Subsonic speed target drone launch control method and subsonic speed target drone launch control system

Abstract

The invention belongs to the technical field of information, and discloses a subsonic target drone launch control method and system, and the method comprises the steps: obtaining target drone training task data, extracting specific scene parameters from task target diversity, and determining a launch trajectory configuration demand; analyzing the dynamic environment adaptation parameters according to a launching track configuration requirement to obtain a time sequence control precision requirement; a control instruction module is processed in an instruction combination mode, and an initial launching control strategy under the modular design capability is generated; for the initial launch control strategy, analyzing dynamic change in complex environment application, and judging a real-time adjustment strategy demand; if the real-time adjustment strategy requirement exists, flexible configuration of a control system is updated according to the dynamic environment adaptation parameters, and an adjusted launching track is obtained; and generating a target drone control scheme with an accurate training task through the adjusted launching track in combination with scene simulation effect data.

Description

Technical Field

[0001] This invention belongs to the field of information technology, specifically relating to a subsonic target drone launch control method and system. Background Technology

[0002] In modern military training and weapons testing, target drone launch control technology is a crucial support for improving the effectiveness of combat simulations and the quality of training. This technology is directly related to the accuracy and diversity of training missions and has irreplaceable value in ensuring the enhancement of combat capabilities. However, current related technologies have revealed significant shortcomings when adapting to complex training needs, mainly in their weak ability to flexibly respond to different scenarios. Especially when facing changing mission objectives and dynamic environments, existing methods often struggle to quickly adjust control strategies, resulting in a significant gap between training effectiveness and actual requirements.

[0003] A deeper analysis of the challenges in this field reveals that the core issue lies in achieving flexibility and adaptability in the control system. Firstly, the configuration of launch trajectories and timing lacks sufficient flexibility, making rapid adjustments impossible based on specific mission requirements. This hinders the target drone from achieving the desired results when simulating different combat scenarios. Furthermore, this problem exposes the overly fixed combination of control commands, making it difficult to address diverse training targets through modularity. This directly limits the system's application capabilities in complex environments.

[0004] Therefore, designing a method that can flexibly configure launch trajectories and timings while supporting modular combinations of control commands has become a key issue in improving the adaptability of subsonic target drone launch control technology. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a subsonic target drone launch control method and system. By combining scenario simulation data, this invention can generate precise target drone control schemes, effectively improving the intelligence level and environmental adaptability of target drone training tasks, and providing a flexible and efficient solution for target drone control under complex conditions.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A method for controlling the launch of a subsonic target drone, the method comprising:

[0008] Acquire target drone training mission data and extract specific scenario parameters to determine launch trajectory configuration requirements;

[0009] Based on the launch trajectory configuration requirements, dynamic environment adaptation parameters are analyzed to obtain the timing control accuracy requirements;

[0010] When the timing control accuracy requirement is higher than the threshold, the corresponding instructions are extracted and combined from the preset instruction module library to generate the initial launch control strategy;

[0011] For the initial launch control strategy, analyze the dynamic changes in complex environment applications and determine the need for real-time strategy adjustment.

[0012] If there is a need to adjust the strategy in real time, the control system is flexibly configured according to the dynamic environment adaptation parameters to obtain the adjusted launch trajectory.

[0013] By adjusting the launch trajectory and combining it with scenario simulation data, a target drone control scheme for the training mission is generated.

[0014] Preferably, the method for acquiring target drone training task data and extracting specific scenario parameters to determine launch trajectory configuration requirements includes:

[0015] By extracting task data from the target machine training system and analyzing the diverse target information contained therein, preliminary task target classification results are obtained.

[0016] Based on the task objective classification results, data is broken down for specific scenarios under each category to determine the key scenario parameters corresponding to each scenario.

[0017] By using a pre-established scene parameter database, the extracted scene parameters are matched to obtain the parameter mapping relationship related to the launch trajectory;

[0018] If there is missing or abnormal data in the parameter mapping relationship, the abnormal data is repaired by data cleaning tools, and it is determined whether the repaired parameters meet the preset threshold range.

[0019] By combining the repaired parameter data with the launch trajectory configuration model, a trajectory configuration scheme adapted to the current mission is calculated, and the final trajectory configuration result is determined.

[0020] After obtaining the final trajectory configuration result, the matching degree between the configuration result and the task data is verified to obtain the verified configuration applicability conclusion.

[0021] If the verified suitability conclusion does not meet the preset standard, the configuration model is optimized by adjusting the parameter weights, and it is determined whether the optimized configuration meets the requirements of the task objective.

[0022] Preferably, the method for obtaining the parameter mapping relationship related to the launch trajectory by matching the currently extracted scene parameters with a pre-established scene parameter database includes:

[0023] The radiance received at the entrance pupil of a spaceborne sensor in the mid-infrared band mainly includes surface radiation and path radiation. The contribution of a point on the surface to the spectral radiance of the entrance pupil can be decomposed as follows:

[0024]

[0025] In the formula, L λ,T Total radiance; and These are surface thermal radiation and path radiation, respectively; τ λ ε represents spectral transmittance. λ λ is the spectral emissivity; h is Planck's constant; c is the speed of light; λ is the wavelength; k is Boltzmann's constant; T is the thermodynamic temperature;

[0026] The forward calculation process for the radiance of a surface point is as follows:

[0027] L λ =f λ (ε λ ,T,H,θ1,θ2);

[0028] In the formula, H is the altitude of the ground point; θ1 is the zenith angle from the sun to the ground point; θ2 is the zenith angle from the ground point to the sensor; f λ The mapping from five elements to radiance includes the Planck function, which is a complex nonlinear function; while the emissivity inversion problem is expressed as:

[0029] ε λ =g λ (L λ ,T,H,θ1,θ2);

[0030] In the formula, the function g λ It characterizes the mapping relationship between various factors affecting the surface emissivity inversion results and the surface emissivity.

[0031] Preferably, the method for obtaining the timing control accuracy requirements by analyzing dynamic environment adaptation parameters based on launch trajectory configuration requirements includes:

[0032] By extracting key data from the configuration requirements of the launch trajectory, initial environmental variables and trajectory parameters are obtained, and the preliminary range of adaptation parameters is determined.

[0033] Based on the preliminary adaptation parameter range, combined with real-time monitoring data in the dynamic environment, the impact of environmental changes on the trajectory is analyzed, and the characteristic distribution of environmental changes is obtained.

[0034] Using the results of the characteristic distribution, parameters are adjusted in response to environmental changes, updated values ​​of the adaptation parameters are derived, and it is determined whether the constraints of the time-series control are met.

[0035] If the updated adaptation parameters meet the constraints of timing control, the adjusted parameters will be applied to trajectory optimization to obtain optimized trajectory data.

[0036] If the conditions are not met, return to the parameter adjustment stage and reanalyze the environmental change characteristics;

[0037] By combining the optimized trajectory data with the control strategy, specific instructions for timing control are generated, and the method for achieving control accuracy is determined.

[0038] The generated timing control instructions are obtained, and the deviations during the execution of the instructions are analyzed in conjunction with the real-time monitored data stream to obtain the verification results of the accuracy requirements.

[0039] Based on the verification results, key parameters in the control strategy were adjusted, the instruction generation logic was optimized, and the final timing control accuracy scheme was derived.

[0040] Preferably, the method for generating the preset instruction module library includes:

[0041] If the timing control accuracy is higher than the preset threshold, the matching instruction is extracted from the instruction module library through the data processing flow to obtain a preliminary instruction set;

[0042] The initial instruction set is filtered through a module matching mechanism to determine the compatibility of the instructions with the current control task and to identify a valid instruction subset.

[0043] The effective instruction subset is sorted and reorganized using instruction generation rules to generate an optimized combination of control instructions;

[0044] Based on the optimized combination of control instructions according to the control execution efficiency evaluation, it is determined whether the real-time requirements are met, and the sequence of execution instructions is obtained.

[0045] The sequence of executed instructions is adjusted by an instruction combination optimization algorithm to determine the final combination of control instructions;

[0046] Obtain the module library update data, determine whether the instruction module library needs to be updated, and obtain the updated instruction module library;

[0047] For the updated instruction module library, a data processing procedure is used to verify its integrity and determine the preset instruction module library.

[0048] Preferably, when the timing control accuracy requirement is higher than a threshold, the method for extracting and combining corresponding instructions from a preset instruction module library to generate an initial launch control strategy includes:

[0049] Obtain a modular instruction set, determine the instruction type and priority by parsing the instruction content, and obtain a categorized instruction sequence;

[0050] Control instructions are extracted from the classified instruction sequence, and the instructions are combined using preset logical rules to generate an initial control instruction set.

[0051] For the initial control instruction set, the decision tree algorithm is applied to analyze the dependencies between instructions, determine the execution order of instructions, and obtain an ordered sequence of control instructions.

[0052] Based on the ordered sequence of control instructions, a state machine model is used to simulate the launch control process, determine the triggering conditions of each instruction, and generate a preliminary launch strategy.

[0053] Modular design parameters are extracted from the initial launch strategy. If the parameters meet the preset threshold, the strategy structure is optimized through dynamic programming algorithm to obtain the optimized launch control strategy.

[0054] For the optimized launch control strategy, obtain feedback on modular design capabilities, determine the compatibility between the strategy and the modular design, and generate the final launch control strategy.

[0055] Preferably, for the initial control instruction set, the method of applying the decision tree algorithm to analyze the dependencies between instructions, determine the instruction execution order, and obtain an ordered control instruction sequence includes:

[0056] A finite set of atomic instructions is predefined. In the instruction parsing part, a large model is used to parse the text instructions into subtasks composed of atomic instructions. In the instruction generation part, a frame-by-frame instruction generation network is designed based on the conditional variational autoencoder (cVAE). In the instruction generation stage, different instruction features are considered to realize the instruction sequence generation task.

[0057] Preferably, the instruction parsing method based on a large model includes:

[0058] Using the GPT-3-davinci interface as the Large Language Model (LLM) planner, for a natural language instruction C, the goal is to obtain a decision tree R representing the task completion process and the final state S for completing the instruction. T ;

[0059] Based on prompting engineering, design reasonable prompts so that the planner can generate acceptable answers to the task;

[0060] The prompt content includes four parts: task content description, output paradigm description, self-check prompts for generated results, and examples. 1) The content describes the specific task of instruction decomposition: parsing the input instructions with the scene into a formatted output that can drive the character to complete the instruction task; the output consists of two parts: first, it includes the character, the scene, the different states that the subsonic target drone may exist in the scene, and the next action the character needs to take in that state to complete the instruction, defined as D. D Secondly, the final state that the subsonic target drone in the scene needs to reach to complete the command is defined as D. S;2) The paradigm defines the specific content of the formatted output, including the structural format and a specific description of the subsonic target drone and its possible states. In particular, it describes the scene category and the subsonic target drones that may exist in the scene in conjunction with the scene content;3) The result self-check prompt requires the planner to confirm that the output result meets the above conditions through a self-check after generating the parsed result;4) Samples are given as a reference to enable the planner to generate similar results.

[0061] For the instruction parsing description D of the large model output D Merging subsonic target drones targeting the same interaction i The different state judgments are merged into a decision tree D. The decision tree contains decision nodes and terminal nodes. The decision node represents the i-th interactive subsonic target machine o that needs to be judged to complete the instruction task. i According to its state s i Different values ​​result in different decision nodes or termination nodes; the termination node contains the character's instruction b and its corresponding sequence style sty; instruction b contains the character's interaction during the subsonic target machine o. i and interactive actions ITR j Yes, or the character's movement animation MV. p The set of interactive subsonic target machines (obj), the set of interactive actions (itr), the set of movement actions (mv), and the set of sequential style (sty); the description of the final state (D) S The text describes the character, the scene, and the final state of the subsonic target drone after completing its mission. Combined with scene awareness, this is transformed into the symbolic description S defined above. T It includes the character's position and orientation, as well as the status, position, and orientation information of each subsonic target drone.

[0062] Preferably, the method for designing a frame-by-frame action generation network based on a cVAE structure includes: the action of the i-th frame is defined as:

[0063]

[0064] In the formula, This indicates the position, rotation, and speed of the subsonic target drone relative to the ground; represents the position of the subsonic target drone relative to the ground in the next second, and j represents the number of commands for the virtual character subsonic target drone; t time steps are sampled in a 2s time window between each frame [-1,1]s; This indicates the position and direction of movement of the current frame relative to the ground in the previous frame; These represent approximate values ​​of the current frame's position relative to the ground in the previous frame and its direction of movement, respectively. This represents the one-hot encoded vector of the instruction label at each time step t. The instruction label includes content and style labels. The 10 content labels and 8 style labels of the instruction are encoded uniformly, i.e., n a =10 × 8 = 80; and Indicates the target's location and orientation; It is the one-hot encoded vector of the action label to be executed at each time step t; Labels indicating the five end connection points of the subsonic target drone;

[0065] The encoder encodes the subsonic target representation O into a latent space vector O through a three-layer fully connected network. lat At the same time, set the current frame state X i and the previous frame's state X i-1 Similarly, it is encoded into a low-dimensional latent space vector Z through a three-layer fully connected network. mot ;

[0066] The decoder consists of a prediction network and a gating network, both of which are three-layer fully connected networks.

[0067] The prediction network is based on the latent space variable O of the encoder output. lat and Z mot By combining the instruction vector from the previous frame, the instructions for the virtual character in the next frame are predicted.

[0068] Gated networks with Z mot and X i-1 As input, dynamically generate mixed weight coefficients ω = {ω1, ω2, ...}, which are mixed with expert weights to obtain the weights of the prediction network. The instruction generation network is generated in an end-to-end manner.

[0069] The present invention also provides a subsonic target drone launch control system, the system being used to implement the aforementioned method, the system comprising: a mission data acquisition module, a scenario parameter extraction module, a timing control analysis module, an initial strategy generation module, a dynamic adjustment judgment module, and a precise scheme generation module;

[0070] The task data acquisition module is used to acquire target drone training task data and extract specific scenario parameters to determine launch trajectory configuration requirements;

[0071] The scenario parameter extraction module is used to analyze dynamic environment adaptation parameters according to the launch trajectory configuration requirements to obtain the timing control accuracy requirements;

[0072] The timing control analysis module is used to extract and combine corresponding instructions from the preset instruction module library to generate an initial launch control strategy when the timing control accuracy requirement is higher than the threshold.

[0073] The initial strategy generation module is used to analyze the dynamic changes in complex environment applications and determine the need for real-time strategy adjustment based on the initial launch control strategy.

[0074] The dynamic adjustment judgment module is used to update the flexible configuration of the control system according to the dynamic environment adaptation parameters if there is a need for real-time adjustment strategy, so as to obtain the adjusted launch trajectory.

[0075] The precision scheme generation module is used to generate a target drone control scheme for the training mission by combining the adjusted launch trajectory with scenario simulation effect data.

[0076] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0077] This invention discloses an intelligent control method for target drone training tasks. By acquiring training task data and extracting specific scenario parameters, the launch trajectory configuration requirements are determined. Based on the requirements, dynamic environment adaptation parameters are analyzed to derive timing control accuracy requirements. When the accuracy requirement exceeds a threshold, corresponding instructions are extracted and combined from a preset instruction module library to generate an initial launch control strategy. For dynamic changes in complex environments, this invention can adjust the strategy in real time, updating the control system configuration to adapt to environmental changes, ultimately obtaining an adjusted launch trajectory. Combined with scenario simulation data, this invention can generate precise target drone control schemes, effectively improving the intelligence level and environmental adaptability of target drone training tasks, and providing a flexible and efficient solution for target drone control under complex conditions. Attached Figure Description

[0078] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0079] Figure 1 This is a schematic flowchart of a subsonic target drone launch control method according to an embodiment of the present invention;

[0080] Figure 2 This is a schematic diagram of a decision tree according to an embodiment of the present invention;

[0081] Figure 3 This is a schematic diagram of the frame-by-frame instruction generation network according to an embodiment of the present invention. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0084] Example 1

[0085] like Figure 1 As shown, the present invention provides a subsonic target drone launch control method, the method comprising:

[0086] Acquire target drone training mission data and extract specific scenario parameters to determine launch trajectory configuration requirements;

[0087] Based on the launch trajectory configuration requirements, dynamic environment adaptation parameters are analyzed to obtain the timing control accuracy requirements;

[0088] When the timing control accuracy requirement is higher than the threshold, the corresponding instructions are extracted and combined from the preset instruction module library to generate the initial launch control strategy;

[0089] For the initial launch control strategy, analyze the dynamic changes in complex environment applications and determine the need for real-time strategy adjustment.

[0090] If there is a need to adjust the strategy in real time, the control system is flexibly configured according to the dynamic environment adaptation parameters to obtain the adjusted launch trajectory.

[0091] By adjusting the launch trajectory and combining it with scenario simulation data, a target drone control scheme for the training mission is generated.

[0092] In this embodiment, the method for acquiring target drone training task data and extracting specific scenario parameters to determine launch trajectory configuration requirements includes:

[0093] By extracting task data from the target machine training system and analyzing the diverse target information contained therein, preliminary task target classification results are obtained.

[0094] Based on the task objective classification results, data is broken down for specific scenarios under each category to determine the key scenario parameters corresponding to each scenario.

[0095] By using a pre-established scene parameter database, the extracted scene parameters are matched to obtain the parameter mapping relationship related to the launch trajectory;

[0096] If there is missing or abnormal data in the parameter mapping relationship, the abnormal data is repaired by data cleaning tools, and it is determined whether the repaired parameters meet the preset threshold range.

[0097] By combining the repaired parameter data with the launch trajectory configuration model, a trajectory configuration scheme adapted to the current mission is calculated, and the final trajectory configuration result is determined.

[0098] After obtaining the final trajectory configuration result, the matching degree between the configuration result and the task data is verified to obtain the verified configuration applicability conclusion.

[0099] If the verified suitability conclusion does not meet the preset standard, the configuration model is optimized by adjusting the parameter weights, and it is determined whether the optimized configuration meets the requirements of the task objective.

[0100] Specifically, the method of using a pre-established scene parameter database to match the currently extracted scene parameters and obtain the parameter mapping relationship related to the launch trajectory includes:

[0101] The radiance received at the entrance pupil of a spaceborne sensor in the mid-infrared band mainly includes surface radiation and path radiation. The contribution of a point on the surface to the spectral radiance of the entrance pupil can be decomposed as follows:

[0102]

[0103] In the formula, L λ,T Total radiance; and These are surface thermal radiation and path radiation, respectively; τ λ ε represents spectral transmittance. λ λ is the spectral emissivity; h is Planck's constant; c is the speed of light; λ is the wavelength (μm); k is the Boltzmann constant; and T is the thermodynamic temperature. Solar radiation is concentrated below 3.5 μm; the longer the wavelength, the smaller its contribution to the total radiance.

[0104] A brief description of the forward calculation process for surface radiance when other factors remain relatively constant in the same scene:

[0105] L λ =f λ (ε λ ,T,H,θ1,θ2);

[0106] In the formula, H is the altitude of the ground point; θ1 is the zenith angle from the sun to the ground point; θ2 is the zenith angle from the ground point to the sensor; f λ The mapping from the above five elements to radiance includes the Planck function, which is a complex nonlinear function. The emissivity inversion problem can be expressed as:

[0107] ε λ =g λ (L λ ,T,H,θ1,θ2);

[0108] In the formula, the function g λ The mapping relationship between various factors affecting the surface emissivity inversion results and the surface emissivity itself was characterized. Based on the above analysis, a method is proposed to represent this mapping relationship using an artificial neural network with extremely strong fitting ability, thereby improving the surface emissivity resolution from the km level to about 100m. The basic principle is to construct a training dataset using points within the field of view that have known surface temperature, altitude, solar zenith angle, sensor zenith angle, apparent radiance, and surface emissivity, and train the parameters of the neural network to obtain an approximate mapping g. λ When the five parameters of the nearest points in the input scene are known, the forward calculation outputs their surface emissivity.

[0109] The contribution of a point on the Earth's surface to the entrance pupil spectral radiance shows that the background entrance pupil radiance depends on the surface temperature, emissivity, atmospheric transmittance, and path radiation. Emissivity is an essential property of matter, while atmospheric transmittance and path radiation are related to parameters such as atmospheric models and the geometric path from the surface point to the sensor.

[0110] Specifically, the configuration model for the launch trajectory is as follows:

[0111] MPC control system:

[0112] x(k+1) = f(x(k), u(k));

[0113] At time k, the system state x(k) is obtained. Based on this, the optimal control sequence u(x(k)) is calculated in the prediction region N, and applied to the sampling period [k, k+1], and then executed cyclically.

[0114] Model predictive control consists of three main processes: prediction, online optimization, and rolling time-domain execution. Essentially, it is also a type of optimal control problem, mathematically structured as follows:

[0115] Control objective:

[0116]

[0117] Cost function:

[0118]

[0119] Optimal control structure:

[0120]

[0121] x u(k+1)=f(x u (k),u(k));

[0122] x u (0) = x0;

[0123] u(k)∈U,x u (k)∈X;

[0124] In the formula, u(k) is the control input; This represents the weighted square norm; Q is the output weight matrix; f(x) represents the weighted square norm; R is the input weight matrix; u (k),u(k)) is the system model; x u For the system output; x r u r The system's reference signal is given by: u = , where u is the thrust vector direction; r is the inertial position; x = , where x is the thrust vector direction; u is the thrust vector direction; r is the inertial position; x is the thrust vector direction. u (k) represents the system state in the direction of the thrust vector; x0 represents the initial system state in the direction of the thrust vector; U and X are real numbers.

[0125] In this embodiment, the method for obtaining the timing control accuracy requirements by analyzing dynamic environment adaptation parameters based on launch trajectory configuration requirements includes:

[0126] By extracting key data from the configuration requirements of the launch trajectory, initial environmental variables and trajectory parameters are obtained, and the preliminary range of adaptation parameters is determined.

[0127] Based on the preliminary adaptation parameter range, combined with real-time monitoring data in the dynamic environment, the impact of environmental changes on the trajectory is analyzed, and the characteristic distribution of environmental changes is obtained.

[0128] Using the results of the characteristic distribution, parameters are adjusted in response to environmental changes, updated values ​​of the adaptation parameters are derived, and it is determined whether the constraints of the time-series control are met.

[0129] If the updated adaptation parameters meet the constraints of timing control, the adjusted parameters will be applied to trajectory optimization to obtain optimized trajectory data.

[0130] If the conditions are not met, return to the parameter adjustment stage and reanalyze the environmental change characteristics;

[0131] By combining the optimized trajectory data with the control strategy, specific instructions for timing control are generated, and the method for achieving control accuracy is determined.

[0132] The generated timing control instructions are obtained, and the deviations during the execution of the instructions are analyzed in conjunction with the real-time monitored data stream to obtain the verification results of the accuracy requirements.

[0133] Based on the verification results, key parameters in the control strategy were adjusted, the instruction generation logic was optimized, and the final timing control accuracy scheme was derived.

[0134] In this embodiment, the method for generating the preset instruction module library includes:

[0135] If the timing control accuracy is higher than the preset threshold, the matching instruction is extracted from the instruction module library through the data processing flow to obtain a preliminary instruction set;

[0136] The initial instruction set is filtered through a module matching mechanism to determine the compatibility of the instructions with the current control task and to identify a valid instruction subset.

[0137] The effective instruction subset is sorted and reorganized using instruction generation rules to generate an optimized combination of control instructions;

[0138] Based on the optimized combination of control instructions according to the control execution efficiency evaluation, it is determined whether the real-time requirements are met, and the sequence of execution instructions is obtained.

[0139] The sequence of executed instructions is adjusted by an instruction combination optimization algorithm to determine the final combination of control instructions;

[0140] Obtain the module library update data, determine whether the instruction module library needs to be updated, and obtain the updated instruction module library;

[0141] For the updated instruction module library, a data processing procedure is used to verify its integrity and determine the preset instruction module library.

[0142] In this embodiment, when the timing control accuracy requirement is higher than a threshold, the method for extracting and combining corresponding instructions from a preset instruction module library to generate an initial launch control strategy includes:

[0143] Obtain a modular instruction set, determine the instruction type and priority by parsing the instruction content, and obtain a categorized instruction sequence;

[0144] Control instructions are extracted from the classified instruction sequence, and the instructions are combined using preset logical rules to generate an initial control instruction set.

[0145] For the initial control instruction set, the decision tree algorithm is applied to analyze the dependencies between instructions, determine the execution order of instructions, and obtain an ordered sequence of control instructions.

[0146] Based on the ordered sequence of control instructions, a state machine model is used to simulate the launch control process, determine the triggering conditions of each instruction, and generate a preliminary launch strategy.

[0147] Modular design parameters are extracted from the initial launch strategy. If the parameters meet the preset threshold, the strategy structure is optimized through dynamic programming algorithm to obtain the optimized launch control strategy.

[0148] For the optimized launch control strategy, obtain feedback on modular design capabilities, determine the compatibility between the strategy and the modular design, and generate the final launch control strategy.

[0149] Specifically, for the initial control instruction set, methods for applying decision tree algorithms to analyze the dependencies between instructions, determine the instruction execution order, and obtain an ordered sequence of control instructions include:

[0150] A finite set of atomic instructions is predefined. In the instruction parsing part, a large model is used to parse the text instructions into subtasks composed of atomic instructions. In the instruction generation part, a frame-by-frame instruction generation network is designed based on the conditional variational autoencoder (cVAE). In the instruction generation stage, different instruction features are considered to realize the instruction sequence generation task.

[0151] Among them, instruction parsing based on a large model:

[0152] This invention employs the GPT-3-davinci interface as a Large Language Model (LLM) planner. For a natural language instruction C, the goal is to obtain a decision tree R representing the instruction's completion process and the final state S upon instruction completion. T .

[0153] Because the output of large models has a high degree of randomness, this invention is based on prompting engineering to design reasonable prompts so that the planner can generate acceptable answers to the task.

[0154] The prompt includes four parts: task description, output paradigm description, self-check prompts for generated results, and sample. 1) The content describes the specific task of instruction decomposition: parsing the input instructions with the scene into a formatted output that drives the character to complete the instruction task. The output consists of two parts: first, it includes the character, the scene, the different states that the subsonic target drone may exist in the scene, and the next action the character needs to take in that state to complete the instruction, defined as D. D Secondly, the final state that the subsonic target drone in the scene needs to reach to complete the command is defined as D. S 2) The paradigm defines the specific content of the formatted output, including the structural format and a detailed description of the subsonic target drone and its possible states. In particular, it describes the scene category and the subsonic target drones that may exist in the scene, in conjunction with the scene content. 3) The result self-check prompt requires the planner to confirm that the output results meet the above conditions through a self-check after generating the parsed results. 4) Samples are provided as a reference to enable the planner to generate similar results. For the instruction parsing description D of the large model output... DMerging subsonic target drones targeting the same interaction i The different state judgments are merged into a decision tree D, such as... Figure 2 As shown.

[0155] A decision tree contains decision nodes and terminal nodes. A decision node represents the i-th interactive subsonic target machine that needs to be judged to complete the instruction task. i According to its state s i Different values ​​result in different decision nodes or termination nodes. The termination node contains the character's instruction b and its corresponding sequence style sty. Instruction b contains the character's interaction style o. i and interactive actions ITR j Yes, or the character's movement animation MV. p The set of interactive subsonic target machines is obj, the set of interactive actions is itr, the set of mobile actions is mv, and the set of sequential styles is sty.

[0156] Description of the final state D S The text describes the character, the scene, and the final state of the subsonic target drone after completing its mission. Combined with scene awareness, this is transformed into the symbolic description S defined above. T It includes the character's position and orientation, as well as the status, position, and orientation information of each subsonic target drone.

[0157] This invention designs a frame-by-frame action generation network based on a cVAE structure, such as... Figure 3 As shown. The action in the i-th frame is defined as:

[0158]

[0159] In the formula, This indicates the position, rotation, and speed of the subsonic target drone relative to the ground; This represents the position of the subsonic target drone relative to the ground in the next second, and j represents the number of commands given to the virtual character's subsonic target drone. T time steps are sampled within a 2-second time window between each frame [-1, 1] s. This indicates the position and direction of movement of the current frame relative to the ground in the previous frame; These represent approximate values ​​of the current frame's position relative to the ground in the previous frame and its direction of movement, respectively. This represents the one-hot encoded vector of the instruction label at each time step t. The instruction label includes content and style labels. The 10 content labels and 8 style labels of the instruction are encoded uniformly, i.e., n a =10 × 8 = 80; and Indicates the target's location and orientation; It is the one-hot encoded vector of the action label to be executed at each time step t; Labels indicating the five end connection points of the subsonic target drone.

[0160] The encoder encodes the subsonic target representation O into a latent space vector O through a three-layer fully connected network. lat At the same time, set the current frame state X i and the previous frame's state X i-1 Similarly, it is encoded into a low-dimensional latent space vector Z through a three-layer fully connected network. mot The decoder consists of a prediction network and a gating network, both of which are three-layer fully connected networks. The prediction network is based on the latent space variable O of the encoder output. lat and Z mot By combining the instruction vector from the previous frame, the instructions for the virtual character in the next frame are predicted. Gated networks with Z mot and X i-1 As input, dynamically generate mixed weight coefficients ω = {ω1, ω2, ...}, which are mixed with expert weights to obtain the weights of the prediction network. The instruction generation network is generated in an end-to-end manner.

[0161] In this embodiment, for the initial launch control strategy, the dynamic changes in complex environment applications are analyzed to determine the need for real-time strategy adjustment, including:

[0162] Acquire multi-source data in complex environments, continuously collect key indicators in the environment, and obtain a preliminary dataset of environmental status through sensor networks and data transmission modules;

[0163] Based on the preliminary dataset, data preprocessing techniques are used to clean and standardize the collected information. If outliers are detected in the data, they are repaired using interpolation methods to determine the processed environmental feature set.

[0164] By analyzing the dynamic trend of the processed environmental feature set, the feature set is modeled using time series analysis. If the trend exceeds a preset threshold, it is marked as a high-risk state, and the evaluation result of the trend is obtained.

[0165] Based on the assessment results of the changing trends, the parameters of the initial launch control strategy are compared. If the assessment results show a high-risk state, the strategy adjustment process is triggered to determine the priority sequence of the adjustment direction.

[0166] By adjusting the priority sequence of directions and combining it with real-time monitored data streams, the support vector machine algorithm is used to optimize the calculation of policy parameters, resulting in an optimized combination of control parameters.

[0167] Based on the optimized combination of control parameters, a set of instructions for real-time adjustment is generated to update the parameters of the launch control system. The updated system response status is then determined through the instruction issuance module.

[0168] By continuously collecting feedback data based on the system response status, the feedback data is compared with the expected target. If there is a deviation, the optimization calculation process is executed repeatedly to obtain the final strategy adaptability result.

[0169] In this embodiment, if a real-time adjustment strategy is required, the control system is flexibly configured based on dynamic environmental adaptation parameters to obtain the adjusted launch trajectory, including:

[0170] If the dynamic environment changes, environmental parameters are collected through a sensor network to generate real-time environmental data;

[0171] Based on real-time environmental data, a preset classification algorithm is used to determine whether a strategy adjustment requirement has been triggered, and the adjustment trigger status is determined.

[0172] If the adjustment trigger state is true, key parameters are extracted from the real-time environmental data to generate a parameter update set;

[0173] Based on the parameter update set, the control system configuration is adjusted using an optimization algorithm to obtain the updated system configuration.

[0174] The updated system configuration generates an initial launch trajectory, obtaining preliminary trajectory data.

[0175] Based on the preliminary trajectory data, simulation algorithms are used to verify the feasibility of the trajectory and generate an optimized launch trajectory;

[0176] The real-time monitoring module continuously tracks the optimized launch trajectory and obtains trajectory execution feedback.

[0177] In this embodiment, by combining the adjusted launch trajectory with scenario simulation effect data, a precise target drone control scheme for the training mission is generated, including:

[0178] By using dynamic data obtained from scenario simulation, a preliminary analysis of the launch path is conducted to obtain the basic dataset for path optimization;

[0179] Based on the basic dataset, the launch path is optimized and adjusted using path planning techniques to determine the adjusted path parameters;

[0180] By inputting the adjusted path parameters into the scene simulation system, feedback data on the simulation effect is obtained, and it is determined whether the simulation effect has reached the preset threshold.

[0181] If the simulation effect does not reach the preset threshold, the feedback data is further processed using data analysis techniques to obtain the key factors affecting the control accuracy.

[0182] Based on key factors, the target drone control logic is dynamically adjusted to determine a control instruction set that meets the mission training requirements.

[0183] By applying the control instruction set to the target drone control system, real-time operating data is obtained, and it is determined whether the operating data meets the accuracy requirements of the training scheme.

[0184] If the running data meets the accuracy requirements, then all data and parameters are integrated to formulate the final accurate training plan.

[0185] Example 2

[0186] The present invention also provides a subsonic target drone launch control system, the system being used to implement the method described in Embodiment 1, the system comprising: a mission data acquisition module, a scenario parameter extraction module, a timing control analysis module, an initial strategy generation module, a dynamic adjustment judgment module, and a precise scheme generation module;

[0187] The mission data acquisition module is used to acquire target drone training mission data and extract specific scenario parameters to determine launch trajectory configuration requirements;

[0188] The scenario parameter extraction module is used to analyze dynamic environment adaptation parameters based on the launch trajectory configuration requirements to obtain the timing control accuracy requirements;

[0189] The timing control analysis module is used to extract and combine corresponding instructions from the preset instruction module library to generate an initial launch control strategy when the timing control accuracy requirement is higher than the threshold.

[0190] The initial strategy generation module is used to analyze dynamic changes in complex environments and determine the need for real-time strategy adjustment based on the initial launch control strategy.

[0191] The dynamic adjustment judgment module is used to update the control system configuration flexibly according to the dynamic environment adaptation parameters if there is a need for real-time adjustment strategy, so as to obtain the adjusted launch trajectory.

[0192] The precision scheme generation module is used to generate target drone control schemes for training missions by combining the adjusted launch trajectory with scenario simulation effect data.

[0193] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for controlling the launch of a subsonic target drone, characterized in that, The method includes: Acquire target drone training mission data and extract specific scenario parameters to determine launch trajectory configuration requirements; Based on the launch trajectory configuration requirements, dynamic environment adaptation parameters are analyzed to obtain the timing control accuracy requirements; When the timing control accuracy requirement is higher than the threshold, the corresponding instructions are extracted and combined from the preset instruction module library to generate the initial launch control strategy; For the initial launch control strategy, analyze the dynamic changes in complex environment applications and determine the need for real-time strategy adjustment. If there is a need to adjust the strategy in real time, the control system is flexibly configured according to the dynamic environment adaptation parameters to obtain the adjusted launch trajectory. By adjusting the launch trajectory and combining it with scenario simulation data, a target drone control scheme for the training mission is generated.

2. The method according to claim 1, characterized in that, Methods for acquiring target drone training mission data and extracting specific scenario parameters to determine launch trajectory configuration requirements include: By extracting task data from the target machine training system and analyzing the diverse target information contained therein, preliminary task target classification results are obtained. Based on the task objective classification results, data is broken down for specific scenarios under each category to determine the key scenario parameters corresponding to each scenario. By using a pre-established scene parameter database, the extracted scene parameters are matched to obtain the parameter mapping relationship related to the launch trajectory; If there is missing or abnormal data in the parameter mapping relationship, the abnormal data is repaired by data cleaning tools, and it is determined whether the repaired parameters meet the preset threshold range. By combining the repaired parameter data with the launch trajectory configuration model, a trajectory configuration scheme adapted to the current mission is calculated, and the final trajectory configuration result is determined. After obtaining the final trajectory configuration result, the matching degree between the configuration result and the task data is verified to obtain the verified configuration applicability conclusion. If the verified suitability conclusion does not meet the preset standard, the configuration model is optimized by adjusting the parameter weights, and it is determined whether the optimized configuration meets the requirements of the task objective.

3. The method according to claim 2, characterized in that, Methods for obtaining parameter mapping relationships related to launch trajectories by matching currently extracted scene parameters with a pre-established scene parameter database include: The radiance received at the entrance pupil of a spaceborne sensor in the mid-infrared band mainly includes surface radiation and path radiation. The contribution of a point on the surface to the spectral radiance of the entrance pupil can be decomposed as follows: In the formula, L λ,T Total radiance; and These are surface thermal radiation and path radiation, respectively; τ λ ε represents spectral transmittance. λ λ is the spectral emissivity; h is Planck's constant; c is the speed of light; λ is the wavelength; k is Boltzmann's constant; T is the thermodynamic temperature; The forward calculation process for the radiance of a surface point is as follows: L λ =f λ (e λ ,T,H,θ1,θ2); In the formula, H is the altitude of the ground point; θ1 is the zenith angle from the sun to the ground point; θ2 is the zenith angle from the ground point to the sensor; f λ The mapping from five elements to radiance includes the Planck function, which is a complex nonlinear function; while the emissivity inversion problem is expressed as: e λ =g λ (L λ ,T,H,θ1,θ2); In the formula, the function g λ It characterizes the mapping relationship between various factors affecting the surface emissivity inversion results and the surface emissivity.

4. The method according to claim 1, characterized in that, The methods for obtaining the timing control accuracy requirements based on launch trajectory configuration requirements and dynamic environment adaptation parameters include: By extracting key data from the configuration requirements of the launch trajectory, initial environmental variables and trajectory parameters are obtained, and the preliminary range of adaptation parameters is determined. Based on the preliminary adaptation parameter range, combined with real-time monitoring data in the dynamic environment, the impact of environmental changes on the trajectory is analyzed, and the characteristic distribution of environmental changes is obtained. Using the results of the characteristic distribution, parameters are adjusted in response to environmental changes, updated values ​​of the adaptation parameters are derived, and it is determined whether the constraints of the time-series control are met. If the updated adaptation parameters meet the constraints of timing control, the adjusted parameters will be applied to trajectory optimization to obtain optimized trajectory data. If the conditions are not met, return to the parameter adjustment stage and reanalyze the environmental change characteristics; By combining the optimized trajectory data with the control strategy, specific instructions for timing control are generated, and the method for achieving control accuracy is determined. The generated timing control instructions are obtained, and the deviations during the execution of the instructions are analyzed in conjunction with the real-time monitored data stream to obtain the verification results of the accuracy requirements. Based on the verification results, key parameters in the control strategy were adjusted, the instruction generation logic was optimized, and the final timing control accuracy scheme was derived.

5. The method according to claim 1, characterized in that, The methods for generating a preset instruction module library include: If the timing control accuracy is higher than the preset threshold, the matching instruction is extracted from the instruction module library through the data processing flow to obtain a preliminary instruction set; The initial instruction set is filtered through a module matching mechanism to determine the compatibility of the instructions with the current control task and to identify a valid instruction subset. The effective instruction subset is sorted and reorganized using instruction generation rules to generate an optimized combination of control instructions; Based on the optimized combination of control instructions according to the control execution efficiency evaluation, it is determined whether the real-time requirements are met, and the sequence of execution instructions is obtained. The sequence of executed instructions is adjusted by an instruction combination optimization algorithm to determine the final combination of control instructions; Obtain the module library update data, determine whether the instruction module library needs to be updated, and obtain the updated instruction module library; For the updated instruction module library, a data processing procedure is used to verify its integrity and determine the preset instruction module library.

6. The method according to claim 5, characterized in that, When the timing control accuracy requirement exceeds a threshold, the method for extracting and combining corresponding instructions from a preset instruction module library to generate an initial launch control strategy includes: Obtain a modular instruction set, determine the instruction type and priority by parsing the instruction content, and obtain a categorized instruction sequence; Control instructions are extracted from the classified instruction sequence, and the instructions are combined using preset logical rules to generate an initial control instruction set. For the initial control instruction set, the decision tree algorithm is applied to analyze the dependencies between instructions, determine the execution order of instructions, and obtain an ordered sequence of control instructions. Based on the ordered sequence of control instructions, a state machine model is used to simulate the launch control process, determine the triggering conditions of each instruction, and generate a preliminary launch strategy. Modular design parameters are extracted from the initial launch strategy. If the parameters meet the preset threshold, the strategy structure is optimized through dynamic programming algorithm to obtain the optimized launch control strategy. For the optimized launch control strategy, obtain feedback on modular design capabilities, determine the compatibility between the strategy and the modular design, and generate the final launch control strategy.

7. The method according to claim 6, characterized in that, For an initial control instruction set, methods for analyzing the dependencies between instructions, determining the instruction execution order, and obtaining an ordered sequence of control instructions using decision tree algorithms include: A finite set of atomic instructions is predefined. In the instruction parsing part, a large model is used to parse the text instructions into subtasks composed of atomic instructions. In the instruction generation part, a frame-by-frame instruction generation network is designed based on the conditional variational autoencoder (cVAE). In the instruction generation stage, different instruction features are considered to realize the instruction sequence generation task.

8. The method according to claim 7, characterized in that, in, Methods for instruction parsing based on large models include: Using the GPT-3-davinci interface as the Large Language Model (LLM) planner, for a natural language instruction C, the goal is to obtain a decision tree R representing the task completion process and the final state S of the completed instruction. T ; Based on prompting engineering, design reasonable prompts so that the planner can generate acceptable answers to the task; The prompt content includes four parts: task content description, output paradigm description, self-check prompts for generated results, and examples. 1) The content describes the specific task of instruction decomposition: parsing the input instructions with the scene into a formatted output that can drive the character to complete the instruction task; the output consists of two parts: first, it includes the character, the scene, the different states that the subsonic target drone may exist in the scene, and the next action the character needs to take in that state to complete the instruction, defined as D. D Secondly, the final state that the subsonic target drone in the scene needs to reach to complete the command is defined as D. S ;2) The paradigm defines the specific content of the formatted output, including the structural format and a specific description of the subsonic target drone and its possible states. In particular, it describes the scene category and the subsonic target drones that may exist in the scene in conjunction with the scene content;3) The result self-check prompt requires the planner to confirm that the output result meets the above conditions through a self-check after generating the parsed result;4) Samples are given as a reference to enable the planner to generate similar results. The instruction parsing description D for the large model output D Merging subsonic target drones targeting the same interaction i The different state judgments are merged into a decision tree D. The decision tree contains decision nodes and terminal nodes. The decision node represents the i-th interactive subsonic target machine o that needs to be judged to complete the instruction task. i According to its state s i Different values ​​result in different decision nodes or termination nodes; the termination node contains the character's instruction b and its corresponding sequence style sty; instruction b contains the character's interaction during the interaction. i and interactive actions ITR j Yes, or the character's movement animation MV. p The set of interactive subsonic target machines (obj), the set of interactive actions (itr), the set of movement actions (mv), and the set of sequential style (sty); the description of the final state (D) S The text describes the character, the scene, and the final state of the subsonic target drone after completing its mission. Combined with scene awareness, this is transformed into the symbolic description S defined above. T It includes the character's position and orientation, as well as the status, position, and orientation information of each subsonic target drone.

9. The method according to claim 7, characterized in that, The method for designing a frame-by-frame action generation network based on cVAE architecture includes: the action of the i-th frame is defined as: In the formula, This indicates the position, rotation, and speed of the subsonic target drone relative to the ground; represents the position of the subsonic target drone relative to the ground in the next second, and j represents the number of commands for the virtual character subsonic target drone; t time steps are sampled in a 2s time window between each frame [-1,1]s; These represent the position of the current frame relative to the ground in the previous frame and the direction of movement, respectively. These represent approximate values ​​of the current frame's position relative to the ground in the previous frame and its direction of movement, respectively. This represents the one-hot encoded vector of the instruction label at each time step t. The instruction label includes content and style labels. The 10 content labels and 8 style labels of the instruction are encoded uniformly, i.e., n a =10 × 8 = 80; and Indicates the target's location and orientation; It is the one-hot encoded vector of the action label to be executed at each time step t; Labels indicating the five end connection points of the subsonic target drone; The encoder encodes the subsonic target representation O into a latent space vector O through a three-layer fully connected network. lat At the same time, set the current frame state X i and the previous frame's state X i-1 Similarly, it is encoded into a low-dimensional latent space vector Z through a three-layer fully connected network. mot ; The decoder consists of a prediction network and a gating network, both of which are three-layer fully connected networks. The prediction network is based on the latent space variable O of the encoder output. lat and Z mot By combining the instruction vector from the previous frame, the instructions for the virtual character in the next frame are predicted. Gated networks with Z mot and X i-1 As input, dynamically generate mixed weight coefficients ω = {ω1, ω2, ...}, which are mixed with expert weights to obtain the weights of the prediction network. The instruction generation network is generated in an end-to-end manner.

10. A subsonic target drone launch control system, said system being used to implement the method according to any one of claims 1-9, characterized in that, The system includes: a task data acquisition module, a scene parameter extraction module, a timing control analysis module, an initial strategy generation module, a dynamic adjustment judgment module, and a precise solution generation module; The task data acquisition module is used to acquire target drone training task data and extract specific scenario parameters to determine launch trajectory configuration requirements; The scenario parameter extraction module is used to analyze dynamic environment adaptation parameters according to the launch trajectory configuration requirements to obtain the timing control accuracy requirements; The timing control analysis module is used to extract and combine corresponding instructions from the preset instruction module library to generate an initial launch control strategy when the timing control accuracy requirement is higher than the threshold. The initial strategy generation module is used to analyze the dynamic changes in complex environment applications and determine the need for real-time strategy adjustment based on the initial launch control strategy. The dynamic adjustment judgment module is used to update the flexible configuration of the control system according to the dynamic environment adaptation parameters if there is a need for real-time adjustment strategy, so as to obtain the adjusted launch trajectory. The precision scheme generation module is used to generate a target drone control scheme for the training mission by combining the adjusted launch trajectory with scenario simulation effect data.

Images