Anti-drone strategy generation method and system
By using a distributed policy memory and a multi-agent self-game mechanism, the problem of insufficient adaptability of existing anti-drone systems in complex environments is solved, enabling continuous accumulation of tactical knowledge and efficient defense, and improving the system's adaptability and stability.
Patent Information
- Application Number
- CN202611114766.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-27
- Publication Date
- 2026-08-25
AI Technical Summary
Existing anti-drone systems lack adaptability, have high computational complexity, lack foresight in tactical generation, and are susceptible to single points of failure due to their centralized architecture when facing complex dynamic environments and intelligent drone intrusions, resulting in insufficient effectiveness in actual combat.
Employing a distributed, dynamically updated policy memory and a multi-agent self-game mechanism, the system achieves continuous accumulation and sharing of tactical knowledge through a policy generation model trained by self-game and confidence-adaptive weight fusion, supporting distributed decision-making and rapid response.
It improves the system's adaptability and stability in complex environments, enhances the innovation and resilience of tactical generation, and ensures efficient defense against intelligent intrusions.
Smart Images
Figure CN122635567A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of anti-drone technology, and in particular to a method and system for generating anti-drone strategies. Background Technology
[0002] With the rapid development and widespread application of drone technology, illegal drone intrusion incidents are increasing, posing a serious threat to critical infrastructure, public places, and privacy security. Counter-drone systems, as a key technology to address this challenge, have received widespread attention from academia and industry in recent years. In existing technologies, counter-drone systems mainly rely on multi-source sensor fusion technologies such as radar, electro-optical, and radio detection to identify target drones, and then implement countermeasures through communication jamming, navigation deception, net interception, and laser destruction. Among these, the tactical decision generation module, as the core component, is responsible for transforming perceived information into effective countermeasure strategies; its performance directly determines the system's combat effectiveness against intelligent drones.
[0003] Currently, anti-drone tactical generation technology mainly adopts an architecture combining a pre-defined script rule base and a finite state machine. This architecture predefines a set of tactical rules covering typical scenarios and triggers corresponding strategies by matching environmental parameters and pre-defined conditions in real time. Typical implementation schemes can be divided into three technical routes: (1) rule reasoning scheme based on expert knowledge base, which constructs a decision tree through radio frequency feature recognition and threat level assessment, but lacks adaptability when facing unknown intrusion patterns; (2) threat fusion scheme based on probabilistic reasoning, which integrates Bayesian networks to process multi-source perception uncertainty, but its static parameter configuration is difficult to adapt to dynamically changing adversarial environments; (3) hierarchical response scheme based on fuzzy logic, which uses membership functions to process perception noise and generate progressive countermeasures, but its rule set update mechanism relies on manual intervention and cannot achieve autonomous evolution.
[0004] Although existing anti-drone systems have shown some effectiveness in specific scenarios, their tactical generation mechanisms have four key limitations, which severely restrict the system's practical value in complex and dynamic environments.
[0005] First, at the learning capability level, existing systems lack mechanisms for continuous learning and knowledge accumulation. Each interception task is treated as an independent event, and lessons learned from historical tasks cannot be effectively distilled into system knowledge. When faced with similar threat patterns, the system cannot invoke historically successful strategies but instead repeats the same decision-making process, failing to autonomously improve from failures. Even systems employing reinforcement learning typically have their training completed offline, and once deployed, the model remains fixed, unable to adapt to environmental changes and new threats, leading to a "separation of learning and application" problem. More seriously, existing methods cannot distinguish whether tactical failures are caused by strategic flaws or execution errors, further hindering the system's autonomous evolution.
[0006] Secondly, regarding the complexity of adversarial operations, current systems are primarily designed for traditional fixed-route drones and are ineffective against intelligent drones with autonomous decision-making capabilities. When intruders employ adaptive strategies similar to reinforcement learning, static rule-based defense systems quickly fail. Furthermore, in multi-drone collaborative intrusion scenarios, existing methods struggle to handle high-dimensional policy spaces and complex interactions, leading to an exponential increase in system computational complexity. This often results in a degradation to simple priority rules, sacrificing tactical optimization quality. This lack of "adversary modeling" capability puts existing systems at a significant disadvantage in intelligent adversarial environments.
[0007] Furthermore, at the decision-making mechanism level, script-based tactical generation methods are inherently reactive, lacking foresight and creativity. The limited number of pre-set scripts and their long update cycles make it difficult to cover increasingly complex drone intrusion scenarios. Especially when facing intelligent drones employing adaptive flight strategies, the system can only select suboptimal solutions from pre-set options, failing to generate innovative countermeasures. The switching logic between scripts is typically based on simple threshold judgments, resulting in abrupt "step-like" changes in tactical behavior, hindering smooth transitions and fine-tuning, and reducing the system's stability under critical conditions.
[0008] Finally, at the system architecture level, existing anti-drone systems generally adopt a centralized, layered design, which poses a risk of single point of failure and poor resilience. Decision-making logic relies entirely on the central controller, with individual execution nodes lacking independent decision-making capabilities and a mode-switching mechanism based on network health. In strong electromagnetic warfare environments, once the central controller's communication link is interfered with or a node fails, the entire system cannot work collaboratively and cannot degrade to a distributed, locally autonomous mode. This architectural flaw leads to a sharp decline in overall defensive effectiveness when some nodes are damaged, failing to meet the requirements of high-reliability operations. Summary of the Invention
[0009] The purpose of this invention is to provide a method and system for generating anti-drone strategies to overcome the limitations of existing technologies.
[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for generating anti-drone strategies includes the following steps: when drone intrusion is detected, the current environmental state vector is obtained, and a preset number of preset strategies corresponding to the current environmental state vector are searched in a preset memory based on the current environmental state vector, so as to determine a first alternative strategy according to the preset strategies. The current environment state vector is input into a preset strategy generation model to obtain a second alternative strategy. The preset strategy generation model is a strategy generation model trained by self-game and equipped with a strategy optimization function. The strategy optimization function includes a strategy performance optimization term, a value function evaluation term, and a strategy entropy term. Based on the first alternative strategy and the second alternative strategy, a target strategy for drone intrusion is generated according to a preset strategy fusion rule.
[0011] Optionally, before the step of searching a preset number of preset strategies corresponding to the current environment state vector in a preset memory based on the current environment state vector, the method further includes: Obtain historical strategies for completed adversarial tasks and execution data corresponding to those historical strategies; The comprehensive evaluation value of the historical strategy is calculated based on the execution data, and the comprehensive evaluation value is matched with a preset threshold to determine whether the historical strategy is effective. When the historical strategy is valid, the local sensitive hash index of the historical strategy is calculated to construct a preset memory.
[0012] Optionally, the step of searching a preset number of preset strategies corresponding to the current environment state vector in a preset memory based on the current environment state vector, and determining the first candidate strategy according to the preset strategies, includes: By calculating the similarity data between each preset environment state vector in the preset memory and the current environment state vector, a preset number of preset environment state vectors are selected from the preset memory in descending order of similarity data as target environment state vectors. Based on the preset strategy corresponding to the target environment state vector, the first alternative strategy is obtained through a similarity-weighted fusion method.
[0013] Optionally, before the step of inputting the current environment state vector into the preset strategy generation model to obtain the second alternative strategy, the method further includes: Construct a virtual simulation adversarial environment that includes an electromagnetic interference model, six-degree-of-freedom kinematic constraints, sensor noise, and communication delay; Define the state space, action space, and reward function. Through self-game training in the virtual simulation adversarial environment, train an initial model based on intelligent algorithms and used for policy generation, including a policy optimization function, to obtain a preset policy generation model.
[0014] Optionally, the step of generating a target strategy for drone intrusion based on the first alternative strategy and the second alternative strategy according to a preset strategy fusion rule includes: According to the preset strategy fusion rules:
[0015] The first and second alternative strategies are merged to obtain a target strategy for drone intrusion, wherein... As the first alternative strategy, As a second alternative strategy, This is the current environment state vector. The confidence level is used for adaptive weights.
[0016] Optionally, the method further includes: The target strategy is evaluated based on predicted interception probability, energy consumption indicators, security indicators, and strategy flexibility indicators to obtain a comprehensive score for the target strategy. If the overall score is lower than the overall threshold, a better strategy is searched in the local neighborhood of the target strategy using a Bayesian optimization algorithm, and the better strategy is output.
[0017] Optionally, the method further includes: Obtain the operational and attribute data of the anti-intrusion system, and calculate the formula based on the status:
[0018] Obtain the status score of the anti-intrusion system; wherein, As a normalized connectivity index, This represents the current number of connected nodes. This represents the total number of nodes in the system. For normalized delay indicators, The average communication delay (ms) This is the delay threshold; For normalized throughput, Current bandwidth (Mbps) Minimum required bandwidth; To prevent division by zero of small constants; Let be the weighting coefficient, satisfying ; when When the threshold value is less than a preset state threshold, the system controls each node within the anti-intrusion system to perform the following steps: when a drone intrusion is detected, obtain the current environmental state vector, and search a preset number of preset strategies corresponding to the current environmental state vector in a preset memory based on the current environmental state vector to determine a first candidate strategy; input the current environmental state vector into a preset strategy generation model to obtain a second candidate strategy, wherein the preset strategy generation model is a strategy generation model trained through self-game theory and equipped with a strategy optimization function, the strategy optimization function including a strategy performance optimization term, a value function evaluation term, and a strategy entropy term; and generate a target strategy for drone intrusion based on the first and second candidate strategies according to a preset strategy fusion rule.
[0019] On the other hand, this application provides an anti-drone strategy generation system, including: The first calculation module is used to obtain the current environmental state vector when the drone intrusion is detected, and search for a preset number of preset strategies corresponding to the current environmental state vector in a preset memory bank based on the current environmental state vector, so as to determine the first alternative strategy according to the preset strategy. The second calculation module is used to input the current environment state vector into the preset strategy generation model to obtain the second alternative strategy. The preset strategy generation model is a strategy generation model that has been trained through self-game and has a strategy optimization function. The strategy optimization function includes a strategy performance optimization term, a value function evaluation term, and a strategy entropy term. The strategy generation module is used to generate a target strategy for drone intrusion based on the first alternative strategy and the second alternative strategy, according to a preset strategy fusion rule.
[0020] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the anti-drone strategy generation method as described above.
[0021] Fourthly, this application provides a computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the anti-drone strategy generation method as described above.
[0022] Beneficial effects: A distributed, dynamically updated strategy memory is constructed to address the issues of "separation of learning and application" and the inability to retain knowledge. By storing and reusing effective tactical strategies from historical adversarial scenarios, continuous accumulation and sharing of decision-making knowledge across nodes is achieved. A decentralized storage architecture avoids single-point-of-failure risks, supports the elastic recovery and migration of tactical knowledge, and ensures the system can quickly invoke high-value strategies under similar threats. A multi-agent self-game mechanism is introduced to address the lack of "adversary modeling" and the difficulty in dealing with intelligent intruders. Agents autonomously evolve diverse and highly adaptable tactical behaviors through repeated games and strategy competition in a distributed simulated adversarial environment. A role rotation and co-evolution framework is designed to enable defenders to deeply understand the opponent's decision-making logic, promote the emergence of novel tactics, and significantly accelerate the tactical innovation process. A hybrid decision-making mechanism addresses the issues of "step-like" changes in tactical behavior and the lack of smooth transitions. Real-time tactical retrieval, optimization, and execution are achieved by dynamically adjusting the ratio of historical experience to online generation through confidence-adaptive weights. In familiar scenarios, experience is relied upon to ensure stability, while in novel scenarios, generation is relied upon to ensure creativity, thereby achieving smooth transitions and fine-tuning of tactical behavior. Attached Figure Description
[0023] Figure 1 This is a flowchart illustrating a method for generating anti-drone strategies according to the present invention. Figure 2 This is a flowchart illustrating another method for generating anti-drone strategies according to the present invention. Figure 3 This is a logical framework diagram of self-game training in an anti-drone strategy generation method of the present invention; Figure 4 This is a schematic diagram of the structure of an anti-drone strategy generation system according to the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device according to the present invention; Figure 6 This is a schematic diagram of the structure of a storage medium according to the present invention. Detailed Implementation
[0024] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0025] like Figures 1-2 As shown, a method for generating anti-drone strategies includes the following steps: S101. When a drone intrusion is detected, the current environmental state vector is obtained, and a preset number of preset strategies corresponding to the current environmental state vector are searched in a preset memory bank based on the current environmental state vector, so as to determine the first alternative strategy according to the preset strategy. For example, the pre-defined memory uses a distributed hash table (DHT) to organize policy knowledge, and each defense node stores a portion of the policy shards, achieving automatic load balancing and high system scalability. Through Locality Sensitive Hash (LSH) indexing and weighted similarity calculation, context-aware millisecond-level policy retrieval is supported, resulting in technical improvements in system response speed and knowledge reuse efficiency.
[0026] S102. Input the current environment state vector into the preset strategy generation model to obtain the second alternative strategy. The preset strategy generation model is a strategy generation model trained by self-game and set with a strategy optimization function. The strategy optimization function includes a strategy performance optimization term, a value function evaluation term, and a strategy entropy term. S103. Based on the first alternative strategy and the second alternative strategy, generate a target strategy for drone intrusion according to a preset strategy fusion rule.
[0027] In one possible implementation, before the step of searching a preset number of preset strategies corresponding to the current environment state vector in a preset memory based on the current environment state vector, the method further includes: Obtain historical strategies for completed adversarial tasks and execution data corresponding to those historical strategies; The comprehensive evaluation value of the historical strategy is calculated based on the execution data, and the comprehensive evaluation value is matched with a preset threshold to determine whether the historical strategy is effective. When the historical strategy is valid, the local sensitive hash index of the historical strategy is calculated to construct a preset memory.
[0028] For example, after a combat mission concludes, to quantitatively evaluate the effectiveness of a tactical strategy, the system calculates a comprehensive evaluation value for that tactic:
[0029] in and For configurable weights, To increase the success rate of interception, For energy efficiency (actual consumption / estimated consumption), The incidental damage index is 0-1. If... ( If a preset threshold (e.g., 0.75) is set, the strategy is deemed effective, stored in the strategy memory, and a Locality Sensitive Hash (LSH) index is created for it to support fast retrieval. The LSH hash function is defined as:
[0030] in A random Gaussian vector, obey Uniform distribution on Control the granularity of hash buckets.
[0031] In one possible implementation, the step of searching a preset number of preset strategies corresponding to the current environment state vector in a preset memory based on the current environment state vector, and determining a first candidate strategy according to the preset strategies, includes: By calculating the similarity data between each preset environment state vector in the preset memory and the current environment state vector, a preset number of preset environment state vectors are selected from the preset memory in descending order of similarity data as target environment state vectors. Based on the preset strategy corresponding to the target environment state vector, the first alternative strategy is obtained through a similarity-weighted fusion method.
[0032] For example, when the system detects a new intrusion target, it extracts the current environmental state vector. Retrieve from memory The most similar historical strategies. Similarity is determined using feature similarity data:
[0033] in This consists of learnable or manually defined feature weight vectors (e.g., position weight is higher than velocity). The final retrieval strategy is obtained through similarity-weighted fusion. .
[0034] In one possible implementation, before the step of inputting the current environment state vector into a preset strategy generation model to obtain a second alternative strategy, the method further includes: Construct a virtual simulation adversarial environment that includes an electromagnetic interference model, six-degree-of-freedom kinematic constraints, sensor noise, and communication delay; Define the state space, action space, and reward function. Through self-game training in the virtual simulation adversarial environment, train an initial model based on intelligent algorithms and used for policy generation, including a policy optimization function, to obtain a preset policy generation model.
[0035] For example, such as Figure 3 As shown, this module constructs virtual intelligent agents for both the attacker and defender in a high-fidelity simulation environment, and continuously evolves the defense strategy through self-play.
[0036] Specifically, a virtual simulation adversarial environment is constructed, incorporating an electromagnetic interference model, six-degree-of-freedom kinematic constraints, sensor noise, and communication delays. The state space is defined. Action space And design a reward function:
[0037] in This is an indicator function (1 is set to 1 if the interception is successful or collateral damage is caused, otherwise 0). This is a normalized value for resource utilization efficiency. These are the weighting coefficients.
[0038] During self-game training, to improve strategy generalization, the system periodically triggers a role rotation mechanism with probability:
[0039] A subset of agents are selected to swap offensive and defensive roles, with the swap ratio dynamically adjusted over training time. After the role swap, the agents inherit a subset of high-quality samples from the original role's experience replay pool, accelerating the learning of new role strategies. During training, the system dynamically adjusts the role rotation frequency based on the strategy diversity index to ensure both training stability and strategy diversity.
[0040] To address the technical challenges of predictable and undiversified policies in adversarial scenarios using traditional single-agent training methods, this invention employs a multi-agent collaborative optimization and update mechanism to simultaneously update the policy networks of both the defender and the intruder. This mechanism features three optimization objectives: a policy performance optimization term ensures stable improvement in interception capabilities through improved defense policies; a value function evaluation term reduces reward prediction errors and enhances policy evaluation accuracy; and a policy entropy maximization term encourages the exploration of novel tactics by increasing policy uncertainty. This collaborative optimization of the three objectives enables the system to maintain a high interception success rate while possessing generalization capabilities to handle unknown intrusion patterns, effectively addressing the problem of drastic performance degradation in traditional methods when facing adaptive adversaries.
[0041] The objective function for optimizing the policy network can be expressed as:
[0042] in As a policy performance optimization term, policy oscillations during training are avoided by limiting the policy update step size; This is a value function evaluation term used to improve the accuracy of long-term return predictions; This is the policy entropy term, which encourages the system to explore diverse tactics by increasing the uncertainty of policy output; and The balancing coefficients are dynamically adjusted based on the training phase. This objective function ensures that the policy maintains sufficient diversity during convergence, preventing it from prematurely falling into local optima.
[0043] In practical implementation, the above optimization objectives can be achieved using the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm, or alternatively, Trust Region Policy Optimization (TRPO), Multi-Agent Deep Deterministic Policy Gradient (MADDPG), or other multi-agent reinforcement learning algorithms can be selected, based on the system's computational resources and real-time requirements. Regardless of the specific algorithm used, the basic architecture of the triple optimization objective is maintained to ensure policy diversity and training stability.
[0044] In one possible implementation, the step of generating a target strategy for drone intrusion based on the first alternative strategy and the second alternative strategy according to a preset strategy fusion rule includes: According to the preset strategy fusion rules:
[0045] The first and second alternative strategies are merged to obtain a target strategy for drone intrusion, wherein... As the first alternative strategy, As a second alternative strategy, This is the current environment state vector. The confidence level is used for adaptive weights.
[0046] For example, define a fusion strategy:
[0047] in The strategy retrieved and fused from the strategy memory is calculated according to formula (4) in step 5.1.2. Generated in real time by the policy network. Given the current environmental state, For confidence-adaptive weights:
[0048] in, This is a confidence index; it is calculated based on the similarity-weighted average return of the k retrieved historical strategies. Representing historical state Compared with the current state similarity, For the first The return value of a historical strategy, Number of retrieval strategies (usually set to) When the scene is familiar, It mainly relies on experience-based strategies; when the scenario is novel, It mainly relies on online generation strategies.
[0049] In one possible implementation, the method further includes: The target strategy is evaluated based on predicted interception probability, energy consumption indicators, security indicators, and strategy flexibility indicators to obtain a comprehensive score for the target strategy. If the overall score is lower than the overall threshold, a better strategy is searched in the local neighborhood of the target strategy using a Bayesian optimization algorithm, and the better strategy is output.
[0050] For example, a multi-dimensional evaluation function is introduced to evaluate the target strategy based on predicted interception probability, energy consumption indicators, security indicators, and strategy flexibility indicators, thereby obtaining a comprehensive score for the target strategy. The multi-dimensional evaluation function is as follows:
[0051] The predicted interception probability is estimated by weighting historical similarity. To normalize energy consumption indicators to the range of 0-1, To take into account collateral damage and regulatory compliance in the safety score, To characterize the policy flexibility of the adjustable space. If ( If the threshold is set, a Bayesian optimization correction process is triggered in the local neighborhood. Better search strategies in China:
[0052] The optimization process is limited to a preset number of iterations, the specific number of which is dynamically determined based on the system's computing power and the urgency of the task. Decision generation meets real-time requirements, and the response time is adaptively adjusted according to the security level requirements of the application scenario to ensure that basic security functions are prioritized in emergency scenarios. In one possible implementation, the method further includes: Obtain the operational and attribute data of the anti-intrusion system, and calculate the formula based on the status:
[0053] Obtain the status score of the anti-intrusion system; wherein, As a normalized connectivity index, This represents the current number of connected nodes. This represents the total number of nodes in the system. For normalized delay indicators, The average communication delay (ms) This is the delay threshold; For normalized throughput, Current bandwidth (Mbps) Minimum required bandwidth; To prevent division by zero of small constants; Let be the weighting coefficient, satisfying ; when When the threshold value is less than a preset state threshold, the system controls each node within the anti-intrusion system to perform the following steps: when a drone intrusion is detected, obtain the current environmental state vector, and search a preset number of preset strategies corresponding to the current environmental state vector in a preset memory based on the current environmental state vector to determine a first candidate strategy; input the current environmental state vector into a preset strategy generation model to obtain a second candidate strategy, wherein the preset strategy generation model is a strategy generation model trained through self-game theory and equipped with a strategy optimization function, the strategy optimization function including a strategy performance optimization term, a value function evaluation term, and a strategy entropy term; and generate a target strategy for drone intrusion based on the first and second candidate strategies according to a preset strategy fusion rule.
[0054] For example, to address the technical problem of overall system collapse when nodes fail in traditional centralized systems, this invention designs a dynamic network health assessment mechanism that adaptively switches decision modes based on real-time communication status:
[0055] in As a normalized connectivity index, This represents the current number of connected nodes. This represents the total number of nodes in the system. For normalized delay indicators, The average communication delay (ms) This is the delay threshold; For normalized throughput, Current bandwidth (Mbps) Minimum required bandwidth; To prevent division by zero of small constants; Let be the weighting coefficient, satisfying .when ( When the threshold is set (to a preset value), the system automatically switches from centralized mode to distributed mode, with each node making independent decisions based on its local memory and local policies.
[0056] In one possible implementation, when a new node joins or an old node fails in the anti-intrusion system, tactical knowledge transfer is achieved through lightweight knowledge distillation. The distillation loss function is:
[0057] in Output for the teacher model (surviving nodes), Output for the student model (new node). This is the logits output of the corresponding model (the unnormalized raw predictions). Temperature parameter controls the smoothness of the distribution; The Kullback-Leibler divergence measures the difference between two distributions. This is achieved by minimizing... New nodes can quickly replicate mature tactical capabilities.
[0058] By constructing a distributed policy memory, the system achieves long-term accumulation and efficient reuse of adversarial experience, effectively overcoming the shortcomings of traditional methods that involve "one-time decisions and inability to accumulate knowledge." When facing similar threats, the system can quickly invoke historical high-value policies, significantly improving response speed and decision-making quality, while supporting cross-node knowledge sharing and elastic recovery. Through role rotation and collaborative training, the defensive agent can not only learn its own optimal behavior but also deeply understand the decision-making logic of potential adversaries, thereby generating more forward-looking and adversarial tactical solutions. This mechanism effectively enhances the system's ability to respond to highly intelligent and adaptive intrusion targets. By adaptively fusing retrieval and generation strategies with confidence levels, the system retains the ability to explore unknown scenarios while ensuring tactical stability. Multi-dimensional evaluation and dynamic correction mechanisms further ensure the comprehensive optimization of tactics in terms of interception effectiveness, resource consumption, and security constraints. The elastic collaborative control system gives the system good environmental adaptability and resilience. In the event of communication limitations or node anomalies, the system can automatically switch to a distributed operation mode and quickly rebuild local tactical capabilities through lightweight knowledge transfer, ensuring that the overall defense effectiveness does not significantly decrease due to local failures.
[0059] In one possible implementation, such as Figure 4 As shown, this application provides an anti-drone strategy generation system, including: The first calculation module 201 is used to obtain the current environmental state vector when the drone intrusion is detected, and search for a preset number of preset strategies corresponding to the current environmental state vector in a preset memory bank based on the current environmental state vector, so as to determine the first alternative strategy according to the preset strategy. The second calculation module 202 is used to input the current environment state vector into a preset strategy generation model to obtain a second alternative strategy. The preset strategy generation model is a strategy generation model that has been trained through self-game and has a strategy optimization function. The strategy optimization function includes a strategy performance optimization term, a value function evaluation term, and a strategy entropy term. The strategy generation module 203 is used to generate a target strategy for drone intrusion based on the first alternative strategy and the second alternative strategy, according to a preset strategy fusion rule.
[0060] In one possible implementation, such as Figure 5 As shown, this application embodiment provides an electronic device 300, including: a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, it performs the following: when a drone intrusion is detected, it acquires the current environmental state vector, and searches a preset number of preset strategies corresponding to the current environmental state vector in a preset memory bank based on the current environmental state vector, so as to determine a first candidate strategy according to the preset strategy; it inputs the current environmental state vector into a preset strategy generation model to obtain a second candidate strategy, wherein the preset strategy generation model is a strategy generation model trained by self-game and set with a strategy optimization function, wherein the strategy optimization function includes a strategy performance optimization term, a value function evaluation term, and a strategy entropy term; and it generates a target strategy for drone intrusion based on the first candidate strategy and the second candidate strategy according to a preset strategy fusion rule.
[0061] In one possible implementation, such as Figure 6 As shown, this application embodiment provides a computer-readable storage medium 400 storing a computer program 411. When executed by a processor, the computer program 411 performs the following: when a drone intrusion is detected, it acquires a current environmental state vector and searches a preset number of preset strategies corresponding to the current environmental state vector in a preset memory based on the current environmental state vector, so as to determine a first candidate strategy according to the preset strategies; it inputs the current environmental state vector into a preset strategy generation model to obtain a second candidate strategy, wherein the preset strategy generation model is a strategy generation model trained by self-game and has a strategy optimization function, the strategy optimization function including a strategy performance optimization term, a value function evaluation term, and a strategy entropy term; based on the first candidate strategy and the second candidate strategy, it generates a target strategy for the drone intrusion according to a preset strategy fusion rule.
[0062] The present invention has been described above through a preferred embodiment. This solution can also be implemented by means of an apparatus or device to perform the methods described in the above embodiments.
[0063] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this solution includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which the embodiments of this solution pertain. The processor performs the various methods and processes described above. For example, the method embodiments of this solution can be implemented as software programs tangibly contained in a machine-readable medium, such as memory. In some embodiments, part or all of the software program can be loaded and / or installed via memory and / or a communication interface. When the software program is loaded into memory and executed by the processor, one or more steps of the methods described above can be performed. Alternatively, in other embodiments, the processor can be configured to perform one of the methods described above by any other suitable means (e.g., by means of firmware).
[0064] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for generating anti-drone strategies, characterized in that, Includes the following steps: When a drone intrusion is detected, the current environmental state vector is obtained, and a preset number of preset strategies corresponding to the current environmental state vector are searched in a preset memory based on the current environmental state vector, so as to determine the first alternative strategy according to the preset strategy. The current environment state vector is input into a preset strategy generation model to obtain a second alternative strategy. The preset strategy generation model is a strategy generation model trained by self-game and equipped with a strategy optimization function. The strategy optimization function includes a strategy performance optimization term, a value function evaluation term, and a strategy entropy term. Based on the first alternative strategy and the second alternative strategy, a target strategy for drone intrusion is generated according to a preset strategy fusion rule.
2. The method for generating anti-drone strategies according to claim 1, characterized in that, Before the step of searching a preset number of preset strategies corresponding to the current environment state vector in a preset memory based on the current environment state vector, the method further includes: Obtain historical strategies for completed adversarial tasks and execution data corresponding to those historical strategies; The comprehensive evaluation value of the historical strategy is calculated based on the execution data, and the comprehensive evaluation value is matched with a preset threshold to determine whether the historical strategy is effective. When the historical strategy is valid, the local sensitive hash index of the historical strategy is calculated to construct a preset memory.
3. The method for generating anti-drone strategies according to claim 1, characterized in that, The step of searching a preset number of preset strategies corresponding to the current environment state vector in a preset memory based on the current environment state vector, and determining a first candidate strategy according to the preset strategies, includes: By calculating the similarity data between each preset environment state vector in the preset memory and the current environment state vector, a preset number of preset environment state vectors are selected from the preset memory in descending order of similarity data as target environment state vectors. Based on the preset strategy corresponding to the target environment state vector, the first alternative strategy is obtained through a similarity-weighted fusion method.
4. The method for generating anti-drone strategies according to claim 1, characterized in that, Before the step of inputting the current environment state vector into the preset strategy generation model to obtain the second alternative strategy, the method further includes: Construct a virtual simulation adversarial environment that includes an electromagnetic interference model, six-degree-of-freedom kinematic constraints, sensor noise, and communication delay; Define the state space, action space, and reward function. Through self-game training in the virtual simulation adversarial environment, train an initial model based on intelligent algorithms and used for policy generation, including a policy optimization function, to obtain a preset policy generation model.
5. The method for generating anti-drone strategies according to claim 1, characterized in that, The step of generating a target strategy for drone intrusion based on the first alternative strategy and the second alternative strategy according to a preset strategy fusion rule includes: According to the preset strategy fusion rules: The first and second alternative strategies are merged to obtain a target strategy for drone intrusion, wherein... As the first alternative strategy, As a second alternative strategy, This is the current environment state vector. The confidence level is used for adaptive weights.
6. The method for generating an anti-drone strategy according to claim 1, characterized in that, The method further includes: The target strategy is evaluated based on predicted interception probability, energy consumption indicators, security indicators, and strategy flexibility indicators to obtain a comprehensive score for the target strategy. If the overall score is lower than the overall threshold, a better strategy is searched in the local neighborhood of the target strategy using a Bayesian optimization algorithm, and the better strategy is output.
7. The method for generating anti-drone strategies according to claim 1, characterized in that, The method further includes: Obtain the operational and attribute data of the anti-intrusion system, and calculate the formula based on the status: Obtain the status score of the anti-intrusion system; wherein, As a normalized connectivity index, This represents the current number of connected nodes. This represents the total number of nodes in the system. For normalized delay indicators, The average communication delay (ms) This is the delay threshold; For normalized throughput, Current bandwidth (Mbps) Minimum required bandwidth; To prevent division by zero of small constants; For the weighting coefficients, satisfying ; when When the threshold value is less than a preset state threshold, the system controls each node within the anti-intrusion system to perform the following steps: when a drone intrusion is detected, obtain the current environmental state vector, and search a preset number of preset strategies corresponding to the current environmental state vector in a preset memory based on the current environmental state vector to determine a first candidate strategy; input the current environmental state vector into a preset strategy generation model to obtain a second candidate strategy, wherein the preset strategy generation model is a strategy generation model trained through self-game theory and equipped with a strategy optimization function, the strategy optimization function including a strategy performance optimization term, a value function evaluation term, and a strategy entropy term; and generate a target strategy for drone intrusion based on the first and second candidate strategies according to a preset strategy fusion rule.
8. A counter-drone strategy generation system, characterized in that, include: The first calculation module is used to obtain the current environmental state vector when the drone intrusion is detected, and search for a preset number of preset strategies corresponding to the current environmental state vector in a preset memory bank based on the current environmental state vector, so as to determine the first alternative strategy according to the preset strategy. The second calculation module is used to input the current environment state vector into the preset strategy generation model to obtain the second alternative strategy. The preset strategy generation model is a strategy generation model that has been trained through self-game and has a strategy optimization function. The strategy optimization function includes a strategy performance optimization term, a value function evaluation term, and a strategy entropy term. The strategy generation module is used to generate a target strategy for drone intrusion based on the first alternative strategy and the second alternative strategy, according to a preset strategy fusion rule.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the anti-drone strategy generation method as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the anti-drone strategy generation method as described in any one of claims 1 to 7.