An intelligent defense control system and method for multi-target multi-batch multi-parameter dynamic scene

Abstract

The application provides a kind of intelligent defense control system and method for multi-target multi-batch multi-parameter dynamic scene, the system includes: scene parameter and three-dimensional construction module;Multi-target protection object management module and other modules;The method includes step 1, constructs scene parameter and generates three-dimensional defense task;Step 2, generate incoming object;Step 3, obtain state observation and decision defense action and other steps.The effect of the application is: the system organizes the number of protection targets, target type, spatial layout and enable state as a set of scene configuration parameters, processes different scale scenes using a fixed upper limit and dynamic enablement, allowing multiple protection targets to share the same scene interface and the same observation structure;And incoming objects can be generated according to batch number, generation time, incoming direction, speed range, height range and attack target number, allowing different batches to enter continuously and overlap each other in the same round, more realistically expressing sustained pressure and multi-directional attack posture.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, 3D simulation, dynamic scene modeling, reinforcement learning and intelligent control technology, and in particular relates to an intelligent defense control system for multi-target, multi-batch and multi-parameter condition changes. Background Technology

[0002] Intelligent defense control in complex and dynamic environments typically requires real-time decision-making under various initial conditions, target behavior patterns, and interactions. Unlike static rule matching tasks, these tasks are characterized by changes in the number of targets, spatial relationships, incoming batches, and continuous changes in motion states. Protection strategies not only need to identify the current threat but also select appropriate response times within a limited timeframe and maintain stable action output amidst continuous situational changes. Traditional rule-based defense methods generally rely on manually set target numbers, distance thresholds, direction judgments, or fixed priorities. While these methods are easy to implement in small-scale, fixed scenarios, their rule coverage is significantly limited when the number of protected targets, incoming direction, target speed, altitude, batch rhythm, and trajectory disturbances change. Repeated manual adjustments are required, making it difficult to develop stable decision-making capabilities for unknown scenarios.

[0003] Reinforcement learning can learn sequential decision-making strategies through continuous interaction between an agent and its environment, making it suitable for handling problems involving continuous state inputs, discrete action selection, and long-term reward optimization. However, if the training environment contains only a single scene or a few fixed rules, the model is prone to becoming dependent on specific initial positions, target numbers, and attack rhythms, leading to decreased success rates, action fluctuations, and insufficient generalization ability under new scene parameter combinations. Furthermore, existing simulation systems and algorithm training processes are often disconnected: the 3D scene focuses on visualization, the training end emphasizes simplified interfaces, and the testing phase uses a different set of decision logic, resulting in poor model training results. For intelligent defense and control tasks, it is even more necessary to unify the organization of scene generation, environment development, state observation, action execution, reward feedback, policy training, and model testing, so that the algorithm training process directly originates from a reproducible 3D dynamic scene. Therefore, there is an urgent need for an intelligent defense and control technology solution that can form a unified process across four levels: scene construction, task modeling, algorithm training, and generalization evaluation, enabling defense strategies to learn reusable target recognition, response timing selection, and defense action decision-making capabilities in multi-condition dynamic environments.

[0004] To address the shortcomings of existing intelligent defense control methods in terms of scenario representation, model training, and cross-condition adaptation, this invention aims to provide an intelligent defense control system and method for multi-target, multi-batch, and multi-parameter dynamic scenarios. This system does not design rules for a single, fixed scenario, but rather unifies the 3D simulation environment, dynamic target model, reinforcement learning training process, and generalization evaluation mechanism into an executable technical chain. To achieve the above objective, this invention mainly solves the following technical problems.

[0005] (1) The problem of parameterization construction of multi-condition three-dimensional defense scenarios: In intelligent defense tasks, the number of protected targets, spatial layout, number of incoming objects, speed, height, direction and batch rhythm may all change. How to uniformly express these changing factors in the same simulation framework and enable scenarios of different scales to share state interfaces and training processes is the first problem that this invention needs to solve;

[0006] (2) The modeling problem of dynamic incoming objects and the interaction process between defense and attack. The incoming objects are not static objects, but move continuously in three-dimensional space and may be affected by disturbances or defensive actions, which may change their trajectory. How to establish a dynamic evolution model that takes into account the direction of movement, historical trajectory, batch generation and disturbance response, so that the scene can provide real, continuous and verifiable interaction samples for policy learning;

[0007] (3) The problem of intelligent decision training for multi-target protection: the protection unit needs to select defense actions according to the local situation, and the algorithm training needs to continuously obtain the state, action, reward and round results from the scene. How to transform the objects in the three-dimensional scene into an observation, action and reward structure that can be processed by reinforcement learning, and realize the unified training of multiple protection units through shared strategies, is the third problem that this invention needs to solve.

[0008] (4) Cross-scene generalization problem: Fixed scene training makes the model dependent on specific numbers, positions or attack rhythms, and it is difficult to adapt to unknown parameter combinations. How to build a scene distribution covering multiple conditions during the training stage, and strengthen the sampling of weak scenes according to the model performance, so that the trained model can complete any task under fixed rules and diversified rules. Summary of the Invention

[0009] To overcome the defects or problems of existing technologies, this invention provides an intelligent defense control system for multi-target, multi-batch, and multi-parameter dynamic scenarios. The system is based on a parameterized three-dimensional scene, uses multiple batches of incoming targets as dynamic interaction objects, uses state, action, and reward interfaces as decision inputs and outputs, supports control parameter learning with a shared policy network and reinforcement learning update process, and improves the system's adaptability to multi-condition dynamic scenarios through feedback from difficult scenarios.

[0010] The intelligent defense control system for multi-target, multi-batch, and multi-parameter dynamic scenarios includes:

[0011] Scene Parameters and 3D Construction Module: This module reads or randomly generates a set of scene configuration parameters and creates protected areas, protected targets, protective units, and round control objects based on these parameters. This module encapsulates the number of protected targets, spatial layout, batch size, batch interval, velocity range, altitude range, incoming direction set, disturbance range, and round duration in a unified manner.

[0012] Multi-target protection object management module: The multi-target protection object management module generates a set of protected targets and a set of protected units based on scene parameters and the enable flag sent by the 3D construction module. For each protected target, it records the target number, target type, 3D coordinates, health status, hit status and corresponding protected unit, and sends the status of the protected target and the status of the protected unit to the status, action and reward interface module.

[0013] Incoming Target Control Module: This module generates incoming targets based on the batch plan, writing a batch number, target number, initial position, velocity, altitude, direction of motion, and disturbance flag for each target. It updates the target's current position, historical position, and trajectory deflection status at each time step, and changes the target's subsequent direction of motion or target point upon receiving valid interference information.

[0014] Status, Action and Reward Interface Module: The status, action and reward interface module uses the protection unit as a reference to calculate the relative position, relative distance, current movement direction, historical movement direction, batch information and resource status of candidate targets, combines the above information into status observation, and can receive action execution results, and generate feedback signals based on target matching, effective interference, ineffective actions, resource consumption and defense results;

[0015] The intelligent defense decision-making module and strategy training module receive state observations and output target selection actions, auxiliary resource actions, and action maintenance information. In training mode, a shared policy network is used to process local decision samples from multiple defense units, and the policy parameters are updated through a near-end policy optimization algorithm. In test or deployment mode, existing policy parameters are loaded and inference is performed.

[0016] Win rate testing and scene update module: The win rate testing and scene update module judges the round result based on the protected target's hit status, the incoming object's miss status, the batch completion status, and the round time, and records scene parameters, interaction trajectories, action events, and feedback signals. It feeds back the failed scene and difficult parameter combination to the scene parameter and 3D construction module.

[0017] The present invention further provides a control method for the intelligent defense control system for multi-target, multi-batch, and multi-parameter dynamic scenarios, the intelligent defense control method for multi-target, multi-batch, and multi-parameter dynamic scenarios comprising the following steps:

[0018] Step 1: Construct scene parameters and generate a 3D defense task;

[0019] Step 2: Generate the incoming attack object;

[0020] Step 3: Obtain state observations and decide on defensive actions;

[0021] Step 4: Execute defensive actions and complete the environmental state transition;

[0022] Step 5: Determine the round status and use PPO to collect experience for training;

[0023] Step 6: Win rate test, scene parameters update and repeat training.

[0024] Furthermore, step 1, which involves constructing scene parameters and generating a 3D defense task, specifically includes:

[0025] Step 1.1 The scene parameters and 3D construction module read the configuration results generated by the task configuration file or random sampler, and organize the number of protected targets, the type of protected targets, the spatial layout, the target generation area, the number of incoming batches, the batch interval, the velocity range, the height range, the set of incoming directions, the disturbance range, and the maximum round duration into a scene configuration parameter set. As shown in equation (1):

[0026] ,

[0027] in, Indicates the number of protected targets activated in the current round. Represents the set of protected target spaces. Indicates the region where the incoming object is generated. Indicates the number of incoming waves. Indicates the target quantity for a single batch or the range of target quantities. Indicates the batch interval. Indicates the speed range. Indicates the height range. Indicate the direction of the attack to gather. Indicates the range of the disturbance. Indicates the maximum round duration;

[0028] Step 1.2 Scene Parameters and 3D Construction Module: Configure Parameter Set Based on Scene. Create protected areas, protected target locations, incoming object generation areas, round control objects, and defense task instances; send the protected target configuration to the multi-target protected object management module; and send the incoming object generation parameters to the incoming object control module.

[0029] Step 1.3 The multi-target protection object management module, based on... and Generate a set of protected targets for the current round, and record the target number, target type, three-dimensional coordinates, activation status, health status and hit status for each activated protected target. To ensure that different target sizes share the input interface, the system reserves a fixed upper limit for the number of protected targets, and controls the actual protected targets participating in the interaction in the current round through the activation flag, as shown in the following formula (2):

[0030] ,

[0031] in, Indicates the first Record the location or status of the protected target. This indicates the maximum number of target slots reserved by the system. The activation flag for unprotected targets is set to 0, their observations are empty, and they do not participate in motion, reward, or round determination.

[0032] Step 1.4 The multi-target protection object management module configures a corresponding protection unit for each enabled protection target, and sets the number of protection units accordingly. The first protection unit and the first A connection is established between each protected target. This connection is used to determine whether an incoming target is directed at the protected target that the current protection unit is responsible for, and serves as the basis for judging the effectiveness of the defense action.

[0033] Step 1.5 Scene parameters and 3D construction module: Before the start of each training or test round, resample the number of protected targets, spatial layout, batch rhythm, speed and height, incoming direction and disturbance range to generate new defense task instances, so that subsequent strategy training does not depend on a single target number, a single spatial layout or a single attack rhythm.

[0034] Furthermore, step 2, generating the incoming target, specifically includes:

[0035] Step 2.1 The incoming object control module receives the scene configuration parameter set. And based on the number of incoming batches Batch interval and single batch target quantity Generate batch plan As shown in equation (3):

[0036]

[0037] in, Indicates the batch number. Indicates the first Batch target generation time. Indicates the first Batch target quantity, This indicates that the incoming direction of this batch can be selected. This indicates the speed range for this batch. Indicates the height range for this batch;

[0038] Step 2.2 The incoming target control module reaches the current time. When creating the first The system processes incoming targets in batches and writes the target number, batch number, attack target number, initial position, speed, altitude, direction of attack, and disturbance flag for each incoming target. The attack target number is determined or randomly generated according to rules such as uniform dispersion, random allocation, priority of nearest target, or intersection of far ends.

[0039] Step 2.3 The incoming target control module records the dynamic status of each incoming target, including the current position, the position at the previous moment, the attack target, the speed, the altitude, the batch to which it belongs, the disturbance flag, the survival status, the out-of-target status, and the interfered status, and sends it to the status, action and reward interface module after updating at each time step;

[0040] Step 2.4 The incoming target control module determines the current position of the incoming target. and the corresponding protected target location Calculate the ideal direction of motion As shown in equation (4):

[0041] ,

[0042] in, Indicates the first The attack target number corresponding to each incoming target. Indicates the location of the protected target corresponding to the attack target;

[0043] Step 2.5, to simulate uncertainties in a dynamic scenario, involves the incoming object control module controlling the disturbance term. Incorporate the ideal direction of motion and follow the target speed. Update the position at the next moment and determine the actual direction of motion after the perturbation is applied. :

[0044]

[0045] in, Indicates direction normalization processing. This indicates the actual direction of motion after the perturbation is applied. This represents a single simulation time step, in which the incoming object exhibits continuous variations in velocity, altitude, orientation, and trajectory deviation.

[0046] Step 2.6 When the previous batch of incoming targets has not yet finished, the incoming target control module still generates the next batch of targets according to the batch plan, so that different batches of incoming targets can coexist in the same time window, forming a continuous pressure, batch superposition and dynamic change defense posture.

[0047] Furthermore, step 3, which involves acquiring state observations and deciding on defensive actions, specifically includes:

[0048] Step 3.1 The Status, Action, and Reward Interface Module receives the status of the protected target, the status of the protected unit, and the dynamic status of the incoming object, and uses the first... Each protection unit serves as a reference to construct candidate target state features. These features extract the relative positions between the protection unit and the candidate incoming targets, the relative distances between them, the current direction of movement, historical direction changes, batch numbers, and target states. The first protection unit and the first State characteristics among candidate incoming targets As shown in equation (6):

[0049]

[0050] in, Indicates relative position, Indicates a change in the direction of history. Indicates relative distance. Indicates the batch number. This indicates whether the target is alive, has missed the target, or has been interfered with.

[0051] Step 3.2 The State, Action, and Reward interface module sorts the candidate target relationship features according to distance, estimated arrival time, and target state, and selects the top... There are 10 candidate targets, which is insufficient. The candidate slots are filled with empty slots, allowing different numbers of incoming objects and different batches to share the same observation dimension. Each protective unit is in constant motion State observation As shown in equation (7):

[0052]

[0053] in, This represents the set of candidate target relation features after sorting and filling. This indicates the resource status, cooling status, and operational status of the protection unit. Indicates the progress of the current round. Indicates splicing;

[0054] Step 3.3 The intelligent defense decision-making module and strategy training module receive state observations. The policy network outputs defensive actions. Defensive actions This includes candidate target selection, whether to enable or disable auxiliary resources, whether to maintain or not execute actions, and defensive actions. As shown in equation (8):

[0055] ,

[0056] in, The parameter is During the training phase, the policy network updates itself based on the collected round experience; during the testing phase, the policy network loads the trained parameters and performs action reasoning only.

[0057] Step 4, which involves performing defensive actions and completing the environmental state transition, specifically includes:

[0058] Step 4.1 The intelligent defense decision-making module and strategy training module will execute defense actions. The action command is sent to the status, action and reward interface module, which translates the action command into defensive behavior in the 3D scene. The defensive behavior in the 3D scene includes target selection, interference triggering, auxiliary protection activation and action maintenance.

[0059] Step 4.2: The status, action, and reward interface module determines the validity of an action based on the association between the protection unit and the protected target, the attack target number of the incoming object, distance conditions, resource status, and cooldown status. When the selected incoming object points to the protected object under the responsibility of the current protection unit and meets the resource, distance, or cooldown conditions, the system generates a valid interference event. When the action target does not match or the resource conditions are not met, the system records an erroneous action event.

[0060] 4.3 When the action is effective, the incoming target control module modifies the subsequent movement direction or target point of the incoming target according to the effective interference event, causing the target trajectory to deflect; when the action is invalid or not executed, the incoming target continues to fly normally according to the above formulas (4) and (5). For the interfered target, the system estimates the miss distance based on the location of the interference, the timing of the response, and the current direction of the target. As shown in equation (9):

[0061] ,

[0062] in, This indicates the distance between the incoming target and the corresponding protected target when the interference occurs. This represents the maximum effective distance used for normalization. This means randomly sampling the deflection amplitude within a given range. In this way, defensive actions can change the subsequent motion state of the target, causing the environment to transition from the current state to the next state.

[0063] Step 4.4 The status, action, and reward interface module generates a single-step feedback signal based on the validity of the current action, the target deflection status, the protected target's hit status, erroneous actions, and resource consumption. Single-step feedback signal The policy is used for policy updates during the training phase and for process recording during the testing phase, as shown in equation (10) below:

[0064]

[0065] in, This indicates feedback on the final or interim defense results. This indicates the feedback on target matching and effective identification. This indicates interference deflection or off-target feedback. This indicates the cost of incorrect actions, resource consumption, or frequent switching. , and The weighting coefficients are used to retain feedback on the success or failure of defense, while also guiding the model to learn appropriate target selection and response timing through process feedback.

[0066] Furthermore, step 5, which involves determining the round state and using PPO to collect experience for training, specifically includes:

[0067] Step 5.1 The win rate test and scenario update module determines whether the current round has ended based on the target being hit, the incoming target missing the target, the batch completion status, and the round time. If there are still unfinished batches or incoming targets that have not been terminated, the system continues to execute steps 3 to 4. If all batches have ended, the target being hit, or the maximum round time has been reached, the current round ends and enters the experience processing process.

[0068] Step 5.2 In training mode, the State, Action, and Reward interface module writes the state observation, action, feedback signal, next time step state observation, round end flag, and scene configuration parameter set of each time step into the sample set. Sample set It can be expressed as the following formula (11):

[0069] ,

[0070] in, Indicates the end of the round. This represents the set of scenario configuration parameters corresponding to the source of the samples, and the sample index satisfies... and Sample set It is used for both strategy updates and failure scenario analysis and difficult parameter combination statistics, where r and a represent rewards and actions;

[0071] Step 5.3 The intelligent defense decision-making module and the strategy training module use a shared policy network to process the local decision sequences of multiple defense units. Although the multiple defense units correspond to different protection targets and different local situations, they use the same observation structure and action structure. Therefore, their interaction experience can be used to update the same policy network. The goal of policy training is to learn the action selection rules on the scene distribution to maximize the benefits. :

[0072] ,

[0073] in, This represents the scene sampling distribution during the training phase. Representation Strategy In the scene configuration parameter set Expected return in the corresponding round;

[0074] Step 5.4 In training the neural network of the intelligent defense decision module, the policy update adopts the near-end policy optimization algorithm, given a sample set collected from multiple scenarios and multiple protection units. The intelligent defense decision-making module and the strategy training module update the policy network based on advantage estimation, and limit the magnitude of each update by pruning terms. Loss represents the amount of loss.

[0075] ,

[0076] in, Indicates the policy network parameters, This represents the ratio of the probability of the new strategy to the probability of the old strategy. Indicates the advantage estimate, The above update process, which represents the pruning threshold, enables the strategy to gradually learn the patterns of target matching, response selection, and resource usage under conditions of multiple targets, multiple locations, multiple batches, and multiple perturbations.

[0077] Step 5.5 After completing multiple rounds of strategy updates, the intelligent defense decision module and the strategy training module save the current strategy parameters and send the round gains, success flags, failure reasons and action records generated during the training process to the win rate test and scenario update module for subsequent testing and scenario parameter adjustment.

[0078] Furthermore, step 6, which involves win rate testing, scene parameter updating, and repeated training, specifically includes:

[0079] Step 6.1 The win rate test and scenario update module calls the saved strategy parameters in test mode and stops the strategy update. During the test phase, the scenario generation, incoming object control, state observation, action execution, environment transition and round determination logic of steps 1 to 5 are still running to ensure that the training process and the testing process use the same set of interactions.

[0080] Step 6.2 The win rate test and scenario update module statistically analyzes the model performance in fixed rule scenarios and diverse random scenarios, including defense win rate, average feedback value, number of effective interferences, number of erroneous actions, and number of times the protected target is attacked. Fixed rule scenarios are used to test whether the model has mastered the clear attack rules, while diverse random scenarios are used to test the model's adaptability to unknown parameter combinations.

[0081] Step 6.3 The win rate testing and scenario update module calculates recent success rates based on the combination of scenario parameters. Average round gains Failure frequency And feedback on fluctuations, and define the difficulty scenario score accordingly. As shown in equation (14):

[0082]

[0083] in, , and These are the weighting coefficients. This indicates normalization. A higher difficulty score indicates that the current strategy is more unstable under this parameter combination;

[0084] Step 6.4 Scene Parameters and 3D Construction Module: Based on the difficult scene records reported by the win rate test and scene update module, adjust the scene sampling distribution in subsequent training. As shown in equation (15):

[0085]

[0086] in, Represents a set of optional scenario parameter combinations. Control the ratio of uniform sampling to sampling in difficult scenarios. Using temperature coefficients, this scene update mechanism maintains overall scene coverage while increasing the frequency of parameter combinations that have high recent failure rates, large reward fluctuations, or difficulties in target recognition.

[0087] Step 6.5 The system re-executes steps 1 to 6 based on the updated scene sampling distribution, forming a cyclical process of scene parameter generation, defense task construction, incoming target control, state observation, defense action decision-making, environmental state transition, round experience training, win rate testing, and scene parameter update. Through continuous repeated training, the intelligent defense strategy can continuously engage with vulnerable scenes that have not yet been fully adapted to, and gradually improve the defense win rate under dynamic conditions of multiple targets, multiple batches, and multiple parameters.

[0088] The intelligent defense control system and method for multi-target, multi-batch, and multi-parameter dynamic scenarios described in this invention have the following superior technical effects:

[0089] 1. The present invention provides an intelligent defense control system and method for multi-target, multi-batch, and multi-parameter dynamic scenarios, which has the ability to uniformly construct multi-target dynamic scenarios. The system organizes the number of protected targets, target types, spatial layout, and activation status into a scenario configuration parameter set, and uses a fixed upper limit and dynamic activation method to handle scenarios of different scales, so that multiple protected targets can share the same scenario interface and the same observation structure.

[0090] 2. The intelligent defense control system and method for multi-target, multi-batch, and multi-parameter dynamic scenarios described in this invention have the ability to continuously generate multiple batches of incoming targets. The system generates incoming targets according to batch number, generation time, direction of attack, speed range, height range, and target number, so that different batches can enter continuously and overlap with each other in the same round, which can more realistically express continuous pressure and multi-directional attack situation.

[0091] 3. The intelligent defense control system and method for multi-target, multi-batch, and multi-parameter dynamic scenarios described in this invention have multi-parameter randomness and dynamic interaction capabilities. The system samples the number of targets, spatial location, batch rhythm, speed and altitude, direction of attack, and disturbance range, and updates the target trajectory, so that scenario changes do not depend on manual rule-by-rule processing, and provides diverse samples for strategy training from the source.

[0092] 4. The intelligent defense control system and method for multi-target, multi-batch, and multi-parameter dynamic scenarios described in this invention have the ability to express status, actions, and feedback. The system extracts the relative position, distance, current direction, historical direction, and batch status of candidate targets with the protection unit as a reference, and generates feedback signals based on target matching, interference results, erroneous actions, and resource consumption, which can reduce the dependence of the strategy on absolute numbering and fixed position.

[0093] 5. The intelligent defense control system and method for multi-target, multi-batch, and multi-parameter dynamic scenarios described in this invention have the ability to perform scenario-based generalized training and automatic strategy learning. The system directly draws training samples from multi-target, multi-batch, and multi-parameter dynamic scenarios, learns the rules for selecting defense actions through a shared policy network, and increases the sampling probability of weak parameter combinations based on recent success rate, average benefit, and failure frequency, so that the strategy has better adaptability under unknown target size, unknown batch rhythm, and unknown disturbance conditions. Attached Figure Description

[0094] Figure 1 This is a schematic diagram of Unity 3D scene construction in a specific embodiment of the present invention;

[0095] Figure 2 This is a schematic diagram of a typical protected target scene in a specific embodiment of the present invention;

[0096] Figure 3 This is a schematic diagram of the generation of the incoming target and the interference ray in a specific embodiment of the present invention;

[0097] Figure 4 This is a schematic diagram illustrating the explosion effect of a failed round defense in a specific embodiment of the present invention;

[0098] Figure 5 This is a schematic diagram of the scenario setup and training process in a specific implementation method;

[0099] Figure 6 This is a schematic diagram of the intelligent defense control method for multi-target, multi-batch, and multi-parameter dynamic scenarios described in this invention. Detailed Implementation

[0100] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can make equivalent substitutions for scene objects, parameter ranges, target types, training algorithms, 3D simulation platforms or system deployment methods; where there is no conflict, the technical features of each embodiment can be combined with each other.

[0101] Example 1

[0102] In one specific embodiment, the present invention can be executed by an electronic device, which includes a processor, a memory, an input / output interface, and a graphics rendering environment. The memory stores a scene parameter generation program, a multi-target protection object management program, an incoming object control program, a state observation program, a defense action decision-making program, a reward calculation program, a strategy training program, a win rate testing program, and a scene parameter update program. When the processor executes the above programs, it implements the intelligent defense control system described in this invention. The electronic device is a standalone workstation, a server, or a cloud training node. The 3D scene is constructed using the Unity environment, and the training interface can connect to a Python training client via ML-Agents.

[0103] This embodiment uses the airport area defense task in the Unity system as an example to illustrate the usage process of the present invention.

[0104] like Figure 1 As shown, Figure 1 The diagram illustrates the construction of a Unity 3D scene. Scene parameters and 3D building modules create a 3D scene in the Unity system, generating scene objects such as roads, runways, aprons, buildings, and oil depots, and creating turn control objects and attack object generation areas.

[0105] like Figure 2 As shown, Figure 2 The typical 3D scene of protected targets is shown. The multi-target protection object management module enables five types of protected targets in the current round: tower, airport, warehouse, building and oil depot. It also configures a corresponding protection unit for each protected target. Unused protected targets remain in an empty state and do not participate in state observation, action execution and round determination.

[0106] The scenario parameters set in this embodiment are as follows: the number of protected targets N in the current round is 5, and the maximum number of protected targets reserved by the system is 10; the types of protected targets include control towers, airports, warehouses, buildings, and oil depots; the incoming target generation area is set within 10 km to 20 km outside the protected area; the number of incoming batches K is 3, and the batch generation times are 0 s, 10 s, and 20 s respectively; the altitude range of incoming targets is 3000 m to 6000 m, the speed range is 0.75 Ma to 2.0 Ma, and the incoming direction set includes north, east, and random directions; the trajectory disturbance angle is set within the range of -3° to 3°; the maximum time per round is set to 120 s, and each auxiliary protection unit can be activated a maximum of 1 time per round;

[0107] Figure 3The diagram illustrates the generation and interference rays for incoming targets. The incoming target control module generates incoming targets based on the parameters described above. The first batch of targets is assigned according to the cross-far end rule; the second batch is assigned according to the nearest target priority rule; and the third batch is assigned randomly. After each incoming target is generated, its target number, batch number, attack target number, initial position, velocity, altitude, current direction, previous position, survival status, miss status, and interference status are recorded. The system updates the incoming target position at each time step and sends the target status to the status, action, and reward interface module.

[0108] The status, action, and reward interface module uses the protection unit as a reference to extract the relative position, relative distance, direction of movement, historical direction changes, batch number, and target status of candidate incoming targets. It then selects the top few candidate targets based on distance or estimated arrival time to form status observations. The intelligent defense decision module and strategy training module receive these status observations and output defense actions, including selecting interference targets, activating auxiliary protection resources, maintaining the action, or not executing the action.

[0109] like Figure 4 As shown, Figure 4 The effect of a failed defense round is shown. When the incoming target selected by the protection unit corresponds to the current protected target and meets the distance, cooldown, and resource conditions, the status, action, and reward interface module determines that the action is a valid interference action. The incoming target control module changes the subsequent flight direction or target point of the target, causing it to deviate from the original protected target. When the action target does not match or the resource conditions are not met, the system records the erroneous action event, and the incoming target continues to fly in the original direction. This process allows the defense action to directly affect the subsequent scene status and round result.

[0110] In one test round of this embodiment of the invention, the system generates a total of 15 incoming targets, of which 15 are effectively interfered with, 0 are erroneous actions, auxiliary protection resources are activated once, the protected target is not attacked, the current round is judged as a successful defense, and the system saves the scene parameters, batch plan, status observation, action events, reward feedback and round results in this round as statistical data for subsequent training and testing.

[0111] Following the above process, the intelligent defense decision module and the strategy training module continuously collect state observations, actions, rewards, next-moment state observations, round-end flags, and scenario parameters in training mode, and update the shared policy network using PPO. The training steps are set to 2,000,000 steps, and the learning rate is set to 3×10⁻⁻⁶. 4The discount factor was set to 0.99, the pruning threshold was set to 0.2, the batch size was set to 64, the experience cache size was set to 512, the final model was saved after training, and the win rate test and scene update module tested the results in fixed rule scenarios and diverse random scenarios.

[0112] Combination Figure 5 , Figure 5 This is a schematic diagram of the scene setup and training process. The test results are further used to adjust the scene parameter sampling ratio in subsequent training.

[0113] To illustrate the testing method of this invention, this embodiment selects four reinforcement learning methods—DQN, PPO, SAC, and DSAC-T—as comparison methods, and tests them with the testing method of this invention under the same state, action, and reward interface. The test set includes fixed rule scenarios and diverse random scenarios, with each setting running for 100 rounds, and the defense win rate is statistically analyzed. The test results are shown in Table 1.

[0114] Table 1. Defense win rate of different methods in the test scenario

[0115] DQN 97.2% 41.9% PPO 98.1% 71.7% SAC 95.1% 70.1% DSAC-T 97.8% 74.2% Test method of the present invention 99.5% 88.1%

[0116] As shown in Table 1, under the test conditions given in this embodiment, the test method of the present invention has a higher defense win rate than the comparative method in both fixed rule scenarios and diverse random scenarios. This is because the test method of the present invention utilizes scene parameters and a 3D construction module, an incoming target control module, and a win rate testing and scene update module to organize training tasks with different target sizes, different batch rhythms, and different incoming directions, enabling the model to learn target recognition and defense action selection rules under more comprehensive scene conditions.

[0117] Example 2

[0118] Figure 6 This is a schematic diagram of the method flow of the present invention. In this embodiment, the method includes steps 1 to 6.

[0119] In a specific step of the method described in this invention, step 1, which involves constructing scene parameters and generating a 3D defense task, includes the following steps:

[0120] Step 1.1 The scene parameters and 3D construction module read the configuration results generated by the task configuration file or random sampler, and organize the number of protected targets, the type of protected targets, the spatial layout, the target generation area, the number of incoming batches, the batch interval, the velocity range, the height range, the set of incoming directions, the disturbance range, and the maximum round duration into a scene configuration parameter set. As shown in equation (1):

[0121] ,

[0122] in, Indicates the number of protected targets activated in the current round. Represents the set of protected target spaces. Indicates the region where the incoming object is generated. Indicates the number of incoming waves. Indicates the target quantity for a single batch or the range of target quantities. Indicates the batch interval. Indicates the speed range. Indicates the height range. Indicate the direction of the attack to gather. Indicates the range of disturbance. Indicates the maximum round duration;

[0123] Step 1.2 Scene Parameters and 3D Construction Module: Configure Parameter Set Based on Scene. Create protected areas, protected target locations, incoming object generation areas, round control objects, and defense task instances; send the protected target configuration to the multi-target protected object management module; and send the incoming object generation parameters to the incoming object control module.

[0124] Step 1.3 The multi-target protection object management module, based on... and Generate a set of protected targets for the current round, and record the target number, target type, three-dimensional coordinates, activation status, health status and hit status for each activated protected target. To ensure that different target sizes share the input interface, the system reserves a fixed upper limit for the number of protected targets, and controls the actual protected targets participating in the interaction in the current round through the activation flag, as shown in the following formula (2):

[0125] ,

[0126] in, Indicates the first Record the location or status of the protected target. This indicates the maximum number of target slots reserved by the system. The activation flag for unprotected targets is set to 0, their observations are empty, and they do not participate in motion, reward, or round determination.

[0127] Step 1.4 The multi-target protection object management module configures a corresponding protection unit for each enabled protection target, and sets the number of protection units accordingly. The first protection unit and the first A connection is established between each protected target. This connection is used to determine whether an incoming target is directed at the protected target that the current protection unit is responsible for, and serves as the basis for judging the effectiveness of the defense action.

[0128] Step 1.5 Scene parameters and 3D construction module: Before the start of each training or test round, resample the number of protected targets, spatial layout, batch rhythm, speed and height, incoming direction and disturbance range to generate new defense task instances, so that subsequent strategy training does not depend on a single target number, a single spatial layout or a single attack rhythm.

[0129] In one specific step of the method described in this invention, step 2, generating the incoming target, includes the following steps:

[0130] Step 2.1 The incoming object control module receives the scene configuration parameter set. And based on the number of incoming batches Batch interval and single batch target quantity Generate batch plan As shown in equation (3):

[0131]

[0132] in, Indicates the batch number. Indicates the first Batch target generation time. Indicates the first Batch target quantity, This indicates that the incoming direction of this batch can be selected. This indicates the speed range for this batch. Indicates the height range for this batch;

[0133] Step 2.2 The incoming target control module reaches [a certain time]. When creating the first The system processes incoming targets in batches and writes the target number, batch number, attack target number, initial position, speed, altitude, direction of attack, and disturbance flag for each incoming target. The attack target number is determined or randomly generated according to rules such as uniform dispersion, random allocation, priority of nearest target, or intersection of far ends.

[0134] Step 2.3 The incoming target control module records the dynamic status of each incoming target, including the current position, the position at the previous moment, the attack target, the speed, the altitude, the batch to which it belongs, the disturbance flag, the survival status, the out-of-target status, and the interfered status, and sends it to the status, action and reward interface module after updating at each time step;

[0135] Step 2.4 The incoming target control module determines the current position of the incoming target. and the corresponding protected target location Calculate the ideal direction of motion As shown in equation (4):

[0136] ,

[0137] in, Indicates the first The attack target number corresponding to each incoming target. This indicates the location of the protected target corresponding to the attack target;

[0138] Step 2.5, to simulate uncertainties in a dynamic scenario, involves the incoming object control module controlling the disturbance term. Incorporate the ideal direction of motion and follow the target speed. Update the position at the next moment and determine the actual direction of motion after the perturbation is applied. :

[0139]

[0140] in, Indicates direction normalization processing. This indicates the actual direction of motion after the perturbation is applied. This represents a single simulation time step, in which the incoming object exhibits continuous variations in velocity, altitude, orientation, and trajectory deviation.

[0141] Step 2.6 When the previous batch of incoming targets has not yet finished, the incoming target control module still generates the next batch of targets according to the batch plan, so that different batches of incoming targets can coexist in the same time window, forming a continuous pressure, batch superposition and dynamic change defense posture.

[0142] In a specific step of the method described in this invention, step 3, which involves acquiring state observations and deciding on defensive actions, includes the following steps:

[0143] Step 3.1 The Status, Action, and Reward Interface Module receives the status of the protected target, the status of the protected unit, and the dynamic status of the incoming object, and uses the first... Each protection unit serves as a reference to construct candidate target state features. These features extract the relative position, relative distance, current direction of movement, historical direction changes, batch number, and target state between the protection unit and the candidate incoming target. The first protection unit and the first State characteristics among candidate incoming targets As shown in equation (6):

[0144]

[0145] in, Indicates relative position, Indicates a change in the direction of history. Indicates relative distance. Indicates the batch number. This indicates whether the target is alive, has missed the target, or has been interfered with.

[0146] Step 3.2 The State, Action, and Reward interface module sorts the candidate target relationship features according to distance, estimated arrival time, and target state, and selects the top... There are 10 candidate targets, which is insufficient. The candidate slots are filled with empty slots, allowing different numbers of incoming objects and different batches to share the same observation dimension. Each protective unit is in constant motion State observation As shown in equation (7):

[0147]

[0148] in, This represents the set of candidate target relation features after sorting and filling. This indicates the resource status, cooling status, and operational status of the protection unit. Indicates the progress of the current round. Indicates splicing;

[0149] Step 3.3 The intelligent defense decision-making module and strategy training module receive state observations. The policy network outputs defensive actions. defensive actions This includes candidate target selection, whether to enable or disable auxiliary resources, whether to maintain or not execute actions, and defensive actions. As shown in equation (8):

[0150] ,

[0151] in, The parameter is During the training phase, the policy network updates itself based on the collected round experience; during the testing phase, the policy network loads the trained parameters and performs action reasoning only.

[0152] In one specific step of the method described in this invention, step 4, which involves performing defensive actions and completing the environmental state transition, includes the following steps:

[0153] Step 4.1 The intelligent defense decision-making module and strategy training module will execute defense actions. The action command is sent to the status, action and reward interface module, which translates the action command into defensive behavior in the 3D scene. The defensive behavior in the 3D scene includes target selection, interference triggering, auxiliary protection activation and action maintenance.

[0154] Step 4.2: The status, action, and reward interface module determines the validity of an action based on the association between the protection unit and the protected target, the attack target number of the incoming object, distance conditions, resource status, and cooldown status. When the selected incoming object points to the protected object under the responsibility of the current protection unit and meets the resource, distance, or cooldown conditions, the system generates a valid interference event. When the action target does not match or the resource conditions are not met, the system records an erroneous action event.

[0155] 4.3 When the action is effective, the incoming target control module modifies the subsequent movement direction or target point of the incoming target according to the effective interference event, causing the target trajectory to deflect; when the action is invalid or not executed, the incoming target continues to fly normally according to the above formulas (4) and (5). For the interfered target, the system estimates the miss distance based on the location of the interference, the timing of the response, and the current direction of the target. As shown in equation (9):

[0156] ,

[0157] in, This indicates the distance between the incoming target and the corresponding protected target when the interference occurs. This represents the maximum effective distance used for normalization. This means randomly sampling the deflection amplitude within a given range. In this way, defensive actions can change the subsequent motion state of the target, causing the environment to transition from the current state to the next state.

[0158] Step 4.4 The status, action, and reward interface module generates a single-step feedback signal based on the validity of the current action, the target deflection status, the protected target's hit status, erroneous actions, and resource consumption. Single-step feedback signal The policy is used for policy updates during the training phase and for process recording during the testing phase, as shown in equation (10) below:

[0159]

[0160] in, This indicates feedback on the final or interim defense results. This indicates the feedback on target matching and effective identification. This indicates interference deflection or off-target feedback. This indicates the cost of incorrect actions, resource consumption, or frequent switching. , and The weighting coefficients are used to retain feedback on the success or failure of defense, while also guiding the model to learn appropriate target selection and response timing through process feedback.

[0161] In a specific step of the method described in this invention, step 5, which involves determining the round state and using PPO to collect experience for training, includes the following steps:

[0162] Step 5.1 The win rate test and scenario update module determines whether the current round has ended based on the target being hit, the incoming target missing the target, the batch completion status, and the round time. If there are still unfinished batches or incoming targets that have not been terminated, the system continues to execute steps 3 to 4. If all batches have ended, the target being hit, or the maximum round time has been reached, the current round ends and enters the experience processing process.

[0163] Step 5.2 In training mode, the State, Action, and Reward interface module writes the state observation, action, feedback signal, next time step state observation, round end flag, and scene configuration parameter set of each time step into the sample set. Sample set It can be expressed as the following formula (11):

[0164] ,

[0165] in, Indicates the end of the round. This represents the set of scenario configuration parameters corresponding to the source of the samples, and the sample index satisfies... and Sample set It is used for both strategy updates and failure scenario analysis and difficult parameter combination statistics, where r and a represent rewards and actions;

[0166] Step 5.3 The intelligent defense decision-making module and the strategy training module use a shared policy network to process the local decision sequences of multiple defense units. Although the multiple defense units correspond to different protection targets and different local situations, they use the same observation structure and action structure. Therefore, their interaction experience can be used to update the same policy network. The goal of policy training is to learn the action selection rules on the scene distribution to maximize the benefits. :

[0167] ,

[0168] in, This represents the scene sampling distribution during the training phase. Representation Strategy In the scene configuration parameter set Expected return in the corresponding round;

[0169] Step 5.4 In training the neural network of the intelligent defense decision module, the policy update adopts the near-end policy optimization algorithm, given a sample set collected from multiple scenarios and multiple protection units. The intelligent defense decision-making module and the strategy training module update the policy network based on advantage estimation, and limit the magnitude of each update by pruning terms. Loss represents the amount of loss.

[0170] ,

[0171] in, Indicates the policy network parameters, This represents the ratio of the probability of the new strategy to the probability of the old strategy. Indicates the advantage estimate, The above update process, which represents the pruning threshold, enables the strategy to gradually learn the patterns of target matching, response selection, and resource usage under conditions of multiple targets, multiple locations, multiple batches, and multiple perturbations.

[0172] Step 5.5 After completing multiple rounds of strategy updates, the intelligent defense decision module and the strategy training module save the current strategy parameters and send the round gains, success flags, failure reasons and action records generated during the training process to the win rate test and scenario update module for subsequent testing and scenario parameter adjustment.

[0173] In a specific step of the method described in this invention, step 6, which involves winning rate testing, scene parameter updating, and repeated training, specifically includes:

[0174] Step 6.1 The win rate test and scenario update module calls the saved strategy parameters in test mode and stops the strategy update. During the test phase, the scenario generation, incoming object control, state observation, action execution, environment transition and round determination logic of steps 1 to 5 are still running to ensure that the training process and the testing process use the same set of interactions.

[0175] Step 6.2 The win rate test and scenario update module statistically analyzes the model performance in fixed rule scenarios and diverse random scenarios, including defense win rate, average feedback value, number of effective interferences, number of erroneous actions, and number of times the protected target is attacked. Fixed rule scenarios are used to test whether the model has mastered the clear attack rules, while diverse random scenarios are used to test the model's adaptability to unknown parameter combinations.

[0176] Step 6.3 The win rate testing and scenario update module calculates recent success rates based on the combination of scenario parameters. Average round gains Failure frequency And feedback on fluctuations, and define the difficulty scenario score accordingly. As shown in equation (14):

[0177]

[0178] in, , and These are the weighting coefficients. This indicates normalization. A higher difficulty score indicates that the current strategy is more unstable under this parameter combination;

[0179] Step 6.4 Scene Parameters and 3D Construction Module: Based on the difficult scene records reported by the win rate test and scene update module, adjust the scene sampling distribution in subsequent training. As shown in equation (15):

[0180]

[0181] in, Represents a set of optional scenario parameter combinations. Control the ratio of uniform sampling to sampling in difficult scenarios. Using temperature coefficients, this scene update mechanism maintains overall scene coverage while increasing the frequency of parameter combinations that have high recent failure rates, large reward fluctuations, or difficulties in target recognition.

[0182] Step 6.5 The system re-executes steps 1 to 6 based on the updated scene sampling distribution, forming a cyclical process of scene parameter generation, defense task construction, incoming target control, state observation, defense action decision-making, environmental state transition, round experience training, win rate testing, and scene parameter update. Through continuous repeated training, the intelligent defense strategy can continuously engage with vulnerable scenes that have not yet been fully adapted to, and gradually improve the defense win rate under dynamic conditions of multiple targets, multiple batches, and multiple parameters.

[0183] This invention is not limited to the above embodiments, which are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the appended claims.

Claims

1. An intelligent defense control system for dynamic scenarios involving multiple targets, multiple batches, and multiple parameters, comprising: Scene Parameters and 3D Construction Module: The Scene Parameters and 3D Construction Module reads or randomly generates a set of scene configuration parameters, and creates protected areas, protected targets, protective units, and round control objects based on the set of scene configuration parameters. This module encapsulates the number of protected targets, spatial layout, number of batches, batch interval, speed range, height range, set of incoming directions, disturbance range, and round duration in a unified manner. Multi-target protection object management module: The multi-target protection object management module generates a set of protected targets and a set of protected units based on scene parameters and the enable flag sent by the 3D construction module. For each protected target, it records the target number, target type, 3D coordinates, health status, hit status and corresponding protected unit, and sends the status of the protected target and the status of the protected unit to the status, action and reward interface module. Incoming Object Control Module: The incoming object control module generates incoming objects according to the batch plan, writes batch number, attack target number, initial position, speed, altitude, direction of movement and disturbance flag for each incoming object, updates the target's current position, historical position and trajectory deflection status at each time step, and changes the target's subsequent direction of movement or target point after receiving valid interference information. Status, Action and Reward Interface Module: The status, action and reward interface module uses the protection unit as a reference to calculate the relative position, relative distance, current movement direction, historical movement direction, batch information and resource status of candidate targets, combines the above information into status observation, and can receive action execution results, and generate feedback signals based on target matching, effective interference, ineffective actions, resource consumption and defense results; The intelligent defense decision-making module and strategy training module receive state observations and output target selection actions, auxiliary resource actions, and action maintenance information. In training mode, a shared policy network is used to process local decision samples from multiple defense units, and the policy parameters are updated through a near-end policy optimization algorithm. In test or deployment mode, existing policy parameters are loaded and inference is performed. Win rate testing and scene update module: The win rate testing and scene update module judges the round result based on the protected target's hit status, the incoming object's miss status, the batch completion status, and the round time, and records scene parameters, interaction trajectories, action events, and feedback signals. It feeds back the failed scene and difficult parameter combination to the scene parameter and 3D construction module.

2. A control method for an intelligent defense control system based on multi-target, multi-batch, and multi-parameter dynamic scenarios, comprising the following steps: Step 1: Construct scene parameters and generate a 3D defense task; Step 2: Generate the incoming attack object; Step 3: Obtain state observations and decide on defensive actions; Step 4: Execute defensive actions and complete the environmental state transition; Step 5: Determine the round status and use PPO to collect experience for training; Step 6: Win rate test, scene parameters update and repeat training.

3. The intelligent defense control method for multi-target, multi-batch, multi-parameter dynamic scenarios according to claim 2, characterized in that, Step 1, which involves constructing scene parameters and generating a 3D defense task, specifically includes: Step 1.1 The scene parameters and 3D construction module read the configuration results generated by the task configuration file or random sampler, and organize the number of protected targets, the type of protected targets, the spatial layout, the target generation area, the number of incoming batches, the batch interval, the velocity range, the height range, the set of incoming directions, the disturbance range, and the maximum round duration into a scene configuration parameter set. As shown in equation (1): ， in, Indicates the number of protected targets activated in the current round. Represents the set of protected target spaces. Indicates the region where the incoming object is generated. Indicates the number of incoming waves. Indicates the target quantity for a single batch or the range of target quantities. Indicates the batch interval. Indicates the speed range. Indicates the height range. Indicate the direction of the attack to gather. Indicates the range of the disturbance. Indicates the maximum round duration; Step 1.2 Scene Parameters and 3D Construction Module: Configure Parameter Set Based on Scene. Create protected areas, protected target locations, incoming object generation areas, round control objects, and defense task instances; send the protected target configuration to the multi-target protected object management module; and send the incoming object generation parameters to the incoming object control module. Step 1.3 The multi-target protection object management module, based on... and Generate a set of protected targets for the current round, and record the target number, target type, three-dimensional coordinates, activation status, health status and hit status for each activated protected target. To ensure that different target sizes share the input interface, the system reserves a fixed upper limit for the number of protected targets, and controls the actual protected targets participating in the interaction in the current round through the activation flag, as shown in the following formula (2): ， in, Indicates the first Record the location or status of the protected target. This indicates the maximum number of target slots reserved by the system. The activation flag for unprotected targets is set to 0, their observations are empty, and they do not participate in motion, reward, or round determination. Step 1.4 The multi-target protection object management module configures a corresponding protection unit for each enabled protection target, and sets the number of protection units accordingly. The first protection unit and the first A connection is established between each protected target. This connection is used to determine whether an incoming target is directed at the protected target that the current protection unit is responsible for, and serves as the basis for judging the effectiveness of the defense action. Step 1.5 Scene parameters and 3D construction module: Before the start of each training or test round, resample the number of protected targets, spatial layout, batch rhythm, speed and height, incoming direction and disturbance range to generate new defense task instances, so that subsequent strategy training does not depend on a single target number, a single spatial layout or a single attack rhythm.

4. The intelligent defense control method for multi-target, multi-batch, multi-parameter dynamic scenarios according to claim 2, characterized in that, Step 2, generating the incoming target, specifically includes: Step 2.1 The incoming object control module receives the scene configuration parameter set. And based on the number of incoming batches Batch interval and single batch target quantity Generate batch plan As shown in equation (3): ， in, Indicates the batch number. Indicates the first Batch target generation time. Indicates the first Batch target quantity, This indicates that the incoming direction of this batch can be selected. This indicates the speed range for this batch. Indicates the height range for this batch; Step 2.2 The incoming target control module reaches [a certain time]. When creating the first The system processes incoming targets in batches and writes the target number, batch number, attack target number, initial position, speed, altitude, direction of attack, and disturbance flag for each incoming target. The attack target number is determined or randomly generated according to rules such as uniform dispersion, random allocation, priority of nearest target, or intersection of far ends. Step 2.3 The incoming target control module records the dynamic status of each incoming target, including the current position, the position at the previous moment, the attack target, the speed, the altitude, the batch to which it belongs, the disturbance flag, the survival status, the out-of-target status, and the interfered status, and sends it to the status, action and reward interface module after updating at each time step; Step 2.4 The incoming target control module determines the current position of the incoming target. and the corresponding protected target location Calculate the ideal direction of motion As shown in equation (4): ， in, Indicates the first The attack target number corresponding to each incoming target. This indicates the location of the protected target corresponding to the attack target; Step 2.5, to simulate uncertainties in a dynamic scenario, involves the incoming object control module controlling the disturbance term. Incorporate the ideal direction of motion and follow the target speed. Update the position at the next moment and determine the actual direction of motion after the perturbation is applied. : ， in, Indicates direction normalization processing. This indicates the actual direction of motion after the perturbation is applied. This represents a single simulation time step, in which the incoming object exhibits continuous variations in velocity, altitude, direction, and trajectory deviation. Step 2.6 When the previous batch of incoming targets has not yet finished, the incoming target control module still generates the next batch of targets according to the batch plan, so that different batches of incoming targets can coexist in the same time window, forming a continuous pressure, batch superposition and dynamic change defense posture.

5. The intelligent defense control method for multi-target, multi-batch, multi-parameter dynamic scenarios according to claim 2, characterized in that, Step 3, which involves acquiring state observations and deciding on defensive actions, specifically includes: Step 3.1 The Status, Action, and Reward Interface Module receives the status of the protected target, the status of the protected unit, and the dynamic status of the incoming object, and uses the first... Each protection unit serves as a reference to construct candidate target state features. These features extract the relative position, relative distance, current direction of movement, historical direction changes, batch number, and target state between the protection unit and the candidate incoming target. The first protection unit and the first State characteristics among candidate incoming targets As shown in equation (6): ， in, Indicates relative position, Indicates a change in the direction of history. Indicates relative distance. Indicates the batch number. This indicates whether the target is alive, has missed the target, or has been interfered with. Step 3.2 The State, Action, and Reward interface module sorts the candidate target relationship features according to distance, estimated arrival time, and target state, and selects the top... There are 10 candidate targets, which is insufficient. The candidate slots are filled with empty slots, allowing different numbers of incoming objects and different batches to share the same observation dimension. Each protective unit is in constant motion State observation As shown in equation (7): ， in, This represents the set of candidate target relation features after sorting and filling. This indicates the resource status, cooling status, and operational status of the protection unit. Indicates the progress of the current round. Indicates splicing; Step 3.3 The intelligent defense decision-making module and strategy training module receive state observations. The policy network outputs defensive actions. Defensive actions This includes candidate target selection, whether to enable or disable auxiliary resources, whether to maintain or not execute actions, and defensive actions. As shown in equation (8): ， in, The parameter is During the training phase, the policy network updates itself based on the collected round experience; during the testing phase, the policy network loads the trained parameters and performs action reasoning only.

6. The intelligent defense control method for multi-target, multi-batch, multi-parameter dynamic scenarios according to claim 2, characterized in that, Step 4, which involves performing defensive actions and completing the environmental state transition, specifically includes: Step 4.1 The intelligent defense decision-making module and strategy training module will execute defense actions. The action command is sent to the status, action and reward interface module, which translates the action command into defensive behavior in the 3D scene. The defensive behavior in the 3D scene includes target selection, interference triggering, auxiliary protection activation and action maintenance. Step 4.2: The status, action, and reward interface module determines the validity of an action based on the association between the protection unit and the protected target, the attack target number of the incoming object, distance conditions, resource status, and cooldown status. When the selected incoming object points to the protected object under the responsibility of the current protection unit and meets the resource, distance, or cooldown conditions, the system generates a valid interference event. When the action target does not match or the resource conditions are not met, the system records an erroneous action event. Step 4.3: When the action is effective, the incoming target control module modifies the subsequent movement direction or target point of the incoming target according to the effective interference event, causing the target trajectory to deflect; when the action is invalid or not executed, the incoming target continues to fly normally according to the above formulas (4) and (5). For the interfered target, the system estimates the miss distance based on the location of the interference, the timing of the response, and the current direction of the target. As shown in equation (9): ， in, This indicates the distance between the incoming target and the corresponding protected target when the interference occurs. This represents the maximum effective distance used for normalization. This means randomly sampling the deflection amplitude within a given range. In this way, defensive actions can change the subsequent motion state of the target, causing the environment to transition from the current state to the next state. Step 4.4 The status, action, and reward interface module generates a single-step feedback signal based on the validity of the current action, the target deflection status, the protected target's hit status, erroneous actions, and resource consumption. Single-step feedback signal The policy is used for policy updates during the training phase and for process recording during the testing phase, as shown in equation (10) below: ， in, This indicates feedback on the final or interim defense results. This indicates the feedback on target matching and effective identification. This indicates interference deflection or off-target feedback. This indicates the cost of incorrect actions, resource consumption, or frequent switching. , and The weighting coefficients are used to retain feedback on the success or failure of defense, while also guiding the model to learn appropriate target selection and response timing through process feedback.

7. The intelligent defense control method for multi-target, multi-batch, multi-parameter dynamic scenarios according to claim 2, characterized in that, Step 5, which involves determining the round state and using PPO to collect experience for training, specifically includes: Step 5.1 The win rate test and scenario update module determines whether the current round has ended based on the target being hit, the incoming target missing the target, the batch completion status, and the round time. If there are still unfinished batches or incoming targets that have not been terminated, the system continues to execute steps 3 to 4. If all batches have ended, the target being hit, or the maximum round time has been reached, the current round ends and enters the experience processing process. Step 5.2 In training mode, the State, Action, and Reward interface module writes the state observation, action, feedback signal, next time step state observation, round end flag, and scene configuration parameter set of each time step into the sample set. Sample set It can be expressed as the following formula (11): ， in, Indicates the end of the round. Describe the state observation, action, reward, and next time step observation respectively. This represents the set of scenario configuration parameters corresponding to the source of the samples, and the sample index satisfies... and Sample set It is used for both strategy updates and failure scenario analysis and statistics of difficult parameter combinations; r and a represent rewards and actions; Step 5.3 The intelligent defense decision-making module and the strategy training module use a shared policy network to process the local decision sequences of multiple defense units. Although the multiple defense units correspond to different protection targets and different local situations, they use the same observation structure and action structure. Therefore, their interaction experience can be used to update the same policy network. The goal of policy training is to learn the action selection rules on the scene distribution to maximize the benefits. As shown in equation (12): ， in, This represents the scene sampling distribution during the training phase. Representation Strategy In the scene configuration parameter set Expected return in the corresponding round; Step 5.4 In training the neural network of the intelligent defense decision module, the policy update adopts the near-end policy optimization algorithm, given a sample set collected from multiple scenarios and multiple protection units. The intelligent defense decision-making module and the strategy training module update the policy network based on advantage estimation, and limit the magnitude of each update by pruning terms. Loss represents the amount of loss. ， in, Represents the loss function. Indicates the policy network parameters, This represents the ratio of the probabilities of the new and old strategies. Indicates the advantage estimate, The above update process, which represents the pruning threshold, enables the strategy to learn target matching, response selection, and resource usage patterns progressively under conditions of multiple targets, multiple locations, multiple batches, and multiple perturbations. Step 5.5 After completing multiple rounds of strategy updates, the intelligent defense decision module and the strategy training module save the current strategy parameters and send the round gains, success flags, failure reasons and action records generated during the training process to the win rate test and scenario update module for subsequent testing and scenario parameter adjustment.

8. The intelligent defense control method for multi-target, multi-batch, multi-parameter dynamic scenarios according to claim 2, characterized in that, Step 6, which involves winning rate testing, scene parameter updating, and repeated training, specifically includes: Step 6.1 The win rate test and scenario update module calls the saved strategy parameters in test mode and stops the strategy update. During the test phase, the scenario generation, incoming object control, state observation, action execution, environment transition and round determination logic of steps 1 to 5 are still running to ensure that the training process and the testing process use the same set of interactions. Step 6.2 The win rate test and scenario update module statistically analyzes the model performance in fixed rule scenarios and diverse random scenarios, including defense win rate, average feedback value, number of effective interferences, number of erroneous actions, and number of times the protected target is attacked. Fixed rule scenarios are used to test whether the model has mastered the clear attack rules, while diverse random scenarios are used to test the model's adaptability to unknown parameter combinations. Step 6.3 The win rate testing and scenario update module calculates recent success rates based on the combination of scenario parameters. Average round gains Failure frequency And feedback on fluctuations, and define the difficulty scenario score accordingly. As shown in equation (14): ， in, , and These are the weighting coefficients. This indicates normalization; a higher difficulty score indicates that the current strategy is more unstable under this parameter combination. Step 6.4 Scene Parameters and 3D Construction Module: Based on the difficult scene records reported by the win rate test and scene update module, adjust the scene sampling distribution in subsequent training. As shown in equation (15): ， in, Represents a set of optional scenario parameter combinations. Control the ratio of uniform sampling to sampling in difficult scenarios. Using temperature coefficients, this scene update mechanism maintains overall scene coverage while increasing the frequency of parameter combinations that have high recent failure rates, large reward fluctuations, or difficulties in target recognition. Step 6.5 The system re-executes steps 1 to 6 based on the updated scene sampling distribution, forming a cyclical process of scene parameter generation, defense task construction, incoming target control, state observation, defense action decision-making, environmental state transition, round experience training, win rate testing, and scene parameter update. Through continuous repeated training, the intelligent defense strategy can continuously engage with vulnerable scenes that have not yet been fully adapted to, and gradually improve the defense win rate under dynamic conditions of multiple targets, multiple batches, and multiple parameters.

Images