Knowledge enhancement decision-making method for multiple unmanned aerial vehicles to intercept low-slow-small targets

Abstract

The invention discloses a knowledge enhancement decision-making method for multiple unmanned aerial vehicles to intercept low-slow small targets. The method can comprise the following steps: extracting key semantic units in a mixed extraction mode, and constructing an entity relationship; constructing a knowledge graph according to the entity relationship; and based on the knowledge graph, determining a hierarchical causal reasoning decision. The target intention is accurately recognized, the efficient and reliable interception strategy is generated, and the method has practicability in the practical application of intercepting and countering low-slow small target tasks.

Description

Technical Field

[0001] The present invention relates to the field of multi-agent decision control under confrontation, and more specifically, to a knowledge-enhanced decision-making method for multiple unmanned aerial vehicles to intercept low, slow and small targets. Background Art

[0002] Traditional decision-making approaches primarily rely on deterministic frameworks based on rule bases and expert systems, constructing task logic through finite state machines (FSMs) or event-condition-action (ECA) models. These approaches were widely used in early countermeasure systems, owing to their transparent logic and efficient response capabilities when event-triggered mechanisms were incorporated. Subsequently, hierarchical decision-making methods, namely the state-event-condition-action (SECA) approach, gradually evolved, integrating the FSM and ECA frameworks. However, when faced with nonlinearly maneuvering targets, the rigidity of pre-set strategies becomes apparent.

[0003] To break through the bottlenecks of traditional methods, reinforcement learning and game theory methods have gradually become research focuses. The multi-agent deep reinforcement learning framework enables drone swarms to autonomously learn collaborative strategies through distributed training and centralized execution mechanisms. For example, the introduction of a trajectory prediction module based on a long short-term memory network can infer maneuvering intentions through historical data when target information is missing, thereby improving the robustness of formation decision-making. At the same time, the game theory model transforms the pursuit problem into a dynamic adversarial deduction, optimizing the pursuit path by constructing a red and blue game tree. However, under some observable conditions, information asymmetry can easily lead to the failure of Nash equilibrium solution. Although these methods have shown potential in simulation environments, problems such as communication delays and sensor noise in actual scenarios can still cause strategy degradation, demonstrating the mismatch between theoretical models and engineering practice.

[0004] Breakthroughs in artificial intelligence (AI) technology have opened up new research paths for multi-UAV collaborative pursuit decision-making systems. Basic models based on large language models, such as ChatGPT and DeepSeek, leverage their near-human-level reasoning and task generalization capabilities. They are expected to flexibly adapt to diverse task requirements and generate efficient planning strategies through contextual learning and few-shot learning, even in complex environments and with targets with uncertain behavior. However, while current large-model-based decision-making methods have shown potential in tasks such as intent reasoning, they generally suffer from weak rule interpretability and a lack of structured domain knowledge, making them unable to meet the safety and reliability requirements of low-speed, small-target interception missions.

[0005] Therefore, it is necessary to develop a knowledge-enhanced decision-making method for multiple UAVs to intercept low, slow and small targets.

[0006] The information disclosed in the background technology section of the present invention is only intended to deepen the understanding of the general background technology of the present invention, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art. Summary of the Invention

[0007] The present invention proposes a knowledge-enhanced decision-making method for multiple UAVs to intercept low-altitude, slow, and small targets. The method can accurately identify the target intention and generate efficient and reliable interception strategies, and is practical in the actual application of interception and countermeasure tasks of low-altitude, slow, and small targets.

[0008] The present disclosure provides a knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets, including:

[0009] Extract key semantic units through hybrid extraction and construct entity relationships;

[0010] Constructing a knowledge graph based on the entity relationships;

[0011] Based on the knowledge graph, hierarchical causal reasoning decisions are determined.

[0012] Preferably, the hybrid extraction is achieved through entity recognition, relationship modeling and position constraint enhancement.

[0013] Preferably, constructing a knowledge graph based on the entity relationships includes:

[0014] Constructing a directed graph structure according to the entity relationship and the topological rules of the semantic network;

[0015] Construct reasoning chains to represent ordered paths;

[0016] Introduce the graph structure embedding method and determine the loss function of graph embedding.

[0017] Preferably, the chain of reasoning is:

[0018]

[0019] The loss function is:

[0020]

[0021] in, For the reasoning chain, each entity e i Represented as a vector is the relational mapping function.

[0022] Preferably, based on the knowledge graph, determining a hierarchical causal reasoning decision includes:

[0023] Introducing the reasoning chain decomposition mechanism to perform structured and hierarchical decomposition of small, slow, and small-scale pursuit and interception tasks;

[0024] A dynamic index scheduling mechanism is introduced to automatically activate and update the reasoning path by sensing data changes, giving the reasoning chain the ability to dynamically reconstruct and prompt switching.

[0025] Preferably, the reasoning chain decomposition mechanism divides the pursuit task from shallow to deep, and decomposes it into four stages: information processing, intention prediction, path planning and instruction generation according to the task processing logic, and constructs a structured task prompt chain through the reasoning chain decomposition module.

[0026] Preferably, in the information processing stage, the target perception unit extracts target state information from the fused multimodal input and outputs basic perception data, wherein the basic perception data includes position, speed, size, and confidence.

[0027] Preferably, in the intention prediction stage, the large model performs causal logic chain reasoning until it generates “ <predict>"mark.

[0028] Preferably, in the path planning stage, the prediction results are combined with the information of the knowledge graph to plan the behavior pattern of the coordinated UAVs, determine the key points of the interception route for each UAV, and generate " <pathplanning>"mark.

[0029] Preferably, during the instruction generation stage, the large model calls an external tool A* algorithm to generate a feasible trajectory for each drone based on key points.

[0030] Its beneficial effects are:

[0031] (1) Intention reasoning and interception decision control technology for low-speed, slow, and small targets: A low-speed, slow, and small target interception decision-making method based on a large language model is proposed to achieve accurate reasoning of the behavioral intentions of non-cooperative low-speed, slow, and small targets in complex environments and generate effective interception strategies.

[0032] (2) Causal knowledge enhanced reasoning framework: To address the problems that exist when large language models are directly used for interception decisions, a professional knowledge map for low-speed and small interception decisions is built, combined with retrieval enhancement generation and reasoning chain decomposition technology, which effectively improves the reliability and explainability of model decisions.

[0033] The method of the present invention has other features and advantages that will be apparent from or will be described in detail in the accompanying drawings and subsequent detailed description incorporated herein, which together serve to explain the specific principles of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention with reference to the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.

[0035] Figure 1 A flowchart showing the steps of a knowledge-enhanced decision-making method for multiple UAVs to intercept low, slow, and small targets according to an embodiment of the present invention is shown.

[0036] Figure 2 A diagram of a target intent reasoning and interception decision architecture according to an embodiment of the present invention is shown.

[0037] Figure 3a and Figure 3b Schematic diagrams respectively show decision results for airport no-fly zones and tall buildings according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The preferred embodiments of the present invention will be described in more detail below. Although the preferred embodiments of the present invention are described below, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0039] Figure 1 A flowchart showing the steps of a knowledge-enhanced decision-making method for multiple UAVs to intercept low, slow, and small targets according to an embodiment of the present invention is shown.

[0040] like Figure 1 As shown, the knowledge-enhanced decision-making method for multiple UAVs to intercept low, slow, and small targets includes:

[0041] Step 101: extract key semantic units through a hybrid extraction method and construct entity relationships;

[0042] Step 102: construct a knowledge graph based on entity relationships;

[0043] Step 103: Determine hierarchical causal reasoning decisions based on the knowledge graph.

[0044] In one example, hybrid extraction is achieved through entity recognition, relationship modeling, and location constraint enhancement.

[0045] In one example, building a knowledge graph based on entity relationships includes:

[0046] Construct a directed graph structure based on entity relationships and the topological rules of the semantic network;

[0047] Construct reasoning chains to represent ordered paths;

[0048] Introduce the graph structure embedding method and determine the loss function of graph embedding.

[0049] In one example, the chain of reasoning is:

[0050]

[0051] The loss function is:

[0052]

[0053] in, For the reasoning chain, each entity e i Represented as a vector is the relational mapping function.

[0054] In one example, determining a hierarchical causal reasoning decision based on a knowledge graph includes:

[0055] Introducing the reasoning chain decomposition mechanism to perform structured and hierarchical decomposition of small, slow, and small-scale pursuit and interception tasks;

[0056] A dynamic index scheduling mechanism is introduced to automatically activate and update the reasoning path by sensing data changes, giving the reasoning chain the ability to dynamically reconstruct and prompt switching.

[0057] In one example, the reasoning chain decomposition mechanism divides the pursuit task from shallow to deep, and decomposes it into four stages according to the task processing logic: information processing, intention prediction, path planning and instruction generation, and constructs a structured task prompt chain through the reasoning chain decomposition module.

[0058] In one example, in the information processing stage, the target perception unit extracts target state information from the fused multimodal input and outputs basic perception data, where the basic perception data includes position, speed, size, and confidence.

[0059] In one example, during the intent prediction phase, the large model performs causal logic chain reasoning until it generates <predict>"mark.

[0060] In one example, during the path planning phase, the prediction results are combined with the information in the knowledge graph to plan the behavior pattern of the coordinated drones, determine the key points of the interception route for each drone, and generate " <pathplanning>"mark.

[0061] In one example, during the instruction generation phase, the large model calls the external tool A* algorithm to generate feasible trajectories for each drone based on key points.

[0062] Specifically, this invention is a decision-making method for intercepting slow, small targets. By introducing a large language model, it enables inference and identification of target intent in complex environments, addressing the rigidity of interception strategies and poor environmental adaptability of traditional methods. It also establishes a causal knowledge-enhanced reasoning framework, significantly improving the model's decision quality and interpretability. The specific technical solution is as follows:

[0063] Figure 2 A diagram of a target intent reasoning and interception decision architecture according to an embodiment of the present invention is shown.

[0064] Establish a low-speed, slow, and small-target intention reasoning and interception decision-making architecture, such as Figure 2 As shown in the figure, the large language model receives target information, environmental information and other data from the outside as input, first establishes global situational awareness to provide the prerequisite for model decision-making; then it infers and identifies the target intention to obtain the prediction results of the target behavior and trajectory; then it generates the corresponding task plan based on the inference results, the position of the friendly interceptor and the task situation, and assigns tasks in the form of structured data to realize the dynamic interception decision-making process for low, slow and small targets.

[0065] Extracting key semantic units (such as targets, obstacles, tracks, and no-fly zones) from input data and their semantic relationships is the foundation for building a knowledge graph. This paper adopts a hybrid extraction method of "entity recognition + relationship modeling + position constraint enhancement."

[0066] Entities are primarily categorized into three types: target entities, such as drone_12 (a low-flying, slow, small drone) and flight_CA123 (a passenger aircraft); environmental entities, such as no-fly-zone_A (a no-fly zone) and building_block_X (a high-rise building); and rule entities, such as proximity_limit (a proximity threshold) and priority_level (a priority level). Entity recognition combines perception results with prior task labels. Based on perception information labels and external task configuration files, key elements such as low-flying, slow, small drones, terrain structures, no-fly zones, and flight paths are identified. By combining entity recognition with spatial annotation methods, the physical information and semantic types of entities can be accurately labeled. For example, the label "drone" obtained from target recognition can be annotated as <low-flying, slow, small target drone> based on location information and trajectory data.

[0067] In terms of relationship modeling, relevant templated semantic relationships are designed according to the task scenario. For example, "close to" represents the proximity of the target to the no-fly zone / important building; "located in" represents the geographical containment relationship between the target / obstacle and the area; "belong to" represents the hierarchical attribution of the task entity to the environmental classification. These are the key connection elements for constructing the semantic structure in the form of basic triples.

[0068] In terms of enhancing location constraints, entity consistency verification is performed by combining multi-modal information such as coordinates, confidence levels, and timestamps to filter out meaningless or incorrect information.

[0069] Finally, each valid semantics forms a semantic triple <h, r, t>, where h is the head entity, r is the relationship type, and t is the tail entity, which will subsequently serve as the basic units of the graph nodes and edges. For example, if the detection result indicates that the target numbered drone_12 is close to the no-fly zone A, then the triple <drone_12, close to, no-fly zone A> can be constructed.

[0070] A partial list of triples is shown in Table 1:

[0071] Table 1 Example list of triples

[0072]

[0073] After the entity relationships are constructed, the system constructs the semantic triples into a directed graph structure G=(V, E) based on the graph database according to the topological rules of the semantic network, where the nodes V represent the task-related entities and the edges E represent the semantic relationships, supporting graph traversal and inference path generation. To support causal reasoning, the graph is further organized into an inference chain represented as a set of ordered paths:

[0074]

[0075] At the same time, a graph structure embedding method is introduced to represent each entity e i as a vector The relationship mapping function is Then the loss function of the graph embedding is defined as:

[0076]

[0077] This loss function is used to optimize the consistency of the graph structure while supporting the causal path transmission from the graph to the larger model. The graph layer consists of a target layer, an environment layer, a rule layer, and a task layer. Multiple edges connect the nodes in each layer. For example, there are "located" and "avoid" relationship edges between flights and "suspicious flying objects," and "contain" or "affect" edges between no-fly zones and terrain. This structure supports embedded learning of models such as graph neural networks and facilitates task-driven graph traversal and path planning. The four-level multi-layered knowledge graph structure is as follows:

[0078] (1) Target layer: including target drones, suspicious flying objects, flights, coordinated drones, etc.

[0079] (2) Environmental layer: including terrain, obstacles, no-fly zones, buildings, etc.;

[0080] (3) Rule layer: including action restrictions, approach levels, path avoidance rules, etc.

[0081] (4) Task layer: including target number, interception instructions, priority allocation, etc.

[0082] The knowledge graph structure is stored in the Neo4j graph database, which supports real-time query and graph traversal operations, and facilitates linkage with large model interfaces.

[0083] In addition, to enhance the knowledge coverage capability of the graph, the system also supports the introduction of external knowledge sources (such as flight manuals and countermeasures regulations) for semantic completion, and automatically aligns synonymous nodes through embedded semantic similarity matching methods (such as BERT embedding cosine) to achieve continuous evolution of the knowledge graph.

[0084] At the application level of reasoning, the knowledge graph provides explicit query support and contextual awareness for the large-scale model's reasoning engine. For example, when determining whether a target is an illegal intrusion, the large-scale model can call the graph interface to retrieve semantic labels such as "near a no-fly zone" and "path crossing a civil aviation route." It then integrates trajectory information to comprehensively infer decision intent, thereby achieving semantically interpretable output for the command chain. This mechanism not only enhances the causal nature of the reasoning process but also significantly improves the interpretability, robustness, and safety of task execution.

[0085] By constructing a knowledge graph, the system converts perception data into semantic information with causal relationships, providing structured knowledge support for the causal reasoning module in the next section.

[0086] To enable the precise pursuit and efficient interception of small, slow, and low-lying targets by multiple drone swarms in complex environments, this paper proposes a hierarchical causal reasoning mechanism as a key component of the Causal Knowledge Augmented Reasoning (CKAR) framework. This mechanism decomposes the task structure, constructs a causal chain to drive step-by-step reasoning within a large model, and integrates perception, knowledge, and decision-making modules through a multi-level prompt engineering and dynamic indexing mechanism, achieving a semantic closed loop from information comprehension to decision execution.

[0087] Faced with pursuit scenarios characterized by high task complexity and a wide range of semantics, the system first introduces a reasoning chain decomposition mechanism to structure and hierarchically decompose low-speed, slow, and small pursuit and interception tasks. This mechanism divides pursuit tasks from shallow to deep, breaking them down into four logical units based on task processing logic: information processing, intent prediction, path planning, and command generation. Furthermore, the reasoning chain decomposition module constructs a structured task prompt chain.

[0088] In the information processing stage, the target perception unit extracts target state information from the fused multimodal input and outputs basic perception data including position, speed, size, and confidence.

[0089] The reasoning chain decomposition mechanism also relies on the tool usage capabilities of the large model. Drawing on the method of integrating the tool usage capabilities of the large model in previous work, in the intention prediction stage, the large model performs causal logic chain reasoning until " <predict>"Tag.Generate" <predict>After the "" mark is placed, the large model stops reasoning and calls the external LSTM tool to predict the future trajectory of the low, slow, small target. This operation uses the LSTM professional tool to model the historical state sequence of the input target, predict its future motion path, and judge its potential behavioral intentions based on this. The ability to schedule external model tools greatly improves the large model's reasoning efficiency and the accuracy of its behavior understanding, while reducing its internal reasoning overhead.

[0090] In the path planning stage, the prediction results are combined with the environment, task priority and other information in the knowledge graph to plan the behavior pattern of the coordinated drones, such as encirclement and interception, circle compression, etc., and accordingly arrange the specific interception route key points of each drone to generate " <pathplanning>"mark.

[0091] generate" <pathplanning>After the " " mark, during the instruction generation phase, the large model calls the external tool A* algorithm to generate feasible trajectories for each drone based on key points. In the entire reasoning chain, there is a clear causal dependency between each layer. The decision-making process starts from the bottom-level perception and progresses layer by layer, forming a complete reasoning path chain.

[0092] Based on the structured subtasks output by the inference chain decomposition mechanism, the system converts the inference chain into a sequence of natural language prompts, embedding the trajectory prediction results and the current task status. This generates a reasoning context with a clear causal path, guiding the large model's logical reasoning and instruction generation. This mechanism leverages hierarchical prompt engineering to embed the relationship chain in the knowledge graph into the prompt context, forming a structured pursuit instruction chain with causal logic.

[0093] For example, if the current mission status is "target drone_12's speed is increasing, its direction is shifting eastward, and its distance from no-fly zone B is less than the safety threshold," the following prompt template can be generated: "The current target number is {drone_id}, its historical trajectory is {[pos1,pos2,…]}, its speed is {vel}, its current direction is {dir}, and it is about to enter the no-fly zone {zone_id}. Please combine the mission knowledge graph with the current context to determine its behavioral intent and generate a pursuit strategy." When this template is passed to the large model, it automatically triggers a structured reasoning path, generating a set of causal chain instructions from target intent recognition and trajectory evaluation to mission execution. An example of model-generated natural language inference output is as follows: "Target drone_12 is expected to enter no-fly zone B within 5 seconds. Its current path crosses route CA123. Based on the current distribution of collaborative units, it is recommended that drone_04 and drone_05 engage in a frontal attack, drone_03 block its retreat, and alert the command center." Guided by structured prompts, the large model not only needs to call on knowledge graphs, perception data, and mission rule information, but also convert the current environmental state sequence into executable language instructions.

[0094] To improve the real-time performance and environmental responsiveness of the inference system, this method proposes a dynamic index scheduling mechanism. While traditional inference processes are limited by static prompts and struggle to adapt to real-time changes, this mechanism automatically activates and updates the inference path by sensing data changes. This enables dynamic reconfiguration and prompt switching within the inference chain, effectively addressing complex situations such as sudden target avoidance and no-fly zone changes.

[0095] The system maintains a trigger index table for each subtask, tied to the perception state. When a target's state changes (such as acceleration, sudden change in direction, or illegal approach to a no-fly zone), it automatically triggers an update to the corresponding inference chain. For example, if a target approaches a no-fly zone, the path risk inference chain is activated; if multiple targets converge, the task division and strategy update chain is activated. This mechanism minimizes ineffective inference overhead and ensures that model inference is always synchronized with the latest perception state, effectively addressing the issues of information timeliness and task coordination during multi-target tracking.

[0096] This method is based on large language model technology. It accurately infers the behavioral intentions of low-speed, small targets based on environmental and state information, and then dynamically generates interception strategies that adapt to the target behavior and environment. It avoids the problems of rigid decision-making and difficulty in dealing with sudden target behavior and environmental interference in traditional methods, and significantly improves the success rate of capturing low-speed, small targets in complex environments. In order to address the problems of lack of knowledge structure and weak interpretability in the field of low-speed, small target interception decision-making of large language models, knowledge graph index-driven and retrieval-enhanced generation technology are used to inject relevant knowledge into them. At the same time, the task of capturing low-speed, small targets is structured and hierarchically decomposed through reasoning chain decomposition, forming a causal knowledge-enhanced reasoning framework, which greatly improves the quality and interpretability of the model generation strategy and effectively makes up for the defects that exist when general large language models are directly used for decision-making tasks.

[0097] To facilitate understanding of the solutions and effects of the embodiments of the present invention, a specific application example is given below. Those skilled in the art should understand that this example is only for facilitating understanding of the present invention, and any specific details thereof are not intended to limit the present invention in any way.

[0098] Example 1

[0099] To validate the effectiveness and practicality of the proposed large-scale model-based dynamic collaborative decision-making system and its core module, the CKAR framework, a multi-scenario, multi-metric simulation experiment was designed and implemented. The system simulated the illegal intrusion of small, slow, and low-lying targets in various complex urban environments and deployed a coordinated swarm of drones to perform pursuit and containment missions. The system's performance was comprehensively evaluated in terms of multimodal fusion accuracy, inference correctness, pursuit success rate, and command timeliness.

[0100] (1) Configuration and evaluation plan

[0101] The core algorithm configuration used in the experiment is as follows: the target model is mainly a rotary-wing drone with diverse behavioral strategies and avoidance capabilities; the knowledge graph engine is stored in JSON-LD format and synchronized with the Neo4j graph database; the causal reasoning module is connected to a pre-trained large model (DeepSeek-32B) and combines Prompt engineering with graph embedding for reasoning.

[0102] In order to comprehensively evaluate the system performance, the present invention designs the following core evaluation indicators, as shown in Table 2.

[0103] Table 2 Decision-making core evaluation indicators

[0104]

[0105] In order to verify the performance advantages of the present invention, the following three comparison schemes are designed:

[0106] ①Baseline-R: Traditional static rule system (no graph, no LLM);

[0107] ②LLM-only: Large language model direct decision system without graph support;

[0108] ③CKAR (the present invention): a complete system integrating graph, reasoning chain and Prompt strategy.

[0109] (2) Experimental results and analysis

[0110] The experiment was conducted in 50 randomly generated groups of the same task environment. The results are shown in Table 3:

[0111] Table 3 Decision-making core evaluation indicators

[0112]

[0113] Results show that the CKAR system outperforms traditional rule-based systems and pure large-model systems in terms of pursuit success rate, reasoning consistency, and command error, with particularly significant improvements in reasoning quality and safety. Notably, the CKAR system's advantages are particularly pronounced in terms of reasoning response time (RT) and error rate (FER). This is attributed in part to the structured control of the causal inference path and in part to the efficient scheduling of the external LSTM prediction module by the large model. Instead of analyzing trajectory trends from scratch, the large model uses the results from the prediction module to rapidly establish the three-step chain of "trajectory → intention → command," significantly reducing the contextual reasoning load and improving overall command quality and time efficiency. The average response time is superior to the LLM-only solution, demonstrating greater real-time performance. Furthermore, in extreme scenarios where targets approach both no-fly zones and flight routes simultaneously, the Baseline-R system frequently experiences decision conflicts and the LLM-only system exhibits misjudgment. However, the CKAR system accurately captures changes in key variables through causal reasoning and outputs a rationalized and interpretable command chain.

[0114] Figure 3a and Figure 3b Schematic diagrams respectively show decision results for airport no-fly zones and tall buildings according to an embodiment of the present invention.

[0115] In order to verify the reasoning effect and instruction rationality of the proposed CKAR system under different task constraints, two scenario-based numerical simulation experiments were designed, simulating the airport no-fly zone defense task and the urban building protection task respectively. Figure 3a 、 Figure 3b As shown in the figure, in each round of tasks, the target's behavioral trajectory is assumed to be the same. Targeted prompts are generated for different tasks, triggering a causal reasoning chain and outputting collaborative decision-making instructions.

[0116] exist Figure 3a In the airport no-fly zone defense mission shown, the system detected that the target aircraft was rapidly approaching a high-value area (circular red area) and generated a more aggressive strategy: allowing the drone to be destroyed in exchange for absolute safety. The system assigned three drones to multiple forward key points in the target's predicted trajectory (red crosses indicate predicted points), forming a time-progressive interception strategy. Drone 1 reached the first predicted point in 0.886 seconds, while Drones 2 and 3 completed interception at their corresponding predicted points in 1.453 and 2.072 seconds, respectively, achieving a multi-layered time-window coordinated suppression effect. The advantages of this strategy are: 1. It creates multi-layered interception redundancy, forming a progressive pursuit chain; 2. It sacrifices a certain drone survival probability in exchange for maximizing mission success, complying with the "unbreakable target" principle in airport scenarios.

[0117] Figure 3b In the urban building protection mission shown, the system reasoned and generated a mild pursuit strategy. Considering that buildings have lower defense levels than airports, the system prioritized avoiding losses during reasoning, generating a more circumventing strategy, despite the higher risk of the predicted trajectory intersecting with buildings. Three drones were assigned to three key nodes along the predicted path, generating a pursuit strategy that encircled the target.

[0118] In summary, the CKAR system can reasonably distinguish the levels of task scenarios based on the input perception state and environmental map, and generate matching collaborative instructions; the instruction sequence has good performance in terms of temporal consistency and logical rationality, demonstrating the dynamic consistency, task adaptability and causal interpretability of the system reasoning results.

[0119] Those skilled in the art should understand that the above description of the embodiments of the present invention is only for the purpose of illustrative purposes only to illustrate the beneficial effects of the embodiments of the present invention, and is not intended to limit the embodiments of the present invention to any given examples.

[0120] While various embodiments of the present invention have been described above, the above description is intended to be illustrative, not exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.< / pathplanning> < / pathplanning> < / predict> < / predict> < / pathplanning> < / predict> < / pathplanning> < / predict>

Claims

1. A knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets, characterized by: include: Extract key semantic units through hybrid extraction and construct entity relationships; Constructing a knowledge graph based on the entity relationships; Based on the knowledge graph, hierarchical causal reasoning decisions are determined.

2. The knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets according to claim 1, wherein: The hybrid extraction is achieved through entity recognition, relationship modeling and position constraint enhancement.

3. The knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets according to claim 1, wherein: Constructing a knowledge graph based on the entity relationships includes: Constructing a directed graph structure according to the entity relationship and the topological rules of the semantic network; Construct reasoning chains to represent ordered paths; Introduce the graph structure embedding method and determine the loss function of graph embedding.

4. The knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets according to claim 3, wherein: The chain of reasoning is: The loss function is: in, For the reasoning chain, each entity e i Represented as a vector is the relational mapping function.

5. The knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets according to claim 1, wherein: Based on the knowledge graph, determining the hierarchical causal reasoning decision includes: Introducing the reasoning chain decomposition mechanism to perform structured and hierarchical decomposition of small, slow, and small-scale pursuit and interception tasks; A dynamic index scheduling mechanism is introduced to automatically activate and update the reasoning path by sensing data changes, giving the reasoning chain the ability to dynamically reconstruct and prompt switching.

6. The knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets according to claim 5, wherein: The reasoning chain decomposition mechanism divides the pursuit task from shallow to deep, and decomposes it into four stages according to the task processing logic: information processing, intention prediction, path planning and instruction generation, and constructs a structured task prompt chain through the reasoning chain decomposition module.

7. The knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets according to claim 6, wherein: In the information processing stage, the target perception unit extracts target state information from the fused multimodal input and outputs basic perception data, where the basic perception data includes position, speed, size, and confidence.

8. The knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets according to claim 6, wherein: In the intention prediction stage, the large model performs causal logic chain reasoning until it generates <predict> "mark.< / predict> 9. The knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets according to claim 6, wherein: In the path planning stage, the prediction results are combined with the information of the knowledge graph to plan the behavior pattern of the coordinated drones, determine the key points of the interception route for each drone, and generate <pathplanning> "mark.< / pathplanning> 10. The knowledge-enhanced decision-making method for multiple UAVs to intercept low-speed, slow, and small targets according to claim 6, wherein: During the instruction generation phase, the large model calls the external tool A* algorithm to generate feasible trajectories for each drone based on key points.

Images