Intelligent control system of hypersonic aircraft based on large language model

Abstract

This invention belongs to the field of hypersonic aircraft control technology and relates to a hypersonic aircraft intelligent control system based on a large language model. The purpose of this invention is to provide intelligent control of an aircraft using natural language instructions. The method includes a system design consisting of an intelligent decision-making layer, a weight mapping layer, an offline optimization database, an interpolation calculation layer, and a control execution layer, achieving stable tracking of reference instructions. This invention can achieve intelligent mapping from natural language instructions to precise control actions, significantly reducing operational complexity and improving the system's mission adaptability and intelligence level. It also has broad application prospects.

Description

Technical Field

[0001] The present invention belongs to the technical field of hypersonic aircraft control, and specifically relates to a hypersonic aircraft intelligent control system based on a Large Language Model (LLM). The system realizes intelligent control of the aircraft through natural language instructions. Background Art

[0002] Hypersonic vehicle control systems face challenges such as strong time-varying parameters within a wide flight envelope, uncertainty caused by complex flight environments, and the tightly coupled nature of flight and propulsion integration. Traditional control methods struggle to meet the demands of full-envelope, high-performance, and high-reliability control. Existing technologies typically require specialized technical personnel to perform complex parameter settings, mission planning, and real-time monitoring, resulting in high operational complexity and limiting the system's flexibility and adaptability in dynamic mission environments.

[0003] The patent "Large Civil Aircraft Flight Control Method Based on Direct Adaptive Control Reconstruction" (CN103235504A) proposes a model-following-based direct adaptive control method. By constructing an adaptive controller that includes controlled object state feedback and reference model state feedforward, this method achieves online adaptation to actuator failures. However, this method only pre-designs for specific failure modes and lacks the ability to understand complex task descriptions, making it incapable of intelligently adjusting control parameters through natural language commands.

[0004] While the patented "Hard-in-the-loop Simulation System for Hypersonic Aircraft Navigation and Control Systems" (CN113658340B) enables flight parameter acquisition and multi-subsystem coordination, it lacks intelligent mission understanding and parameter optimization capabilities. The system requires operators to manually set complex control parameters and is unable to automatically adjust control strategies based on mission requirements described in natural language.

[0005] The dynamic test system proposed in the patent "Method and Apparatus for Processing Aerodynamic Signals of a Variant Aircraft Wind Tunnel Dynamic Test Balance" (CN117909659A) can simulate aircraft attitude changes, but it fails to address the issues of multimodal data fusion and intelligent control strategy generation. The system also lacks the ability to understand the operator's intent, preventing intelligent human-machine interaction.

[0006] The patent "A simulation test platform and control method for hypersonic aircraft assessment" (CN104182272B) proposes a simulation test platform, but as a software assessment and verification platform for control algorithms, it only covers known dynamic models and aerodynamic parameter simulations, and lacks the ability to adapt to complex mission environments and intelligent decision-making functions.

[0007] The main deficiencies of existing technologies are as follows: lack of natural language understanding capabilities, inability to directly process the operator's natural language instructions, resulting in inefficient human-computer interaction, and the operator needs to control the system through a complex professional interface. Adjusting control parameters requires a lot of professional knowledge and experience, and the operation is extremely complex, which not only increases the cost of personnel training, but also limits the popularization and application of the system. At the same time, the existing system has poor task adaptability and it is difficult to dynamically optimize the control strategy according to different task requirements. When faced with complex and changing flight environments, manual redesign and adjustment of the control algorithm are often required. In addition, traditional control systems lack multimodal data fusion and intelligent decision-making capabilities. They cannot fully utilize the multi-source and differently structured data generated during the flight process, nor can they realize the generation of intelligent control strategies based on task preferences, which seriously affects the intelligence level and control performance of the system. Summary of the Invention

[0008] The present invention provides a hypersonic aircraft intelligent control system based on a large language model. This system aims to achieve intelligent mapping from natural language instructions to precise control actions by combining the natural language understanding and reasoning capabilities of the large language model with high-precision control algorithms, significantly reducing operational complexity and improving the system's mission adaptability and intelligence level.

[0009] The technical solutions of the present invention are as follows:

[0010] A hypersonic aircraft intelligent control system based on a large language model includes an intelligent decision-making layer, a weight mapping layer, an offline optimization database, an interpolation calculation layer, and a control execution layer, as follows:

[0011] (1) Building an intelligent decision-making layer to achieve natural language understanding and security verification

[0012] The intelligent decision-making layer is the system's core intelligent unit. Built on a large language model, it is responsible for understanding natural language commands and generating control strategy weights and flight trajectories. This layer includes a command preprocessing module, a semantic safety guardrail module, a trajectory generation module, and a weight generation module.

[0013] Instruction preprocessing module:

[0014] The command preprocessing module, the first step in the intelligent decision-making layer, receives natural language input and standardizes it, including format standardization, unit conversion, and contextual information integration. The system supports various command input formats. A typical flight command format is: "Within T seconds, adjust the altitude to H kilometers, the speed to M Mach, and the required performance preference to P," where T is the time constraint, H is the target altitude, M is the target speed, and P is the performance preference description. This module outputs the standardized command information and transmits it to the semantic safety guardrail module.

[0015] Semantic Security Guardrail Module:

[0016] The semantic safety guardrail module receives the standardized command information output by the command preprocessing module and performs multi-level safety verification on the input command based on the RoboGuard framework. The feasibility of the command is determined by calculating the required average climb rate and acceleration. The average climb rate calculation formula is:

[0017] (1)

[0018] in, is the average climb rate (m / s), is the target height (m), is the current altitude (m), is the specified time (s). The average acceleration is calculated as:

[0019] (2)

[0020] in, is the average acceleration (Ma / s), is the target speed (Ma), is the current speed (Ma).

[0021] The system compares the calculated average climb rate and acceleration values with the aircraft's performance envelope to identify physically unattainable commands. The module also performs logical consistency checks to identify logical inconsistencies within commands. For example, if the operator simultaneously requests "fastest arrival" and "minimum fuel consumption," the system can detect this conflict and ask the operator to clarify the priority. Furthermore, the safety boundary protection function verifies whether commands exceed predefined safety envelope limits, including key safety parameters such as the altitude safety corridor (20 km, 35 km) and the speed safety range (4 Ma, 8 Ma), ensuring that all verified commands are within the aircraft's safe operating range.

[0022] Trajectory generation module:

[0023] The trajectory generation module receives the command information verified by the semantic safety guardrail module and generates a continuous altitude-speed trajectory sequence based on the target state and time constraints in the natural language command to obtain the desired trajectory. For flight missions, the quintic polynomial method is used:

[0024] Altitude trajectory :

[0025] (3)

[0026] Speed trajectory :

[0027] (4)

[0028] in, are the coefficients of the height trajectory polynomial, is the velocity trajectory polynomial coefficient, and t is the time variable (s).

[0029] The boundary conditions for trajectory generation include position, velocity, and acceleration constraints at the start and end times. Specifically, these are the initial and final altitudes, velocity values, and their first- and second-order derivatives. The polynomial coefficients are determined by solving a system of linear equations. The generated trajectory information is then passed to the control execution layer.

[0030] Weight generation module:

[0031] The weight generation module receives the instruction information verified by the semantic safety guardrail module and converts the natural language instruction into a five-dimensional task weight vector through semantic parsing of the large language model:

[0032] (5)

[0033] Where, is the overshoot weight, To adjust the time weight, is the steady-state error weight, is the weight of the smoothness of the rudder surface change, is the robust margin weight, all weights are non-negative. The weight vector satisfies the normalization constraint:

[0034] (6)

[0035] The large language model identifies the operator's task preferences through a pre-established knowledge base of keyword-to-weight mapping. For time-priority tasks, when the instructions contain keywords such as "as soon as possible", "quickly", and "urgent", the system will significantly increase the adjustment time weight. For tasks with a smooth priority, when keywords such as "smooth", "gentle", and "comfortable" are detected, the system will increase the overshoot weight. The value of , ensures that the overshoot in the control process is minimized and a smooth state transition is achieved. For precision-priority tasks, keywords such as "precision", "accurate", and "error-free" will trigger the steady-state error weight. For safety-priority tasks, keywords such as “safety”, “stability”, and “conservative” correspond to the robustness margin weights. For comfort-first tasks, keywords such as "comfort", "smoothness", and "load reduction" will increase the weight of the smoothness of the rudder changes. The value of ensures the smoothness of the control surface changes. The weight vector generated Passed to the weight map layer.

[0036] (2) Design a weight mapping layer to achieve accurate conversion of weight vectors to control parameters

[0037] Create a set of weight anchor points:

[0038] The weight mapping layer is responsible for converting the task weight vector output by the intelligent decision layer into Matching with the offline optimization database to achieve accurate mapping from weight space to control parameter space. The core mechanism of this layer is to maintain a predefined set of weight anchor points , these anchor points form a convex hull in the weight space through convex combination theory, which is mathematically expressed as:

[0039] (7)

[0040] in, is the convex combination coefficient of the i-th weight anchor point, and N is the total number of anchor points.

[0041] When large language models generate arbitrary task weight vectors When , the weight mapping layer first determines whether the vector is within the predefined convex hull. , the system determines a set of convex combination coefficients by solving a convex optimization problem , so that the weight vector can be expressed as a linear combination of anchor points:

[0042] (8)

[0043] Design weight projection mechanism:

[0044] For weight vectors that are not within the convex hull, the system uses the optimal projection strategy to project them onto the nearest convex hull boundary point. The projection calculation formula is:

[0045] (9)

[0046] in, is the weight vector after projection, represents the two-norm, for The minimum value in .

[0047] This design ensures that all weight vectors can find a reasonable mapping relationship within the predefined convex hull range, providing a mathematically rigorous guarantee for subsequent interpolation calculations, while avoiding the problem of weight vectors exceeding the coverage of the offline optimization database. Passed to the interpolation calculation layer.

[0048] (3) Build an offline optimization database to provide full coverage of optimal control parameters

[0049] The offline optimization database is the core knowledge storage unit of the system. It organizes data using the three-dimensional structure of "weight-operating condition-optimization extreme point" to provide a comprehensive optimization parameter basis for the intelligent control system.

[0050] Design a three-dimensional database architecture:

[0051] The weight dimension of the database corresponds to N predefined weight anchor points , each anchor point represents a specific control performance preference combination. The operating condition dimension only includes altitude and speed parameters, with the altitude range set to 20km to 35km and the speed range covering 4Ma to 8Ma. The sampling interval is designed based on the rationality of flight pressure to ensure the physical rationality of the operating condition points. The optimized pole dimension corresponds to the optimal pole configuration of each weight anchor point-operating condition combination. These pole data are obtained through offline CMA-ES (Covariance Matrix Adaptation Evolution Strategy) algorithm optimization, ensuring optimal control performance under specific weight preferences and flight conditions.

[0052] Establish database index structure:

[0053] The database index structure is represented in the form of a two-dimensional mapping:

[0054] (10)

[0055] in, represents the i-th weight anchor point, represents the jth operating point, represents the corresponding optimal pole configuration.

[0056] Implement offline multi-objective optimization:

[0057] The offline optimization process adopts a multi-objective optimization method, and its cost function is designed as:

[0058] (11)

[0059] in, Assign vectors to the poles, is the overshoot penalty, To adjust the time cost, is the steady-state error cost, is the cost of smoothness of rudder surface changes, is the robustness margin cost.

[0060] This cost function comprehensively considers five key performance indicators, achieves balanced optimization between different performance objectives by adjusting the weights, and provides high-quality reference data for subsequent interpolation calculations.

[0061] (4) Develop an interpolation calculation layer to achieve optimal pole configuration for dual-domain interpolation

[0062] The interpolation calculation layer implements a dual-domain interpolation algorithm, receiving the convex combination coefficients output by the weight mapping layer and the current flight state, and obtains the optimal pole configuration through interpolation calculation. This layer adopts a hierarchical and progressive interpolation strategy, first performing interpolation matching in the weight domain, then performing refined interpolation in the operating condition domain, ultimately achieving an organic combination of the two domains.

[0063] Implement the weighted domain interpolation algorithm:

[0064] In the weight domain interpolation stage, the system uses the convex combination coefficients passed by the weight mapping layer to interpolate the predefined weight anchor points. Specifically, for any task weight vector output by the large language model , the system determines the optimal convex combination coefficients by solving the least squares optimization problem:

[0065] (12)

[0066] This process needs to satisfy the normalization constraints and non-negativity constraints , ensuring the physical rationality and mathematical rigor of the interpolation results. The core idea of weight domain interpolation is to discretize the continuous weight space into a finite number of anchor points, and use convex combination theory to ensure that any weight vector can be represented by a linear combination of the existing anchor points.

[0067] Design case domain interpolation mechanism:

[0068] In the interpolation phase of the working condition domain, the system And each weight anchor , select K nearest neighbor operating points from the corresponding pole database, and use the radial basis function interpolation method to obtain the local optimal pole configuration. The interpolation calculation formula is:

[0069] (13)

[0070] in, is the local optimal point obtained by interpolation of the working condition domain; is the distance-weighted interpolation coefficient, is the bandwidth parameter of the radial basis function, which is used to control the smoothness and locality of the interpolation. This method can fully utilize the similarity between operating points and achieve smooth extreme transition through distance weighting, thus avoiding control performance jumps caused by sudden changes in operating conditions.

[0071] Perform final pole synthesis:

[0072] The final pole synthesis stage organically integrates the interpolation results of the weight domain and the working condition domain. The system calculates the final optimal pole configuration through the formula:

[0073] (14)

[0074] in, is the convex combination coefficient obtained by interpolation in the weight domain.

[0075] This dual-domain interpolation strategy not only ensures the accurate reflection of weight preferences, but also takes into account the actual constraints of the current flight state, achieving the best balance between personalized control requirements and physical feasibility. Passed to the control execution layer.

[0076] (5) Establish a control execution layer to achieve precise flight control

[0077] The control execution layer includes the outer loop control module and the inner loop control module, which is responsible for converting the output of the intelligent decision-making layer into specific control actions.

[0078] Design the outer loop control module:

[0079] The outer loop control module receives the desired trajectory output by the intelligent decision-making layer trajectory generation module and uses the dynamic inversion method to establish a mapping relationship from the desired trajectory to the inner loop command:

[0080] (15)

[0081] in, is the angle of attack command (°), is the equivalent ratio instruction, is the altitude change rate instruction (m / s), is the speed change rate instruction (Ma / s), and is the aircraft dynamics function.

[0082] Construct the inner loop control module:

[0083] The inner loop control module receives the angle of attack and equivalence ratio instructions output by the outer loop control module, and uses the optimal pole configuration provided by the interpolation calculation layer based on the adaptive pole configuration method. , achieving precise attitude and propulsion system control.

[0084] Pole scheduling uses a progressive adjustment method:

[0085] (16)

[0086] in, is the adjustment coefficient, which controls the pole update speed. To control the number of steps.

[0087] Beneficial effects of the present invention:

[0088] 1. Leveraging the natural language understanding capabilities of a large language model, we achieve intelligent mapping from natural language commands to precise control parameters, significantly reducing operational complexity and enabling non-professionals to efficiently interact with complex control systems using natural language. By designing a dual safety mechanism of semantic safety guardrails and control safety guardrails, we provide comprehensive security protection at both the command and control execution levels, ensuring safe and reliable system operation under various circumstances.

[0089] 2. The offline optimization database structure of "weight-operating condition-optimized pole" is adopted, combined with a dual-domain interpolation algorithm, to achieve real-time calculation of the optimal pole configuration under any weight vector. The system can dynamically adjust the control strategy according to different task preferences (time priority, stability priority, precision priority, etc.), and compared with traditional fixed parameter control methods, the task adaptability is improved.

[0090] 3. The offline optimization database covers the wide flight envelope of hypersonic aircraft (altitude 20-35km, speed 4-8Ma). The physical rationality of the interpolation results is guaranteed by convex combination theory, realizing intelligent control within a wide range of flight envelope. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 This is the overall architecture diagram of the hypersonic aircraft intelligent control system based on a large language model;

[0092] Figure 2 It is the CMA-ES optimization flow chart;

[0093] Figure 3 It is the three-dimensional graph of the real part of the conjugate poles of the offline optimization database;

[0094] Figure 4 It is the three-dimensional graph of the imaginary part of the conjugate poles of the offline optimization database;

[0095] Figure 5 It is the three-dimensional graph of the first real extreme point of the offline optimization database;

[0096] Figure 6 It is the three-dimensional graph of the second real extreme point of the offline optimization database;

[0097] Figure 7This is the control execution layer architecture diagram;

[0098] Figure 8 It is the system height response curve;

[0099] Figure 9 It is the system speed response curve;

[0100] Figure 10 is the curve diagram of angle of attack change;

[0101] Figure 11 It is the rudder deviation change curve;

[0102] Figure 12 It is the fuel equivalence ratio change curve. DETAILED DESCRIPTION

[0103] The following further describes the specific implementation of the present invention in conjunction with the accompanying drawings and technical solutions. Figure 1 As shown, the large language model-based hypersonic aircraft intelligent control system of the present invention realizes a complete mapping from natural language instructions to precise control actions through the collaborative work of five main levels.

[0104] Example: Layered implementation of slowly adjusting tasks

[0105] The operator inputs natural language instructions: "Slowly adjust the aircraft from the current state (altitude 25km, speed 5Ma) to the target state (altitude 26km, speed 6Ma) within one minute."

[0106] The system is executed in the following five levels:

[0107] The first layer: the implementation process of the intelligent decision-making layer

[0108] Command pre-processing module execution: The system first parses the input command and extracts key information: time constraint T=60 seconds, starting altitude =25km, target altitude =26km, starting speed =5Ma, target speed = 6Ma, performance preference keyword = "slow". The instruction preprocessing module normalizes "one minute" to 60 seconds and identifies "slow" as a control preference that prioritizes smoothness.

[0109] Verification of semantic safety guardrail module: The system calculates the average climbing rate according to formula (1): , which is within the performance envelope of hypersonic aircraft [5m / s, 50m / s]. The average acceleration is calculated according to formula (2): Ma / s, which is also within the safety range [0.01Ma / s, 0.1Ma / s]. The system verified that the altitude target value of 26km and the speed target value of 6Ma were both within the safety envelope, and the command passed the safety verification.

[0110] Implementation of trajectory generation module: Generate the height trajectory according to formula (3), and the boundary conditions are set as: (0)=25km, (60)=26km, (0)=0, (60)=0, (0)=0, (60)=0. The polynomial coefficients are obtained by solving the six-variable linear equation system: =25, =0, =0, =0.0046, =-0.00015, =0.0000017. Similarly, the velocity trajectory is generated according to formula (4) and the corresponding polynomial coefficient combination is obtained to ensure that the trajectory is smooth and continuous throughout the entire 60 seconds.

[0111] like Figure 2 The CMA-ES optimization process shown ensures the optimal pole configuration for each weight anchor point and load case combination in the offline database. Figure 3 、 Figure 4 、 Figure 5 、 Figure 6 The offline optimization database is presented, in which the real and imaginary parts of the conjugate poles and the two real poles show obvious distribution patterns with the changes of Mach number and altitude, verifying the integrity and consistency of the offline optimization database.

[0112] The weight generation module performs: The large language model recognizes the keyword "slow" and determines that this is a task of the stability priority type based on the preset mapping knowledge base. The system generates a weight vector =[0.35, 0.25, 0.15, 0.20, 0.05], where the overshoot weight =0.35 (highest weight, ensuring a smooth process), adjust the time weight =0.25 (moderate weight, meeting time constraints), smoothness weight =0.20 (higher weight, ensuring smooth changes in the rudder), steady-state error weight =0.15, robustness weight = 0.05. The weight vector satisfies the normalization constraint and sums to 1.0.

[0113] Layer 2: Implementation of the Weight Mapping Layer

[0114] Weight anchor matching: The weight mapping layer receives the weight vector output by the intelligent decision layer = [0.35, 0.25, 0.15, 0.20, 0.05] and matches it with the predefined weight anchor set. The system database stores multiple representative weight anchors, of which the four most relevant to the current weight vector include: High Stability Anchor =[0.40, 0.20, 0.10, 0.25, 0.05], balanced anchor point =[0.30, 0.30, 0.20, 0.15, 0.05], time-priority anchor =[0.25, 0.35, 0.15, 0.20, 0.05] and very high stability anchor points =[0.45, 0.15, 0.10, 0.25, 0.05].

[0115] Convex combination coefficient calculation: According to formula (8), the system solves the convex combination coefficient by the least square method. After optimization calculation, we get: =0.3, =0.2, =0.1, =0.4, satisfying the constraint and Verification result: 0.3× +0.2× +0.1× +0.4× ≈[0.35, 0.25, 0.15, 0.20, 0.05]= , the mapping accuracy meets the requirements.

[0116] The third layer: the implementation process of the interpolation calculation layer

[0117] Interpolation execution in the working condition domain: The interpolation calculation layer is based on the current working condition state =[25km, 5Ma], select K=6 nearest neighbor working condition points for each weight anchor point from the offline database for interpolation calculation. For example, the nearest operating points selected by the system include:

[0118] =[24.5km, 4.8Ma] (distance = 0.71)

[0119] =[25.2km, 5.1Ma] (distance = 0.22)

[0120] =[24.8km, 5.2Ma] (distance = 0.28)

[0121] =[25.5km, 4.9Ma] (distance = 0.52)

[0122] =[24.6km, 5.3Ma] (distance = 0.57)

[0123] =[25.1km, 4.7Ma] (distance = 0.32)

[0124] Radial basis function interpolation calculation: During the radial basis function interpolation calculation, the system sets the bandwidth parameter according to formula (13) =0.5, calculate the interpolation weight of each operating point:

[0125] = exp(-0.71² / (2×0.5²))=0.12

[0126] =exp(-0.22² / (2×0.5²))=0.85

[0127] =exp(-0.28² / (2×0.5²))=0.73

[0128] = exp(-0.52² / (2×0.5²))=0.32

[0129] = exp(-0.57² / (2×0.5²))=0.27

[0130] = exp(-0.32² / (2×0.5²))=0.62

[0131] After normalization, the system performs weighted average calculation to obtain the local optimal point The system has three other weight anchor points. 、 、 Perform the same interpolation process and calculate the corresponding local optimal extremes 、 、 .

[0132] Final pole synthesis: In the final pole synthesis stage, the system performs the fusion calculation of the weight domain and the working condition domain according to formula (14): =0.3× +0.2× +0.1× +0.4× After dual-domain interpolation calculation, the system obtains the final optimal pole configuration =[-2.1-0.8j, -2.1+0.8j, -4.5, -5.1]. This pole configuration fully takes into account the stability requirements of the "slow" mission and the characteristics of the current flight conditions, providing an optimized control parameter basis for the control execution layer.

[0133] The fourth layer: control the implementation process of the execution layer

[0134] like Figure 7 In the control execution layer architecture shown in the figure, the outer-loop controller receives the altitude and speed trajectories generated by the intelligent decision layer, and the inner-loop controller implements precise control based on the optimal pole configuration provided by the interpolation calculation layer.

[0135] Outer loop controller execution: The outer loop controller of the control execution layer receives the altitude and speed trajectory generated by the intelligent decision layer, and uses the dynamic inversion method to convert the trajectory command into the angle of attack and fuel equivalence ratio command. At the initial time of the mission t=0, the commands received by the system include: altitude command (0)=25km, speed command (0)=5Ma, altitude change rate command (0)=0m / s, speed change rate instruction (0)=0Ma / s. According to the dynamic inverse calculation of formula (15), the system obtains the initial control instruction: angle of attack instruction (0)=3.113°, fuel equivalence ratio command (0)=0.286. At the midpoint of the mission execution, t=30 seconds, the system's command status is updated to: altitude command (30) = 25.5 km, speed command (30)=5.5Ma, the corresponding control command calculation result is: angle of attack command (30)=2.85°, fuel equivalence ratio command (30)=0.340.

[0136] Inner loop controller execution: The inner loop controller receives the attack angle and equivalence ratio instructions output by the outer loop controller, and uses the progressive pole scheduling method according to formula (16) to achieve smooth control parameter transition. System setting adjustment coefficient =0.08 to ensure the stability of the pole update, the initial pole configuration is set to p(0)=[-2.1-0.8j,-2.1+0.8j,-4.5,-5.1]. In the first pole update, the system calculates P(1)=P(0)+0.08×( -P(0))=[-2.108-0.808j, -2.108+0.808j, -4.524, -5.132]. Through continuous progressive adjustment, the pole configuration converges smoothly to the target value after about 15 control cycles. , achieving disturbance-free transition of control parameters and ensuring the smoothness of the flight process.

[0137] Layer 5: Control system execution response

[0138] Highly responsive performance analysis: e.g. Figure 8 As shown, the system demonstrated excellent altitude response performance, successfully adjusting the altitude from 25 km to 26 km within the specified time of 60 seconds. The entire response process exhibited an ideal S-shaped curve. The response process can be divided into three distinct phases: the first 15 seconds is a slow start phase, with the altitude change rate gradually increasing from zero; the middle 30 seconds is the main climb phase, with the altitude change rate reaching a maximum of approximately 30 m / s, providing the main altitude increment; and the final 15 seconds is a stable convergence phase, with the altitude change rate gradually decreasing from the maximum to zero, ensuring smooth arrival at the target altitude. The system's final control accuracy indicators showed a steady-state error of less than 5 m, a total adjustment time of 59.2 seconds, and zero overshoot, fully meeting the "slow" adjustment performance requirements specified in natural language commands.

[0139] Speed response performance analysis: Figure 9 As shown, the speed response performance is also excellent, with the Mach number increasing smoothly from 5 to 6 Mach. The response curve remains highly synchronized with the altitude change, demonstrating the superiority of multivariable coordinated control. The maximum tracking error during speed tracking is kept within 0.02 Mach, and the entire change process is smooth and continuous, without any sudden changes or oscillations. The time constant of the speed response perfectly matches the altitude response, verifying the superior performance of the outer loop controller in multivariable coupled control and ensuring speed-altitude coordination during the aircraft's climb.

[0140] Control surface response analysis: Control surface response analysis shows that the changes of each control surface meet the requirements of smooth control. Figure 10 As shown in the figure, the angle of attack decreases steadily from the initial 3.113° to the final 2.452° during the entire flight process. The change range is moderate and the change rate is always controlled within 0.3° / s, ensuring a smooth transition of the aircraft's attitude. Figure 11As shown in the figure, the elevator deflection angle changes from 9.419° to 9.941°, and the change process is continuous and smooth, without saturation and oscillation, which reflects the good control quality of the rudder surface. Figure 12 As shown in the figure, the fuel equivalence ratio gradually increases from an initial 0.286 to a final 0.392. Its variation pattern perfectly matches the altitude change of the flight trajectory, providing necessary thrust support during the critical stage of the climb. At the same time, the fuel consumption growth rate is controlled within an economically reasonable range, fully demonstrating the system's intelligent understanding and precise execution of the "slow" instruction.

[0141] Through this layered implementation, the proposed large-scale language model-based intelligent control system for hypersonic aircraft can effectively understand performance preferences such as "slowness" in natural language commands. Through semantic parsing at the intelligent decision-making layer, precise conversion at the weight mapping layer, optimized configuration at the interpolation calculation layer, and precise control at the control execution layer, it achieves a complete closed loop from natural language to flight control. The coordinated operation of all layers ensures intelligent, safe, and high-performance control, providing a comprehensive technical solution for the intelligent development of hypersonic aircraft control systems.

Claims

1. A hypersonic aircraft intelligent control system based on a large language model, characterized in that: It includes intelligent decision-making layer, weight mapping layer, offline optimization database, interpolation calculation layer and control execution layer, as follows: (1) Building an intelligent decision-making layer to achieve natural language understanding and security verification The intelligent decision-making layer is built on a large language model and is responsible for understanding natural language instructions and generating control strategy weights and flight trajectories. This layer includes an instruction preprocessing module, a semantic safety guardrail module, a trajectory generation module, and a weight generation module. Instruction preprocessing module: The instruction preprocessing module receives the input natural language instructions and performs standardization processing on them, including format standardization, unit conversion, and context information integration. This module outputs the standardized instruction information and passes it to the semantic security guardrail module. Semantic Security Guardrail Module: The semantic safety guardrail module receives the standardized command information output by the command preprocessing module and performs multi-level safety verification on the input command based on the RoboGuard framework. It determines the feasibility of the command by calculating the required average climb rate and acceleration. The average climb rate calculation formula is: (1) , where is the average rate of climb, is the target height, is the current altitude, is the specified time; the average acceleration calculation formula is: (2) , where is the average acceleration, is the target speed, is the current speed; The system compares calculated average climb rate and acceleration values with the aircraft's performance envelope to identify physically unattainable commands. The module also performs logic consistency checks to identify logical inconsistencies within commands. A safety margin protection function verifies that commands do not exceed predefined safety envelope limits, ensuring that all verified commands are within the aircraft's safe operating range. Trajectory generation module: The trajectory generation module receives the command information verified by the semantic safety guardrail module and generates a continuous altitude-speed trajectory sequence based on the target state and time constraints in the natural language command to obtain the desired trajectory. For flight missions, the quintic polynomial method is used: Altitude trajectory : (3) , velocity trajectory : (4) , where are the coefficients of the height trajectory polynomial, is the velocity trajectory polynomial coefficient, t is the time variable; Weight generation module: The weight generation module receives the instruction information verified by the semantic safety guardrail module and converts the natural language instruction into a five-dimensional task weight vector through semantic parsing of the large language model: (5) , where is the overshoot weight, To adjust the time weight, is the steady-state error weight, is the weight of the smoothness of the rudder surface change, is the robust margin weight, all weights are non-negative; the weight vector satisfies the normalization constraint: (6) The large language model identifies the operator's task preference through a pre-established keyword-to-weight mapping knowledge base; the generated weight vector Passed to the weight mapping layer; (2) Design a weight mapping layer to achieve accurate conversion of weight vectors to control parameters Create a set of weight anchor points: The weight mapping layer is responsible for converting the task weight vector output by the intelligent decision layer into Matching with the offline optimization database to achieve accurate mapping from weight space to control parameter space; by maintaining a predefined set of weight anchor points , these anchor points form a convex hull in the weight space through convex combination theory, which is mathematically expressed as: (7) , where is the convex combination coefficient of the i-th weight anchor point, and N is the total number of anchor points; When large language models generate arbitrary task weight vectors When , the weight mapping layer first determines whether the vector is within the predefined convex hull; if , the system determines a set of convex combination coefficients by solving a convex optimization problem , so that the weight vector can be expressed as a linear combination of anchor points: (8) Design a weight projection mechanism: For weight vectors that are not in the convex hull, the system adopts the optimal projection strategy to project them onto the nearest convex hull boundary point. The projection calculation formula is: (9) , where is the weight vector after projection, represents the two-norm, for The minimum value in ; (3) Build an offline optimization database to provide full coverage of optimal control parameters Design a three-dimensional database architecture: The weight dimension of the database corresponds to N predefined weight anchor points , each anchor point represents a specific control performance preference combination; the operating condition dimension only includes altitude and speed parameters; the optimization pole dimension corresponds to the optimal pole configuration of each weighted anchor point-operating condition combination; Establish database index structure: The database index structure is represented in the form of a two-dimensional mapping: (10) , where represents the i-th weight anchor point, represents the jth operating point, represents the corresponding optimal pole configuration; Implement offline multi-objective optimization: The offline optimization process adopts a multi-objective optimization method, and its cost function is designed as: (11), among which, Assign vectors to the poles, is the overshoot penalty, To adjust the time cost, is the steady-state error cost, is the cost of smoothness of rudder surface changes, is the robust margin cost; (4) Develop an interpolation calculation layer to achieve optimal pole configuration for dual-domain interpolation The interpolation calculation layer implements a dual-domain interpolation algorithm. It receives the convex combination coefficients output by the weight mapping layer and the current flight state, and obtains the optimal pole configuration through interpolation calculation. This layer adopts a hierarchical and progressive interpolation strategy, first performing interpolation matching in the weight domain, then performing refined interpolation in the operating condition domain, ultimately achieving an organic combination of the two domains. Implement the weighted domain interpolation algorithm: In the weight domain interpolation stage, the system uses the convex combination coefficients passed by the weight mapping layer to interpolate the predefined weight anchor points; specifically, for any task weight vector output by the large language model , the system determines the optimal convex combination coefficients by solving the least squares optimization problem: (12) This process needs to satisfy the normalization constraint and non-negativity constraints ; Design case domain interpolation mechanism: In the interpolation phase of the working condition domain, the system And each weight anchor , select K nearest neighbor operating points from the corresponding pole database, and use the radial basis function interpolation method to obtain the local optimal pole configuration; the interpolation calculation formula is: (13) Among them, is the local optimal point obtained by interpolation of the working condition domain; is the distance-weighted interpolation coefficient, is the bandwidth parameter of the radial basis function, which is used to control the smoothness and locality of the interpolation. is the jth nearest neighbor operating point; Perform final pole synthesis: The final pole synthesis stage organically integrates the interpolation results of the weight domain and the working condition domain; the system calculates the final optimal pole configuration through the formula: (14) Among them, is the convex combination coefficient obtained by weight domain interpolation; the final pole configuration Passed to the control execution layer; (5) Establish a control execution layer to achieve precise flight control The control execution layer includes the outer loop control module and the inner loop control module, which is responsible for converting the output of the intelligent decision-making layer into specific control actions; Design the outer loop control module: The outer loop control module receives the desired trajectory output by the intelligent decision-making layer trajectory generation module and uses the dynamic inversion method to establish a mapping relationship from the desired trajectory to the inner loop command: (15) Among them, is the angle of attack command, is the equivalent ratio instruction, is the altitude change rate instruction, is the speed change rate instruction, and is the aircraft dynamics function; Construct the inner loop control module: The inner loop control module receives the angle of attack and equivalence ratio instructions output by the outer loop control module, and uses the optimal pole configuration provided by the interpolation calculation layer based on the adaptive pole configuration method. , achieving precise attitude and propulsion system control; Pole scheduling uses a progressive adjustment method: (16) Among them, is the adjustment coefficient, which controls the pole update speed. To control the number of steps.

2. The hypersonic aircraft intelligent control system based on a large language model according to claim 1, characterized in that: The command input of the command preprocessing module has the following flight command format: "Within T seconds, adjust the altitude to H kilometers and the speed to M Mach, and require a performance preference of P", where T is the time constraint, H is the target altitude, M is the target speed, and P is the performance preference description.

3. The hypersonic aircraft intelligent control system based on a large language model according to claim 1, characterized in that: The process of the large language model identifying the operator's task preference is as follows: For time-priority tasks, when the instructions contain the keywords "as soon as possible", "quickly", and "urgent", the system increases the adjustment time weight. For tasks with a smooth priority, when the keywords "smooth", "gentle", and "comfortable" are detected, the system increases the overshoot weight. For precision-priority tasks, the system increases the weight of steady-state error for the keywords "precision", "accuracy", and "error-free". Improvement; for safety priority tasks, "safety", "stability", and "conservative", the system increases the robustness margin weight For comfort-priority tasks, the system increases the weight of the smoothness of the rudder changes for the keywords "comfort", "smoothness", and "load reduction". The numerical value of .

4. The hypersonic aircraft intelligent control system based on a large language model according to claim 1, characterized in that: In the working condition dimension of the three-dimensional database architecture, the altitude range is set to 20km to 35km, and the speed range covers 4Ma to 8Ma; the pole data is obtained through offline CMA-ES algorithm optimization.

Images