

[0032] 2. Provides a structured situational awareness representation method that organizes information such as enemy and friendly numbers, remaining missile quantity, radar lock-on, missile warning, attack window, and boundary risks into inputs that can be understood by a large language model, thereby improving the stability and interpretability of strategic judgment.

[0033] 3. Provide a mapping mechanism from strategic intent to tactical coordination layer, transforming high-level tasks into executable information such as skill selection, target allocation, search anchor points, and retreat direction, thereby achieving effective connection between strategic decision-making and lower-level collaborative control.

[0034] 4. A large language model distillation deployment scheme is provided, which trains a lightweight strategic decision-making model through a teacher-student approach, reduces the latency of online calls to large language models, and improves system real-time performance and engineering deployment feasibility. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the layered architecture of a multi-UAV collaborative air combat strategic decision-making system provided in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the overall process of the strategic task allocation method provided in the embodiments of the present invention;

[0037] Figure 3 A schematic diagram illustrating the mapping relationship between strategic intent and tactical coordination layer provided in this embodiment of the invention;

[0038] Figure 4 This is a schematic diagram of the teacher-student model distillation and lightweight deployment process provided in an embodiment of the present invention. Detailed Implementation

[0039] 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.

[0040] This invention provides a multi-UAV cooperative air combat strategic decision-making method based on large language model distillation, such as... Figure 1As shown, this method is deployed in the high-level strategic task allocation module of a multi-UAV collaborative air combat decision-making system, and is connected to the simulation environment layer, tactical coordination layer, and skill execution layer. The simulation environment provides information such as the status of both friendly and enemy UAVs, target status, radar detection status, missile warning status, weapon usage status, and combat results; the strategic task allocation module generates team-level strategic tasks based on the overall situation; the tactical coordination layer completes target allocation, skill selection, search anchor point generation, and retreat direction selection among multiple UAVs based on the strategic tasks; the skill execution layer invokes corresponding search, attack, or evasion skills according to the task instructions output by the tactical coordination layer, and applies the final action instructions to the air combat environment.

[0041] The strategic task allocation module, located above the tactical coordination layer, does not directly output low-level continuous control parameters such as the UAV's heading, speed, and altitude. Instead, it outputs team-level strategic tasks such as attack suppression, defense evasion, search and recapture, and situational readjustment, further forming task constraints or decision biases identifiable by the tactical coordination layer. In this way, a hierarchical connection is established between high-level strategic judgment, tactical coordination decision-making, and low-level skill execution, enabling different decision-making layers to complete their respective tasks within their respective time scales.

[0042] Figure 2 This is the overall flowchart of strategic task allocation for this invention. It includes steps such as battlefield situation acquisition, strategic situation feature screening, state normalization processing, structured semantic text generation, cue word constraints, large language model strategic intent generation, strategic intent parsing and mapping, task issuance, and replanning judgment. During task execution, the system first acquires the global battlefield situation of both sides from the air combat environment or situational awareness system, and filters out key features related to strategic judgment. These features are then converted into structured semantic text and input into the large language model for strategic task judgment. After the large language model outputs the strategic intent, the system performs format verification and task parsing, and passes the parsing results to the lower-level tactical coordination layer. If situations such as missile terminal guidance warning, aircraft loss, target loss, sudden changes in radar lock-on relationship, insufficient missile quantity, or increased boundary risk occur during task execution, the system triggers strategic replanning and re-enters the strategic task allocation process.

[0043] Figure 3This is a mapping diagram between the strategic intent and the tactical coordination layer of this invention. It illustrates the correspondence between strategic intents such as concentrated fire annihilation, dispersed suppression, disengagement and reset, and conversion of advantages, and the output variables of the tactical coordination layer. Specifically, the intent to concentrate fire annihilation corresponds to increasing the priority of attack skills, prioritizing the allocation of locked targets, restricting ineffective search behavior, and maintaining a favorable attack formation; the intent to disperse and suppress corresponds to reducing the reward for focusing on a single target, encouraging multi-target allocation, restricting ineffective deep tracking behavior, and maintaining a safe distance between platforms; the intent to disengage and reset corresponds to increasing the priority of evasion skills, issuing retreat points or safe directions, reducing the priority of attack behavior, and maintaining support or alert tasks; the intent to convert advantages corresponds to increasing the priority of attack skills, improving parallel attack capabilities, encouraging target pursuit, and maintaining a favorable attack formation. Through the above mapping relationship, the high-level semantic tasks output by the large language model can be further transformed into task constraints or decision bias information that can be directly executed by the lower-level tactical coordination layer.

[0044] Figure 4 This is a flowchart illustrating the teacher-student distillation and lightweight deployment process of this invention. It explains the process by which a large language model acts as the teacher model to generate strategic task samples, and how the system records, cleans, and trains these samples to obtain a lightweight student model. During the deployment phase, the system prioritizes using the student model for rapid strategic task allocation. When the student model's output confidence is insufficient or the current situation significantly exceeds the coverage of the training samples, the large language model is then invoked for auxiliary judgment, and newly generated samples are added to the training dataset for subsequent updates to the student model.

[0045] The software component of the strategic task allocation module described in this invention includes a situational information acquisition submodule, a strategic situational processing submodule, a structured semantic encoding submodule, a cue word constraint submodule, a large language model strategy generation submodule, a strategic intent parsing and mapping submodule, a replanning triggering submodule, a distillation training submodule, a lightweight deployment submodule, and a data recording submodule. These submodules work collaboratively in the following order: "raw situational input, key feature filtering, structured text generation, strategic intent output, result parsing and mapping, post-execution monitoring and replanning, sample recording and distillation training." Specifically, the situational information acquisition submodule receives platform status and battlefield alarm information from both sides; the strategic situational processing submodule filters key features related to high-level strategic judgment from the raw situation; the structured semantic encoding submodule converts these features into text input with fixed fields and clear semantics; the cue word constraint submodule limits the output category and format of the large language model; and the large language model strategy generation submodule... The generation submodule is used to output team-level strategic tasks. The strategic intent parsing and mapping submodule is used to convert strategic tasks into executable constraint information for the lower-level tactical coordination layer. The replanning triggering submodule is used to determine whether to update the strategic tasks in advance based on conditions such as terminal guidance warnings, aircraft losses, target loss, boundary risks, and changes in ammunition quantity. The distillation training submodule is used to train a lightweight student model based on historical samples. The lightweight deployment submodule is used to prioritize calling the student model to complete rapid decision-making during the online phase. The data recording submodule is used to record the original situation, model output, final adopted tasks, and execution results to support retrospective analysis and subsequent updates.

[0046] The situational information acquisition submodule receives situational information from both friendly and enemy forces in the air combat environment or situational awareness system, and organizes it into raw situational data required for subsequent strategic judgment. Specifically, this submodule acquires at least the friendly UAV's number, survival status, current position, speed, altitude, heading, remaining missile quantity, radar detection status, missile warning status, whether it is currently engaged in attack, search, evasion, or retreat missions, and its relative position and orientation with the enemy target; it also acquires the enemy target's number, survival status, estimated position, speed, altitude, heading, whether it is detected by friendly radar, whether it is locked by friendly radar, whether it is within an attack window, and whether there is a terminal guidance threat. For missile warning information, it further distinguishes whether it is a terminal guidance warning, the target number from which the warning originates, and the duration of the warning. The above raw situational information is uniformly packaged into structured data and output to the strategic situational processing submodule for subsequent calculations of friendly and enemy quantity comparison, threat intensity, boundary risks, attack opportunities, and search opportunities.

[0047] The strategic situation processing submodule is used to filter out key features related to high-level strategic task allocation from the original battlefield situation and convert them into structured numerical indicators that can be judged by the large language model. Specifically, this submodule first counts the number of our surviving drones. Number of surviving enemy targets Total remaining ammunition of our side Number of terminal guidance alarms Number of enemy targets captured by the team and the minimum boundary security margin of our platform .in, This indicates the number of our drones that are currently still in service. This indicates the number of enemy targets that are still alive. This represents the sum of remaining ammunition for all our drones. This indicates the number of terminal-guided missile warnings issued against our platform in the current battlefield. This indicates the number of enemy targets that our radar has established a stable acquisition relationship with. This represents the minimum normalized safety margin from our platforms to the battlefield boundary. Based on this, the submodule further normalizes key features to obtain the state vector used as model input. :

[0048] in,

[0049] in, To score points for attacking the window, Due to the urgency of evacuation, This represents the current decision-making step. The above normalization process is used to eliminate the influence of inconsistent dimensions of different indicators, enabling large language models or student models to make judgments within a unified feature space. Among these, the attack window score... Calculate as follows:

[0050] This formula indicates that the attack window score increases when the attacker has established a significant number of target acquisition relationships; and further increases when the number of terminal guidance alarms is low. In the formula, This indicates the proportion of captured targets to the maximum possible number. This indicates a threat to the remaining space. This is used to prevent this item from becoming negative when there are too many alarms. Used to limit the results Within the range.

[0051] urgency of evacuation Calculate as follows:

[0052] in, Reflects the intensity of the terminal guidance alarm. This reflects the additional risks arising from insufficient boundary safety margins. When the number of final guidance alarms increases or the platform approaches the boundary... As the system grows larger, it becomes more inclined to select strategic tasks such as defense and evasion or situational reorganization.

[0053] The submodule also calculates the relative positions and approach speeds of the enemy and friendly formation centers. Let the coordinates of the friendly formation center be... The enemy formation's center coordinates are Then the relative displacement between the two is:

[0054] The distance between the centers of the two formations is:

[0055] If the center distances of two adjacent decision moments are respectively and The approach velocity can then be expressed as:

[0056] in, This indicates that the two sides are approaching each other. This indicates that both sides are moving away. This indicator is used to determine whether the current battlefield situation is entering a convergence phase or a phase of disengagement.

[0057] Finally, the submodule generates phase trigger flags based on the aforementioned characteristics, including whether disengagement is needed, whether it is suitable to transfer advantages, whether termination conditions are met, and whether regrouping is needed. Specifically, when... or When, the exit condition is triggered; when , and When, an advantage conversion judgment is triggered; when and When, the termination condition is triggered; when and When this occurs, a reorganization decision is triggered. The aforementioned triggering conditions are used to drive subsequent strategic task selections such as attack suppression, defense evasion, search and recapture, and situational reorganization.

[0058] The structured semantic encoding submodule is used to convert the numerical battlefield features output by the strategic situation processing submodule into structured semantic text that can be directly understood by the large language model. Specifically, this submodule generates input text according to a preset field order. The fields include at least the number of remaining friendly drones, the number of remaining enemy targets, the level of remaining friendly ammunition, the number of terminal guidance warnings, the number of targets captured by the team, the minimum boundary safety margin, the relative distance to the formation center, the attack window score, the urgency of withdrawal, the current decision step, the phase trigger flag, and candidate attack target information. Among them, the candidate attack target information includes at least the target number, the closest distance, the number of times it was captured, the number of times it was in the attack zone, and the priority score. The structured semantic text is output in a fixed key-value pair format, and the numerical values ​​are uniformly processed to retain the number of digits to avoid inconsistencies in the expression of the same semantics in different rounds. In this way, the original situation data is organized into text input with stable fields, clear semantics, and easy parsing, enabling the large language model to complete strategic task judgments in a unified context.

[0059] The prompt word constraint submodule is used to construct input prompts for the large language model based on structured semantic text and to impose category and format constraints on the model output. Specifically, this submodule pre-writes the task background, field meanings, optional strategic task categories, output format requirements, and security constraints into the prompt words. The optional strategic task categories include at least four types: concentrated fire annihilation, dispersed suppression, disengagement and reset, and conversion of advantage. Concentrated fire annihilation aims to concentrate superior forces to strike high-value targets locked by radar, striving to shoot them down. Dispersed suppression is suitable for confrontations in a balanced situation, simultaneously attacking multiple targets and providing a foundation for finding subsequent strike opportunities. Disengagement and reset means that the overall formation is at a disadvantage, requiring each aircraft to actively withdraw from the current combat area to evade multiple missile interceptions. Conversion of advantage typically occurs when the enemy formation suffers losses, the formation is superior in strength, and the enemy aircraft are in a poor air combat situation, encouraging the friendly aircraft to actively pursue the remaining enemy aircraft to achieve final victory. The prompts also require the large language model to output only one of the preset categories, prohibiting the output of free text unrelated to the category, and returning it in JSON format, including at least the strategic task category, the reason for selection, and the suggested target. If the output content does not conform to the preset format, it is considered an invalid result, and the model is requested to output according to the specified format again. In this way, the output of the large language model is transformed from open natural language into a parsable, verifiable, and distributable strategic task result.

[0060] The large language model strategy generation submodule receives the prompt text output by the prompt word constraint submodule and calls the large language model to generate team-level strategic intent. Specifically, this submodule takes structured semantic text, task category constraints, and security constraints as model input. The model output includes at least the strategic task category, the reason for task selection, and the suggested target. The reason for task selection explains the basis for choosing the strategic task under the current situation, and the suggested target indicates which type of enemy target or which group of friendly platforms the task mainly targets. After receiving the model output, the submodule first performs format validation, checking whether it meets the preset JSON structure, whether it contains required fields, and whether the field values ​​are within the allowed range. If the output is missing fields, has an incorrect format, or the strategic category exceeds the preset set, the output is considered invalid, and the same structured input is resent to the large language model, requiring it to re-output according to the specified format. If a valid result cannot be obtained after multiple retries, a preset rule strategy is used as a backup result to ensure that the strategic task allocation process is not interrupted. In this way, the output of the large language model is no longer an open natural language, but a structured strategic intent that can be parsed, verified, and distributed.

[0061] The strategic intent parsing and mapping submodule is used to transform the strategic intent generated by the large language model into task constraints or decision biases that can be used by the lower-level tactical coordination layer. If the strategic intent is concentrated fire annihilation, this submodule increases the priority of attack skill invocation and guides the tactical coordination layer to prioritize high-value targets that have been locked by radar, are within the attack window, and are at a suitable distance. It also suggests increasing the upper limit of platforms that can be deployed for attacks and setting a single target allocation limit to avoid resource congestion. If the strategic intent is distributed suppression, this submodule reduces the single target focus weight and encourages simultaneous pressure on multiple targets. It achieves multi-target targeting by reducing the maximum number of single target allocations and adjusting the target scoring strategy. The allocation is designed to maintain continuous containment of the enemy. If the strategic intent is to disengage and reset, this submodule prioritizes evasion and retreat skills, generates retreat points or safe directions based on the enemy's threat direction, boundary risks, and our platform's location, and issues constraints such as prohibiting deep pursuit and setting higher minimum boundary safety margins to ensure platform survival. If the strategic intent is to convert advantages, this submodule sets a more aggressive risk bias, moderately relaxes the single-target allocation limit, and increases the team attack limit. It prioritizes identifying and allocating resources to pursue remaining enemy aircraft to facilitate the conversion of advantages into annihilation results, thereby transforming the semantic strategic tasks output by the large language model into executable information at the tactical coordination layer.

[0062] The replanning triggering submodule is used to determine whether the current strategic task needs to be updated in advance, and supports initial capture triggering, fixed periodic triggering, terminal alarm triggering, and threat score triggering. Specifically, if an initial radar capture relationship has not yet been formed after the system has just completed a reset, replanning will not be initiated; the system will only allow entry into the strategic decision-making process when any platform of the Red Team forms a valid radar capture of any platform of the Blue Team. Let the current decision-making step number be... The basic strategic decision-making cycle is Then, periodic replanning is triggered when tmodT=0 is satisfied, where This is the current effective replanning cycle. If the dynamic cycle mechanism is not enabled, then... If the dynamic cycle mechanism is enabled, the team capture count will be calculated first. and threat score The cycle is determined based on whether the situation is one of combat or not.

[0063] in, To capture a threshold number of items, Threat score threshold, The decision-making cycle under combat situation, This is the default period in a non-combat situation. The threat score... The calculation is based on both the final guidance alarm and the boundary margin, specifically as follows:

[0064] in,

[0065] in, This indicates the number of terminal guidance warnings currently in use on the battlefield. This represents the minimum boundary security margin among our platforms. Therefore, the more terminal guidance warnings and the closer the boundary is, the higher the threat score. When the system enables terminal warning replanning, it triggers early replanning as soon as any enemy missile is detected in a terminal guidance warning state and the target is pointing towards our platform; when the system enables threat score replanning, as long as... Reaching the preset threshold This also immediately triggers advance replanning. Through the above methods, the system can update strategic tasks at fixed intervals and recalculate immediately when terminal guidance alarms are enhanced or boundary risks increase, thereby avoiding the use of outdated strategies.

[0066] The distillation training submodule is used to convert the large language model decision samples recorded during online operation into training data for a lightweight student model, and to complete the training of the strategic task classification model. Specifically, the system reads samples from historical logs, retaining only records whose sources contain multi-option decision labels and do not belong to the fallback results; each sample consists of a normalized feature vector. and strategic mission tags Composition, in which The values ​​are taken from four sets of plans:

[0067] {0,1,2,3}

[0068] ↔

[0069] {FOCUS_KILL,SPREAD_PRESSURE,SURVIVE_RESET,CONVERT_ADVANTAGE}

[0070] The feature vector is 10-dimensional and is defined as follows:

[0071] in, This indicates the number of our surviving drones. Indicates the number of surviving enemy targets. This indicates the total number of ammunition remaining on our side. Indicates the number of end-guided alarms. Indicates the number of targets captured by the team. Indicates the minimum boundary safety margin. Indicates the score for attacking the window. Indicates the urgency of the evacuation. This represents the current decision step. The student model is denoted as... Its output is the probability distribution of four types of strategic tasks:

[0072] in, Indicates sample Belongs to the The probability of a task class, satisfying:

[0073] The training objective uses the cross-entropy loss function:

[0074] in, The total number of training samples, For the indicator function, when the sample label equals the first... The value is 1 if the condition is met, and 0 otherwise. This is achieved by minimizing the loss function. The student model learns the strategic judgment patterns of the large language model under different battlefield situations. During online deployment, the system first calls the student model to output strategic tasks; if the student model's output does not meet the requirements, the large language model or rules are then called for auxiliary judgment.

[0075] The lightweight deployment submodule is used to replace the large language model in quickly allocating strategic tasks during the system's online operation phase. During deployment, the system prioritizes calling the lightweight student model to output the strategic task category. When the highest category confidence score output by the student model exceeds a set threshold, the system directly adopts that strategic task and sends it to the strategic intent parsing and mapping submodule. When the student model outputs an illusion, or the current situation clearly exceeds the coverage of the training data, the system then calls the large language model or rules for auxiliary judgment. New samples generated by the large language model can continue to be stored in the data recording submodule for subsequent updates to the student model. In this way, the present invention retains the high-level strategic judgment capabilities of the large language model while reducing online inference latency and improving the system's real-time deployment capabilities.

[0076] The data recording submodule is used to record key data during system operation, including the original battlefield situation, strategic situation characteristics, structured semantic text, output results of the large language model, output results of the student model, the final adopted strategic task, the reasons for replanning triggers, and task execution results. This submodule is used for task process tracing and manual review, and also for continuously expanding the distillation training dataset. Strategic task samples with poor execution performance or obvious unreasonableness can be marked and removed during subsequent data cleaning to improve the training quality of the student model.

[0077] During mission execution, the simulation training system or command and control personnel first set the parameters for the multi-UAV collaborative air combat mission, including the number of participating UAVs, the number of enemy targets, battlefield boundaries, initial deployment positions, radar detection rules, weapon usage rules, and strategic decision-making cycles. After the mission begins, the situation information acquisition submodule continuously receives status information from both friendly and enemy forces; the strategic situation processing submodule extracts key strategic features; the structured semantic encoding submodule generates a fixed-format battlefield situation text; the cue word constraint submodule combines this text with strategic mission rules and inputs it into a large language model; the large language model strategy generation submodule outputs team-level strategic intent; the strategic intent parsing and mapping submodule converts this intent into constraint information usable by the tactical coordination layer; the tactical coordination layer then completes target allocation, skill selection, and mission scheduling among multiple UAVs based on these constraints; and the skill execution layer finally outputs specific flight control or action commands, forming a closed-loop interaction with the air combat environment.

[0078] In one specific implementation, when the system detects that our remaining missile supply is sufficient, multiple enemy targets have been locked by radar, our drones are not under serious missile threat, and the attack window is favorable, the structured semantic encoding submodule organizes the above situation into text input, and the large language model strategy generation submodule outputs the attack suppression strategic task. Based on this, the strategic intent parsing and mapping submodule prioritizes the invocation of attack skills and guides the tactical coordination layer to prioritize the allocation of locked targets.

[0079] In another specific implementation, when the system detects that multiple friendly drones are threatened by missile terminal guidance, or that some drones have approached the battlefield boundary, the large language model strategy generation submodule outputs a defense and evasion strategy task. Based on this, the strategic intent parsing and mapping submodule generates withdrawal points or safe directions and guides the tactical coordination layer to prioritize the use of evasion skills.

[0080] In another specific implementation, when an enemy target is out of radar detection range for an extended period, or when our lock-on relationship with the enemy target significantly decreases, the large language model strategy generation submodule outputs a search and recapture strategy task. The strategic intent parsing and mapping submodule generates search anchor points based on the target's last detected position and our UAV distribution status, guiding multiple UAVs to re-establish detection relationships.

[0081] In another specific implementation, when our ammunition is insufficient, formation is scattered, or target allocation is chaotic, the large language model strategy generation submodule outputs a situational reorganization strategy task. The strategic intent parsing and mapping submodule lowers the attack priority and constrains the tactical coordination layer to readjust target allocation, formation relationships, and resource usage, enabling the team to restore a more stable state of coordination.

Claims

1. A multi-UAV cooperative air combat strategic decision-making method based on large language model distillation, characterized in that, Includes the following steps: Step 1: Obtain battlefield situation information from both sides and extract key features relevant to high-level strategic judgments. Step 2: Convert the key features into structured semantic text that the large language model can understand, and constrain the output of the large language model based on the preset prompt word template to generate parsable team-level strategic intent; Step 3: Analyze the strategic intent and map it into executable task constraints or decision biases of the lower-level tactical coordination layer before issuing it; Step 4: During task execution, determine whether the strategic task needs to be updated in advance based on a mechanism that combines fixed cycles and event triggers. If so, re-enter the strategic task allocation process. Step 5: Using a teacher-student distillation method, the large language model is used as the teacher model to generate strategic task samples to train the lightweight student model. During the online deployment phase, the student model is prioritized for rapid strategic task allocation.

2. The method according to claim 1, characterized in that, The key features mentioned in step 1 include at least the number of our surviving drones, the number of enemy surviving targets, the total number of our remaining ammunition, the number of terminal guidance warnings, the number of enemy targets captured by the team, and the minimum boundary safety margin of our platform; and the key features are normalized to calculate the attack window score and the urgency of withdrawal to form a state vector for model input.

3. The method according to claim 1, characterized in that, Step 2 constrains the output of the large language model based on a preset prompt word template, including: pre-writing the task background, field meaning, optional strategic task categories, output format requirements, and security constraints in the prompt words; the optional strategic task categories include concentrated fire annihilation, dispersed suppression, disengagement reset, and conversion of advantages; constraining the large language model to output the strategic intent in a fixed key-value pair format, the strategic intent including at least the strategic task category, the reason for selection, and the suggested target.

4. The method according to claim 3, characterized in that, Step 3, mapping the strategic intent to task constraints or decision biases, includes: If the strategic intent is to concentrate fire and annihilate, then increase the priority of attack skills and guide the priority allocation of locked high-value targets; If the strategic intent is to exert pressure in a decentralized manner, then the focus on a single objective should be reduced, and pressure on multiple objectives should be encouraged simultaneously. If the strategic intent is to escape the reset, then increase the priority of evasion skills, generate a retreat point or safe direction, and issue a constraint prohibiting deep pursuit. If the strategic intent is to convert an advantage, then increase the priority of attack skills and relax the single-target allocation limit to encourage the pursuit of remaining enemy aircraft.

5. The method according to claim 1, characterized in that, The conditions for triggering the event described in step 4 include: initial radar acquisition triggering, terminal guidance warning triggering, aircraft loss triggering, target loss triggering, sudden change in radar lock-on relationship triggering, insufficient missile quantity triggering, or increased boundary risk triggering; when the periodic replanning condition or any of the event triggering conditions are met, the system immediately triggers strategic replanning.

6. The method according to claim 1, characterized in that, In step 5, the online deployment phase prioritizes the use of the student model for rapid strategic task allocation, including: when the highest category confidence score output by the student model is higher than a set threshold, the strategic task is directly adopted; when the confidence score output by the student model is insufficient or the current situation exceeds the coverage of the training data, the large language model or preset rules are invoked for auxiliary judgment, and the new samples generated by the large language model are stored in the dataset for subsequent updates to the student model.

7. A multi-UAV cooperative air combat strategic decision-making system based on large language model distillation, characterized in that, It includes a situation information acquisition submodule, a strategic situation processing submodule, a structured semantic encoding submodule, a cue word constraint submodule, a large language model strategy generation submodule, a strategic intent parsing and mapping submodule, a replanning triggering submodule, a distillation training submodule, a lightweight deployment submodule, and a data recording submodule; The system is used to perform the method as described in any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

Images