Spacecraft cluster target allocation method based on large language model
By combining large language models with physical rules and optimization algorithms, the complexity and security issues of target allocation in spacecraft swarms are solved, achieving efficient and safe target allocation decisions and improving the intelligence and adaptability of spacecraft swarms.
Patent Information
- Application Number
- CN202511603811.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-04
- Publication Date
- 2026-02-10
Smart Images
Figure CN121503222A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of spacecraft swarm collaborative control and artificial intelligence. It mainly relates to an autonomous collaborative mission planning method for spacecraft swarms based on a large language model, and in particular, an intelligent method that applies the natural language understanding and logical reasoning capabilities of a large language model to multi-spacecraft, multi-target mission allocation decisions. Background Technology
[0002] With the increasing demands for space exploration and applications, such as large-scale Earth observation, on-orbit servicing, space station maintenance, and deep space exploration, swarm systems composed of multiple spacecraft have become an important development direction for future space activities due to their significant advantages in flexibility, robustness, and functional distribution. One of the core technologies of spacecraft swarms is the multi-objective allocation problem, which involves rationally allocating a limited number of mission objectives to individual spacecraft within the swarm while satisfying various constraints, in order to optimize overall mission effectiveness.
[0003] Currently, traditional spacecraft swarm target allocation methods mainly rely on classical methods such as mathematical programming, intelligent optimization algorithms, and game theory.
[0004] 1. Mathematical programming methods: such as mixed-integer linear programming and nonlinear programming. These methods can find the theoretically optimal solution when the model is accurate, but their computational complexity increases exponentially with the size of the problem (number of spacecraft and targets), making it difficult to meet the online, real-time decision-making requirements of space missions. Furthermore, they are highly dependent on the accuracy of the model and cannot effectively handle the fuzziness and uncertainty inherent in missions.
[0005] 2. Intelligent optimization algorithms: such as genetic algorithms, particle swarm optimization, and ant colony optimization. These methods have strong global search capabilities and are suitable for complex nonlinear problems. However, they generally suffer from slow convergence speed, susceptibility to getting trapped in local optima, and significant impact of parameter settings on performance, limiting their application in dynamic environments requiring rapid response.
[0006] 3. Game theory-based approach: This method treats spacecraft in a cluster as rational agents and uses game equilibrium to allocate tasks. While this approach reflects individual autonomy, convergence is difficult to guarantee, and it places strict requirements on the behavioral models of the agents.
[0007] Existing methods are typically designed for specific task scenarios and lack generalization ability. When task objectives or constraints change, it is often necessary to remodel and redesign the algorithm, resulting in poor adaptability. Traditional algorithms lack the ability to understand the task context, physical common sense, and high-level tactical intentions, and are unable to perform deep logical reasoning and causal judgment.
[0008] Large Language Models (LLMs), such as the GPT series and BERT, have achieved revolutionary breakthroughs in the field of natural language processing in recent years, demonstrating powerful semantic understanding, knowledge integration, logical reasoning, and code generation capabilities. This provides a completely new technical path for solving the aforementioned challenges. However, directly applying LLMs to the high-reliability and high-safety requirements of spacecraft control faces problems such as uncontrollable output results, lack of physical common sense, and the potential for "illusions." Therefore, how to deeply integrate the cognitive advantages of LLMs with the precise physical rules and safety requirements of the aerospace field to design a reliable and efficient target allocation framework has become an urgent technical challenge. Summary of the Invention
[0009] The primary objective of this invention is to overcome the shortcomings of existing technologies and provide a target allocation method for spacecraft swarms based on a large language model. This method aims to enable direct and efficient command of spacecraft swarms using natural language, improving the intelligence, real-time performance, and adaptability of target allocation decisions. Simultaneously, by introducing physical rule constraints, it ensures the security and feasibility of the generated schemes.
[0010] Technical Solution: This invention proposes a spacecraft swarm target allocation method based on a large language model (LLM). It utilizes a large language model (LLM) to generate different models to handle the spacecraft swarm target allocation task planning problem. The method is characterized by the following steps:
[0011] Step 1: Obtaining Task Commands and Cluster Status
[0012] It receives mission instructions from command personnel or the command system. These instructions include mission objectives, priorities, and constraints described in natural language. Simultaneously, it acquires real-time status information of the spacecraft cluster, primarily including: the number of orbital elements, attitude, fuel reserves, payload operational status, satellite value, inter-satellite relative positions, and communication link status for each spacecraft.
[0013] Step 2: Information Structuring and Prompt Word Construction
[0014] The natural language task instructions and structured cluster status information obtained in step 1 are fused and preprocessed to construct a structured prompt word.
[0015] Step 3: Preliminary Scheme Reasoning Based on Large Language Model
[0016] The prompts constructed in step 2 are input into a large language model that has been fine-tuned with domain knowledge. This large language model is responsible for deep semantic understanding and task reasoning of the task instructions, and outputs a preliminary target allocation scheme.
[0017] The fine-tuning process of the large language model is key to the domain adaptation of this invention, and specifically includes:
[0018] 1. Domain Corpus Construction
[0019] The main tasks include constructing a model for calculating the fuel consumption cost of transferring between spacecraft orbits, collecting orbital data of both target spacecraft, and designing the objective function and constraints for optimizing the target allocation of spacecraft swarms.
[0020] (1) Based on the six elements of satellite orbit parameters, construct the initial position parameter matrix S of our spacecraft. i Let D represent the initial position six-element parameter matrix of the i-th friendly spacecraft, and construct the initial position six-element parameter matrix D of the enemy target. j The matrix representing the six parameters of the initial position of the j-th enemy target:
[0021] S = S i , i = 1, ... m
[0022] D = D j j = 1, ... n
[0023] (2) Calculate the fuel cost of orbital transfer between enemy and friendly spacecraft. Using the Lambert orbital transfer method, design a method to calculate the maneuver cost of enemy and friendly spacecraft and solve for the maneuver time Δt and the speed increment Δv required for the maneuver.
[0024] The specific calculation method for the maneuvering cost of our spacecraft is as follows: At the initial moment, the position vector and velocity vector of our spacecraft are denoted as r. p0 and v p0 The position vector and velocity vector of the enemy target spacecraft are denoted as r. e0 and v e0 Given that after a finite time Δt, the position and velocity vectors of the two spacecraft are the same, denoted as r1 and v1, then the initial and final velocities of the Lambert transfer orbit are v1 and v2, respectively. t0 and v t1 Then the velocity increment can be expressed as:
[0025]
[0026] To determine the optimal maneuver scheme within the operational timeframe of each spacecraft, i.e., the maneuver scheme with the minimum speed increment, the specific implementation is as follows: Select the maneuver time t as the iteration variable, iterate through the mission window to calculate the speed increment Δv corresponding to different times t, compare Δv to select the optimal maneuver scheme, and record and save the optimal maneuver scheme data, including the speed increment Δv required for the optimal maneuver. min and the corresponding maneuver time t trans , recorded as [t trans Δv minThat is, fuel consumption is c. ij .
[0027] (3) Establish a heterogeneous weapon system model for a spacecraft cluster. Assume that our side has m spacecraft, and define the weapon type set as W = {w1, w2, ..., w...} m The process involves a one-to-many task allocation, ultimately determining the optimal allocation scheme for m spacecraft. Assume there are n target spacecraft, defined as the target set T = {t1, t2, ..., tn}. n The operation requires m spacecraft to conduct strikes, and each target spacecraft must be struck at least once by our spacecraft. Our spacecraft can strike at least one enemy target spacecraft. The constraints are expressed as follows:
[0028]
[0029] Where k represents the maximum number of enemy target spacecraft that a single friendly spacecraft can strike.
[0030] Assuming all enemy targets are attacked, x ij =1 indicates that the i-th spacecraft strikes the j-th enemy target; otherwise, x ij =0. Therefore, when attacking the j-th target, its overall damage probability is:
[0031]
[0032] The expected operational benefit of attacking all n targets is:
[0033]
[0034] Considering that when all our spacecraft carry out strike missions, the mission balance should not be too high, otherwise resources will be wasted. The standard deviation of the number of enemy targets struck by each of our spacecraft is calculated to measure the balance of the final target allocation scheme, and it is expressed as follows:
[0035]
[0036] in,
[0037] The objective function to be optimized is:
[0038]
[0039] Here, f1 represents the expected operational benefits of the weapon striking the target; f2 represents the fuel consumption of the weapon during the entire process of striking the target; and f3 represents the overall mission load balance when assessing each weapon striking all targets.
[0040] 2. Supervised fine-tuning
[0041] The general-purpose LLM was fine-tuned using the corpus described above, enabling it to gain a deeper understanding of the Lambert problem and specialized concepts such as computational speed increment Δv and J2 perturbation, as well as the underlying physical principles.
[0042] 3. Intelligent Algorithm Optimization
[0043] Furthermore, an intelligent optimization algorithm is introduced to construct an optimization objective function based on indicators such as task balance, fuel consumption, and combat benefits. This optimizes the model so that its output scheme, while satisfying natural language instructions, is more inclined to select the physically feasible scheme with the optimal fuel consumption and the highest combat benefits.
[0044] Step 4: Physical rule verification and numerical optimization
[0045] Since large language models may produce solutions that violate precise physical quantity constraints or are not optimal, this step performs safety verification and performance improvement on the preliminary solution.
[0046] 1. Feasibility Verification: The preliminary scheme output by LLM is used to calculate the orbital maneuver fuel consumption required for each spacecraft to perform the corresponding target mission sequence using the Lambert problem, so as to obtain its velocity increment Δv and mission time Δt, and verify whether the scheme meets all constraints, including fuel constraints, time constraints, etc.
[0047] 2. Solution Optimization: If the initial solution is infeasible or suboptimal, a numerical optimizer (NSKOA algorithm) is initiated. This optimizer uses the initial solution generated by LLM as a high-quality initial solution and performs local search and fine-tuning within its neighborhood. The optimization objective function can be to minimize fuel consumption, mission balance, and maximize combat benefits, with constraints being all the aforementioned physical and mission constraints. This design can significantly accelerate the convergence speed of the optimizer and avoid getting trapped in local optima.
[0048] Step 5: Final instruction generation and distribution
[0049] The final target allocation scheme, after verification and optimization, will be converted into specific command sequences executable by each spacecraft, including orbital maneuver commands, payload power-on / off commands, and data transmission commands. These commands will then be distributed to the corresponding spacecraft in the cluster via space-to-ground links or inter-satellite communication networks.
[0050] Step 6: Task Execution and Status Monitoring
[0051] Each spacecraft receives and executes commands, and feeds back its execution status and mission results to the system in real time, forming a closed loop. The system can then determine whether target reassignment is necessary based on the feedback information.
[0052] Beneficial effects:
[0053] This invention solves the target allocation problem for spacecraft swarms using a large language model. It primarily enables direct command of the spacecraft swarm using natural language, significantly lowering the operational threshold and improving command efficiency, allowing non-command personnel to perform complex target allocation task planning. The forward inference process of the large language model is extremely short, providing a high-quality initial solution for large-scale swarm target allocation problems. Combined with rapid optimization starting from this initial solution, the overall decision-making speed is far faster than traditional optimization algorithms that start from scratch. By introducing an independent physical rule verification and optimization module, the inherent uncertainty of LLM as a generative model is effectively addressed, ensuring the physical feasibility and optimality of the final output solution, meeting the requirements of space mission planning. Attached Figure Description
[0054] Figure 1 The flowchart below shows the target allocation algorithm for spacecraft swarms based on a large language model according to the present invention. Detailed Implementation
[0055] 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.
[0056] This invention discloses a target allocation method for spacecraft swarms based on a large language model. The method first receives natural language task instructions and swarm status, and constructs structured prompts. Then, it inputs these prompts into a large language model fine-tuned with aerospace knowledge to generate a preliminary target allocation scheme. Next, using a physics-based verification and optimization module, combined with the Lambert problem and numerical optimization, the preliminary scheme is validated for feasibility and its performance improved, generating final executable instructions. Finally, the instructions are distributed to the swarm spacecraft for execution. To highlight the objectives, technical solutions, and advantages of this invention, the following detailed description, in conjunction with specific embodiments, further illustrates the invention.
[0057] like Figure 1 As shown, this invention proposes a spacecraft cluster target allocation method based on a large language model, mainly involving the construction of a domain corpus, including:
[0058] An optimization model is constructed based on the decision variables, constraints, and optimization objectives of spacecraft cluster target allocation;
[0059] Based on the optimization model, the Lambert problem model was used to calculate the fuel consumption and transfer time for orbital transfers between spacecraft.
[0060] A large language model fine-tuned with deep domain knowledge can understand the deeper meaning of instructions and reason based on its internalized knowledge to output a structured initial target allocation scheme. Combining the Lambert problem and optimization models, the orbital maneuver fuel consumption, transfer time, mission balance, and operational benefits required for each spacecraft to perform its assigned task in the initial target allocation scheme are evaluated, and the scheme is verified to satisfy all constraints. If the initial allocation scheme is feasible and the constraints are satisfied, it is adopted as the final scheme; otherwise, a numerical optimizer (NSKOA algorithm) is initiated. Using the NSKOA algorithm, an LLM-generated scheme is used as a high-quality initial solution, and fast local search and fine-tuning are performed within its neighborhood. The goal is to optimize the preset objective function while satisfying all constraints to obtain the final target allocation scheme.
[0061] Furthermore, the decision variable is the decision state of our spacecraft i against the target spacecraft j; where the decision state includes: executing the attack mission and not executing the attack mission;
[0062] The constraints include: the fuel consumption for orbital transfer between each spacecraft and the target cannot exceed its maximum fuel limit; the number of our spacecraft assigned to the target cannot exceed the existing number of our spacecraft; and each target can only be attacked once and cannot be repeatedly attacked.
[0063] The optimization objectives include: minimizing fuel consumption, minimizing mission balance, and maximizing operational benefits.
[0064] Specifically, in this embodiment, the decision variable is: x ij This represents the decision-making state of our spacecraft i against enemy target j;
[0065]
[0066] Constraints:
[0067] The fuel consumption for each weapon during orbital transfer between its weapon and its target must not exceed its maximum fuel limit.
[0068] The number of weapons allocated to a target cannot exceed the existing number of weapons.
[0069] Each target can only be attacked once and cannot be attacked repeatedly.
[0070] Optimize the objective function:
[0071]
[0072] Here, f1 represents the expected operational benefits of the weapon striking the target; f2 represents the fuel consumption of the weapon during the entire process of striking the target; and f3 represents the overall mission load balance when assessing each weapon striking all targets.
[0073] In this embodiment, the spacecraft swarm target allocation algorithm based on a large language model is summarized in the figure below. Figure 1 As shown, the system first receives natural language task instructions and cluster status, constructs a large language model based on prompt words and fine-tuning with aerospace knowledge, and generates an initial target allocation scheme. Then, combining the Lambert problem with the optimization model and target allocation constraints, it performs feasibility analysis and numerical optimization on the initial allocation scheme to generate the final target allocation scheme. The specific implementation steps of this embodiment are as follows:
[0074] 1. Determine the mathematical description of the research problem, identify the decision variables, constraints, and optimization objectives of the optimization problem, and construct the optimization model.
[0075] 2. Based on the constructed optimization model, fine-tune the large language model with spacecraft knowledge domain.
[0076] 3. Obtain task instructions and cluster status, and construct information structure and prompt words.
[0077] 4. Input the constructed prompt words and information data into the fine-tuned large language model to generate an initial target allocation scheme.
[0078] 5. For the initial target allocation scheme, combine the Lambert problem and the optimization model to calculate the initial target allocation scheme and determine whether it meets the constraints.
[0079] 6. If the constraints are met, the target allocation scheme is output directly; otherwise, the initial target allocation scheme is used as the initial solution and input into the numerical optimizer for solving.
[0080] 7. Determine if the set value has been reached during the iteration. If not, return to step 6 to continue optimization. Otherwise, stop optimization, output the final target allocation scheme, and continue fine-tuning the large language model based on the above allocation results.
[0081] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A target allocation method for spacecraft swarms based on a large language model, characterized in that, Specifically, the steps include the following: Step a: Obtain input data containing mission instructions described in natural language and real-time status information of the spacecraft cluster; Step b: The input data is processed in a structured manner to construct prompt words for the large language model; Step c: Input the prompt words into a large language model that has been fine-tuned with knowledge from the aerospace field. The large language model performs semantic understanding and task reasoning, and outputs an initial target allocation scheme. Step d: Input the initial target allocation scheme into the verification and optimization module. Based on the Lambert problem model and preset target allocation constraints, verify the feasibility of the initial scheme and perform numerical optimization to generate the final target allocation instruction. Step e involves distributing the final target allocation instruction to each spacecraft in our spacecraft's respective mission sequence to control them to execute the corresponding target mission.
2. The spacecraft cluster target allocation method based on a large language model according to claim 1, characterized in that: In step c, the large language model, fine-tuned with knowledge from the aerospace field, is implemented as follows: The fine-tuning process of the large language model is key to the domain adaptation of this invention, and specifically includes: (1) Domain Corpus Construction The main tasks include constructing a model for calculating the fuel consumption cost of transferring between spacecraft orbits, collecting orbital data of both target satellites, and designing the objective function and constraints for target allocation optimization in a spacecraft swarm. 1) Based on the six key parameters of satellite orbit, construct the initial position parameter matrix S of our spacecraft. i Let D represent the initial position six-element parameter matrix of the i-th friendly spacecraft, and construct the initial position six-element parameter matrix D of the enemy target. j The matrix representing the six parameters of the initial position of the j-th enemy target: S=S i ,i=1,...m D=D j ,j=1 ,...n 2) Calculate the fuel cost of orbital transfer between enemy and friendly spacecraft. Using the Lambert orbital transfer method, design a method to calculate the maneuver cost of enemy and friendly spacecraft and solve for the maneuver time Δt and the speed increment Δv required for the maneuver. The specific calculation method for the maneuver cost of our spacecraft is as follows: At the initial moment, the position vector and velocity vector of our spacecraft are denoted as r. p0 and v p0 The position vector and velocity vector of the enemy target spacecraft are denoted as r. e0 and v e0 Given that after a finite time Δt, the position and velocity vectors of the two spacecraft are the same, denoted as r1 and v1, then the initial and final velocities of the Lambert transfer orbit are v1 and v2, respectively. t0 and v t1 Then the velocity increment can be expressed as: To determine the optimal maneuver scheme within the operational timeframe of each spacecraft, i.e., the maneuver scheme with the minimum speed increment, the specific implementation is as follows: Select the maneuver time t as the iteration variable, iterate through the mission window to calculate the speed increment Δv corresponding to different times t, compare Δv to select the optimal maneuver scheme, and record and save the optimal maneuver scheme data, including the speed increment Δv required for the optimal maneuver. min and the corresponding maneuver time t trans , recorded as [t trans Δv min That is, fuel consumption is c. ij . 3) Establish a heterogeneous weapon system model for a spacecraft cluster. Assume our side has m spacecraft, and define the weapon type set as W = {w1, w2, ..., w...} m The process involves a one-to-many task allocation, ultimately determining the optimal allocation scheme for m spacecraft. Assume there are n target spacecraft, defined as the target set T = {t1, t2, ..., tn}. n The operation requires m spacecraft to conduct strikes, and each target spacecraft must be struck at least once by our spacecraft. Our spacecraft can strike at least one enemy target spacecraft. The constraints are expressed as follows: Where k represents the maximum number of enemy target spacecraft that a single friendly spacecraft can strike. Assuming all enemy targets are attacked, x ij =1 indicates that the i-th spacecraft strikes the j-th enemy target; otherwise, x ij =0. Therefore, when attacking the j-th target, its overall damage probability is: The expected operational benefit of attacking all n targets is: To avoid excessively high task imbalance when all weapons are engaged in strike missions, which would otherwise waste resources, the standard deviation of the number of enemy targets struck by each weapon is calculated to measure the balance of the final target allocation scheme. This is expressed as follows: in, The objective function to be optimized is: Here, f1 represents the expected operational benefits of the weapon striking the target; f2 represents the fuel consumption of the weapon during the entire process of striking the target; and f3 represents the overall mission load balance when assessing each weapon striking all targets. (2) Supervised fine-tuning The general-purpose LLM was fine-tuned using the corpus described above, enabling it to gain a deeper understanding of the Lambert problem and specialized concepts such as computational speed increment Δv and J2 perturbation, as well as the underlying physical principles. (3) Intelligent algorithm optimization Furthermore, an intelligent optimization algorithm is introduced to construct an optimization objective function based on indicators such as task balance, fuel consumption, and combat benefits. This optimizes the model so that its output scheme, while satisfying natural language instructions, is more inclined to select feasible target allocation schemes with optimal fuel consumption and highest combat benefits.
3. The spacecraft cluster target allocation method based on a large language model according to claim 2, characterized in that: In step d, the specific implementation method is as follows: (1) Based on the Lambert orbital problem model, the initial target allocation scheme is simulated and calculated to calculate the fuel consumption, transfer time, combat benefits, and mission balance of each spacecraft, and to verify whether it meets all the constraints. (2) If the initial scheme does not meet the constraints or the performance is not optimal, then numerical optimization is initiated. The initial scheme is used as the initial solution. Local search is performed under all constraints to optimize the predetermined objective function and obtain the final objective allocation scheme.
4. The spacecraft cluster target allocation method based on a large language model according to claim 3, characterized in that: Following step e, the process also includes receiving mission execution status data from the spacecraft cluster and using this data to continuously fine-tune the large language model.