Hypersonic aircraft intelligent control system based on large language model

Abstract

The invention belongs to the technical field of hypersonic flight vehicle control, and relates to a hypersonic flight vehicle intelligent control system based on a large language model. The purpose of the invention is to provide intelligent control of an aircraft through natural language instructions. The method comprises system design of an intelligent decision-making layer, a weight mapping layer, an offline optimization database, an interpolation calculation layer and a control execution layer, and stable tracking of a reference instruction is realized. According to the method, intelligent mapping from a natural language instruction to an accurate control action can be realized, the operation complexity is remarkably reduced, and the task adaptability and the intelligent level of the system are improved. And the method has a wide application prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of hypersonic vehicle control, and specifically relates to an intelligent control system for hypersonic vehicles based on a large language model (LLM). This system realizes the intelligent control of the vehicle through natural language instructions. Background Art

[0002] The control system of hypersonic vehicles faces challenges such as strong time-variation of parameters under a wide flight envelope, uncertainties caused by complex flight environments, and strong coupling characteristics under flight propulsion integration. Traditional control methods are difficult to meet the control requirements of the entire flight envelope, high performance, and high reliability. Existing technologies usually require professional technicians to perform complex parameter settings, task planning, and real-time monitoring, with high operation complexity, which limits the flexibility and adaptability of the system in dynamic task environments.

[0003] The patent "Flight control method for large civil aircraft based on direct adaptive control reconstruction" (CN103235504A) proposes a direct adaptive control method based on model following. By constructing an adaptive controller including terms such as state feedback of the controlled object and state feedforward of the reference model, online adaptation to actuator failures is achieved. However, this method is only pre-designed for specific fault modes and lacks the ability to understand complex task descriptions, and cannot adjust control parameters intelligently through natural language instructions.

[0004] The patent "A semi-physical simulation system for hypersonic vehicle navigation and control system" (CN113658340B) can realize flight parameter acquisition and multi-subsystem cooperation, but lacks intelligent task understanding and parameter optimization capabilities. The system requires operators to manually set complex control parameters and cannot automatically adjust control strategies according to task requirements described in natural language.

[0005] The dynamic test system proposed in the patent "Method and device for processing aerodynamic force signals of a balance in a wind tunnel for a variable aircraft" (CN117909659A) can simulate the attitude change of the aircraft, but does not solve the problems of multi-modal data fusion and intelligent control strategy generation. The system lacks the ability to understand the intentions of operators and cannot realize intelligent human-computer interaction.

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

[0007] The main deficiencies of the prior art are as follows: It lacks natural language understanding ability and cannot directly process the natural language instructions of operators, resulting in low human-computer interaction efficiency. Operators need to control the system through complex professional interfaces. Adjusting control parameters requires a lot of professional knowledge and experience, and the operation complexity is extremely high. This not only increases the personnel training cost but also limits the popularization and application of the system. At the same time, the existing system has poor task adaptability and is difficult to dynamically optimize control strategies according to different task requirements. When facing complex and changeable flight environments, it often requires manual re-design and adjustment of control algorithms. In addition, traditional control systems lack multi-modal data fusion and intelligent decision-making capabilities, cannot make full use of multi-source and different-structured data generated during flight, and cannot generate intelligent control strategies based on task preferences, seriously affecting the intelligence level and control performance of the system. Summary of the Invention

[0008] The present invention provides an intelligent control system for hypersonic vehicles based on large language models, aiming to combine the natural language understanding and reasoning abilities of large language models with high-precision control algorithms to achieve an intelligent mapping from natural language instructions to precise control actions, significantly reducing operation complexity and improving the task adaptability and intelligence level of the system.

[0009] The technical solution of the present invention is as follows:

[0010] An intelligent control system for hypersonic vehicles based on large language models includes an intelligent decision-making layer, a weight mapping layer, an offline optimization database, an interpolation calculation layer, and a control execution layer, specifically as follows:

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

[0012] The intelligent decision-making layer is the core intelligent unit of the system, constructed based on large language models, 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.

[0013] Instruction preprocessing module:

[0014] As the first link in the intelligent decision-making layer, 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. The system supports various forms of instruction input. A typical flight instruction format is: "Within T seconds, adjust the altitude to H kilometers, the speed to M Mach, and the required performance preference is P", where T is the time constraint, H is the target altitude, M is the target speed, and P is the performance preference description. The module outputs the standardized instruction information and transmits it to the semantic safety guardrail module.

[0015] Semantic Safety Fence Module:

[0016] The Semantic Safety Fence Module receives the standardized instruction information output by the Instruction Preprocessing Module and performs multi-level security verification on the input instructions based on the RoboGuard framework. It determines the feasibility of the instructions by calculating the required average climb rate and acceleration. The formula for calculating the average climb rate is:

[0017] (1)

[0018] Where is the average climb rate (m / s), is the target altitude (m), is the current altitude (m), is the specified time (s). The formula for calculating the average acceleration is:

[0019] (2)

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

[0021] The system compares the calculated values of the average climb rate and acceleration with the aircraft performance envelope to identify physically unreachable instructions. At the same time, this module also performs a logical consistency check to identify logical contradictions within the instructions. For example, when the operator requests both "fastest arrival" and "minimum fuel consumption" simultaneously, the system can detect this conflict and request the operator to clarify the priority. In addition, the safety boundary protection function verifies whether the instructions exceed the predefined safety envelope limits, including key safety parameters such as the altitude safety corridor [20 km, 35 km] and the velocity safety range [4 Ma, 8 Ma], ensuring that all verified instructions are within the safe operating range of the aircraft.

[0022] Trajectory Generation Module:

[0023] The Trajectory Generation Module receives the instruction information verified by the Semantic Safety Fence Module and generates a continuous altitude-velocity trajectory sequence based on the target state and time constraints in the natural language instructions to obtain the desired trajectory. For flight missions, the fifth-order polynomial method is used:

[0024] Altitude trajectory :

[0025] (3)

[0026] Velocity trajectory :

[0027] (4)

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

[0029] The boundary conditions for trajectory generation include position, velocity, and acceleration constraints at the starting and ending times, specifically the height and velocity values at the initial and terminal points and their first and second derivatives. The polynomial coefficients are determined by solving a system of linear equations. The generated trajectory information is transmitted 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] In the formula, is the overshoot weight, is the settling time weight, is the steady-state error weight, is the smoothness weight of the rudder surface change, is the robust margin weight, and 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 mappings. For time-priority tasks, when keywords such as "as soon as possible", "quickly", "urgently" are included in the instruction, the system will significantly increase the value of the settling time weight , and at the same time appropriately reduce other weight components to achieve the control goal of fast response. For smoothness-priority tasks, when keywords such as "smooth", "gentle", "comfortable" are detected, the system will increase the value of the overshoot weight to ensure the minimum overshoot during the control process and achieve a smooth state transition. For accuracy-priority tasks, keywords such as "accurate", "precise", "error-free" will trigger an increase in the steady-state error weight to guarantee the final control accuracy. For safety-priority tasks, keywords such as "safe", "stable", "conservative" correspond to an increase in the robust margin weight to improve the robustness and safety margin of the system. For comfort-priority tasks, keywords such as "comfortable", "smooth", "load reduction" will increase the smoothness weight of the rudder surface change The numerical value ensures the smoothness of the control surface change. The generated weight vector is passed to the weight mapping layer.

[0036] (2) Design the weight mapping layer to achieve an accurate conversion from the weight vector to the control parameters

[0037] Establish a set of weight anchor points:

[0038] The weight mapping layer is responsible for matching the task weight vector output by the intelligent decision-making layer with the offline optimization database to achieve an accurate mapping from the weight space to the control parameter space. The core mechanism of this layer is to maintain a predefined set of weight anchor points , and these anchor points form a convex hull in the weight space through the convex combination theory, and its mathematical expression is:

[0039] (7)

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

[0041] When the large language model generates an arbitrary task weight vector , the weight mapping layer first determines whether this 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 the anchor points:

[0042] (8)

[0043] Design the weight projection mechanism:

[0044] For the weight vector outside the convex hull, the system adopts the optimal projection strategy to project it onto the nearest convex hull boundary point, and the projection calculation formula is:

[0045] (9)

[0046] where is the projected weight vector, represents the second norm, is the minimum value in.

[0047] This design ensures that all weight vectors can find a reasonable mapping relationship within the predefined convex hull range, provides a strict mathematical guarantee for the subsequent interpolation calculation, and at the same time avoids the problem that the weight vector exceeds the coverage range of the offline optimization database. The convex combination coefficient is passed to the interpolation calculation layer.

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

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

[0050] Design a three-dimensional database architecture:

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

[0052] Establish a database index structure:

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

[0054] (10)

[0055] where represents the i-th weight anchor point, represents the j-th operating condition point, represents the corresponding optimal pole configuration.

[0056] Implement offline multi-objective optimization:

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

[0058] (11)

[0059] where is the pole configuration vector, is the overshoot cost, is the settling time cost, is the steady-state error cost, is the smoothness cost of control surface change, is the robust margin cost.

[0060] This cost function comprehensively considers five key performance indicators, and realizes the balanced optimization between different performance goals through the adjustment of weights, providing high-quality reference data for subsequent interpolation calculations.

[0061] (4) Develop an interpolation calculation layer to achieve the optimal pole configuration of bi-domain interpolation

[0062] The interpolation calculation layer implements the bi-domain interpolation algorithm, 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, interpolation matching is performed in the weight domain, then refined interpolation is performed in the working condition domain, and finally the organic combination of the two domains is realized.

[0063] Implement the interpolation algorithm in the weight domain:

[0064] In the weight domain interpolation stage, the system performs interpolation calculations on the predefined weight anchor points using the convex combination coefficients transmitted by the weight mapping layer. Specifically, for any task weight vector output by the large language model , the system determines the optimal convex combination coefficients by solving a least squares optimization problem:

[0065] (12)

[0066] This process needs to satisfy the normalization constraint condition and the non-negativity constraint , 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 ensure that any weight vector can be represented by the linear combination of existing anchor points through the convex combination theory.

[0067] Design the interpolation mechanism in the working condition domain:

[0068] In the working condition domain interpolation stage, the system selects K nearest neighbor working condition points from the corresponding pole database for the current flight state and each weight anchor point , and uses the radial basis function interpolation method to obtain the local optimal pole configuration. The interpolation calculation formula is:

[0069] (13)

[0070] Among them, is the local optimal pole obtained by interpolation in the working condition domain; is the interpolation coefficient based on distance weighting, is the bandwidth parameter of the radial basis function, which is used to control the smoothness and locality of interpolation, is the j-th nearest neighbor operating point. This method can make full use of the similarity between operating points and achieve smooth pole transition through distance weighting, avoiding the control performance jump caused by sudden changes in operating conditions.

[0071] Execute the final pole synthesis:

[0072] In the final pole synthesis stage, the interpolation results of the weight domain and the operating condition domain are organically combined. The system calculates the final optimal pole configuration through the formula:

[0073] (14)

[0074] where 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 considers the actual constraints of the current flight state, achieving the best balance between personalized control requirements and physical feasibility. The final pole configuration is passed to the control execution layer.

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

[0077] The control execution layer includes an outer loop control module and an inner loop control module, which are 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 trajectory generation module of the intelligent decision-making layer and uses the dynamic inversion method to establish the mapping relationship from the desired trajectory to the inner loop command:

[0080] (15)

[0081] where is the angle of attack command (°), is the equivalence ratio command, is the altitude rate of change command (m / s), is the speed rate of change command (Ma / s), and are the aircraft dynamics functions.

[0082] Construct the inner loop control module:

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

[0084] The pole scheduling adopts a progressive adjustment method:

[0085] (16)

[0086] Among them, is the adjustment coefficient, which controls the pole update speed, is the control step number.

[0087] Advantages of the present invention:

[0088] 1. Through the natural language understanding ability of the large language model, the intelligent mapping from natural language instructions to precise control parameters is realized, significantly reducing the operation complexity, enabling non-professionals to efficiently interact with complex control systems through natural language. Through the dual safety mechanisms of designing semantic safety fences and control safety fences, comprehensive safety protection is provided from the instruction level and the control execution level, ensuring the safe and reliable operation of the system in various situations.

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

[0090] 3. The offline optimization database covers the wide flight envelope of hypersonic vehicles (altitude 20 - 35 km, speed 4 - 8 Ma). By the convex combination theory, the physical rationality of the interpolation results is guaranteed, realizing intelligent control within a large range of flight envelopes. Description of the Drawings

[0091] Figure 1 is the overall architecture diagram of the intelligent control system of hypersonic vehicles based on large language models;

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

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

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

[0095] Figure 5 is the three-dimensional diagram of the first real pole of the offline optimization database;

[0096] Figure 6 is the three-dimensional diagram of the second real pole of the offline optimization database;

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

[0098] Figure 8 It is the system's high-response curve graph;

[0099] Figure 9 It is the system's speed-response curve graph;

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

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

[0102] Figure 12 It is the fuel equivalence ratio change curve graph. Specific implementation mode

[0103] The following further illustrates the specific implementation mode of the present invention in combination with the attached drawings and technical solutions. As Figure 1 shown, the intelligent control system of the hypersonic vehicle based on the large language model of the present invention realizes the complete mapping from natural language instructions to precise control actions through the collaborative work of five main levels.

[0104] Example: Hierarchical implementation process of the slow adjustment task

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

[0106] The system executes in the following five levels in sequence:

[0107] First level: Implementation process of the intelligent decision-making layer

[0108] The instruction preprocessing module executes: The system first parses the input instruction and extracts key information: time constraint T = 60 seconds, starting altitude = 25 km, target altitude = 26 km, starting speed = 5 Ma, target speed = 6 Ma, performance preference keyword = "slow". The instruction preprocessing module standardizes "one minute" to 60 seconds and recognizes "slow" as a control preference with priority for smoothness.

[0109] The semantic safety fence module verifies: The system calculates the average climb rate according to formula (1): , and this value is within the performance envelope range of the hypersonic vehicle [5 m / s, 50 m / s]. Calculate the average acceleration according to formula (2): Ma / s, and this value is also within the safe range [0.01 Ma / s, 0.1 Ma / s]. The system verifies that the altitude target value of 26 km and the speed target value of 6 Ma are both within the safety envelope, and the command passes the safety verification.

[0110] Trajectory generation module implementation: Generate the altitude trajectory according to formula (3), and set the boundary conditions as: h(0) = 25 km, h(60) = 26 km, h'(0) = 0, h'(60) = 0, h''(0) = 0, h''(60) = 0. By solving the six - element linear equations, the polynomial coefficients are obtained: a0 = 25, a1 = 0, a2 = 0, a3 = 0.0046, a4 = - 0.00015, a5 = 0.0000017. Similarly, generate the speed trajectory according to formula (4) to obtain the corresponding combination of polynomial coefficients, ensuring that the trajectory is smooth and continuous throughout the 60 seconds.

[0111] As Figure 2 shown, the CMA - ES optimization process ensures the optimal pole configuration for each weight anchor point and operating condition combination in the offline database. Figure 3 、 Figure 4 、 Figure 5 、 Figure 6 Show the offline optimization database, where the real and imaginary parts of the conjugate poles and the two real poles show obvious distribution laws with the change of Mach number and altitude, verifying the integrity and consistency of the offline optimization database.

[0112] Weight generation module execution: The large - language model recognizes the keyword "slow", and according to the preset mapping knowledge base, determines that this is a task of the steady - state priority type. The system generates the weight vector w = [0.35, 0.25, 0.15, 0.20, 0.05], where the overshoot weight w1 = 0.35 (the highest weight to ensure a stable process), the settling - time weight w2 = 0.25 (a moderate weight to meet the time constraint), the smoothness weight w3 = 0.20 (a relatively high weight to ensure smooth change of the rudder surface), the steady - state error weight w4 = 0.15, and the robustness weight w5 = 0.05. The weight vector satisfies the normalization constraint, and the sum is 1.0.

[0113] The second layer: The implementation process of the weight mapping layer

[0114] Weight anchor matching: The weight mapping layer receives the weight vector output by the intelligent decision-making layer =[0.35, 0.25, 0.15, 0.20, 0.05] and matches it with a predefined set of weight anchors. Multiple representative weight anchors are stored in the system database. The 4 anchors most relevant to the current weight vector include: High smoothness anchor =[0.40, 0.20, 0.10, 0.25, 0.05], Balanced anchor =[0.30, 0.30, 0.20, 0.15, 0.05], Time priority anchor =[0.25, 0.35, 0.15, 0.20, 0.05] and Extremely high smoothness anchor =[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 squares method. After optimization calculation, it is obtained that: = 0.3, = 0.2, = 0.1, = 0.4, satisfying the constraint conditions and . Verification result: 0.3 × + 0.2 × + 0.1 × + 0.4 × ≈ [0.35, 0.25, 0.15, 0.20, 0.05] = , and the mapping accuracy meets the requirements.

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

[0117] Operating condition domain interpolation execution: The interpolation calculation layer selects K = 6 nearest neighbor operating condition points from the offline database for interpolation calculation based on the current operating condition state =[25km, 5Ma]. Taking the weight anchor as an example, the nearest neighbor operating condition points selected by the system include:

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

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

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

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

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

[0123] = [25.1 km, 4.7 Ma] (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, and calculates the interpolation weights 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 processing, the system performs weighted average calculation to obtain the local optimal pole . The system performs the same interpolation process on the other three weight anchor points 、 、 respectively calculates the corresponding local optimal poles 、 、 .

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

[0133] Layer 4: Implementation process of the control execution layer

[0134] As Figure 7 shown in the control execution layer architecture, the outer-loop controller receives the altitude and speed trajectories generated by the intelligent decision-making layer, and the inner-loop controller achieves precise control according to the optimal pole configuration provided by the interpolation calculation layer.

[0135] Execution of the outer-loop controller: The outer-loop controller of the control execution layer receives the altitude and speed trajectories generated by the intelligent decision-making layer, and uses the dynamic inversion method to convert the trajectory commands into angle-of-attack and fuel equivalence ratio commands. At the initial moment t = 0 of the task, the commands received by the system include: altitude command (0) = 25 km, speed command (0) = 5 Ma, altitude rate-of-change command (0) = 0 m / s, speed rate-of-change command (0) = 0 Ma / s. According to the dynamic inversion calculation of formula (15), the system obtains the initial control commands: angle-of-attack command (0) = 3.113°, fuel equivalence ratio command (0) = 0.286. At the midpoint moment t = 30 s of the task execution, the command status of the system is updated to: altitude command (30) = 25.5 km, speed command (30) = 5.5 Ma, and the corresponding control command calculation results are: angle-of-attack command (30) = 2.85°, fuel equivalence ratio command (30) = 0.340.

[0136] Execution of the inner-loop controller: The inner-loop controller receives the angle-of-attack and equivalence ratio commands output by the outer-loop controller, and uses the progressive pole scheduling method according to formula (16) to achieve a smooth transition of control parameters. The system sets the adjustment coefficient = 0.08 to ensure the stability of pole updates. The initial pole configuration is set as 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 adjustments, the pole configuration converges smoothly to the target value after approximately 15 control cycles , achieving a disturbance-free transition of the control parameters and ensuring the smoothness of the flight process.

[0137] Layer 5: Control system execution response

[0138] Altitude response performance analysis: As Figure 8 shown, the system altitude response performance is excellent. It successfully realizes the altitude adjustment from 25 km to 26 km within the specified time of 60 seconds. The entire response process presents the characteristics of an ideal S-shaped curve. The response process can be divided into three distinct stages: the first 15 seconds is the slow start stage, where the altitude change rate gradually increases from zero; the middle 30 seconds is the main climbing stage, where the altitude change rate reaches a maximum of approximately 30 m / s, providing the main altitude increment; the last 15 seconds is the stable convergence stage, where the altitude change rate gradually decreases from the maximum to zero, ensuring a smooth arrival at the target altitude. The final control accuracy index of the system shows that the steady-state error is less than 5 m, the total adjustment time is 59.2 seconds, and the overshoot is zero, fully meeting the performance requirements of "slow" adjustment in the natural language instructions.

[0139] Velocity response performance analysis: As Figure 9 shown, the velocity response performance is also excellent. The Mach number smoothly increases from 5 Ma to 6 Ma, and the response curve is highly synchronized with the altitude change, demonstrating the superiority of multivariable coordinated control. The maximum tracking error during the velocity tracking process is controlled within 0.02 Ma, and the entire change process is smooth and continuous without any sudden changes or oscillations. The time constant of the velocity response is perfectly matched with the altitude response, verifying the excellent performance of the outer loop controller in multivariable coupling control and ensuring the velocity-altitude coordination of the aircraft during the climbing process.

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

[0141] Through the above hierarchical implementation verification, the intelligent control system of the hypersonic vehicle based on the large language model of the present invention can effectively understand performance preferences such as "slow" in natural language commands, and through semantic parsing of the intelligent decision-making layer, precise conversion of the weight mapping layer, optimized configuration of the interpolation calculation layer, and precise control of the control execution layer, a complete closed loop from natural language to flight control is realized. Each layer of the system works together to ensure the intelligence, safety and high performance of the control process, providing a complete technical solution for the intelligent development of the hypersonic vehicle control system.

Claims

1. An intelligent control system for hypersonic vehicles based on large language models, characterized in that, It 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: (1) Construct an intelligent decision-making layer to achieve natural language understanding and security verification The intelligent decision-making layer is constructed based on a large language model, responsible for understanding natural language instructions and generating control strategy weights and flight trajectories; this layer includes an instruction preprocessing module, a semantic security fence 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 standardized processing on them, including format standardization, unit conversion, and context information integration; the standardized instruction information output by this module is transmitted to the semantic security fence module; Semantic security fence module: The semantic security fence module receives the standardized instruction information output by the instruction preprocessing module and performs multi-level security verification on the input instructions based on the RoboGuard framework; it judges the feasibility of the instructions by calculating the required average climb rate and acceleration, where the formula for calculating the average climb rate is: (1), where is the average climb rate, is the target altitude, is the current altitude, is the specified time; The formula for calculating the average acceleration is: (2), where is the average acceleration, is the target speed, is the current speed; The system compares the calculated values of the average climb rate and acceleration with the aircraft performance envelope to identify physically unreachable instructions; at the same time, this module also performs logical consistency checks to identify logical contradictions within the instructions; the safety boundary protection function verifies whether the instructions exceed the predefined safety envelope limits to ensure that all verified instructions are within the safe operating range of the aircraft; Trajectory generation module: The trajectory generation module receives the instruction information verified by the semantic security fence module and generates a continuous height-speed trajectory sequence according to the target state and time constraints in the natural language instructions to obtain the desired trajectory; for flight missions, the fifth-order polynomial method is adopted: Height trajectory : (3), speed trajectory : (4), where is the coefficient of the height trajectory polynomial, is the coefficient of the velocity trajectory polynomial, and t is the time variable; Weight generation module: The weight generation module receives the instruction information verified by the semantic security fence module and converts the natural language instructions into a five-dimensional task weight vector through semantic parsing of the large language model: (5), where is the overshoot weight,[[]] is the settling time weight,[[]] is the steady-state error weight,[[]] is the control surface change smoothness weight,[[]] is the robust margin weight, and all weights are non-negative; the weight vector satisfies the normalization constraint: (6), the large language model identifies the task preferences of the operator through a pre-established keyword-to-weight mapping knowledge base; the generated weight vector is passed to the weight mapping layer; (2) Design a weight mapping layer to achieve an accurate conversion from the weight vector to control parameters Establish a set of weight anchor points: The weight mapping layer is responsible for matching the task weight vector output by the intelligent decision-making layer with the offline optimization database to achieve an accurate mapping from the weight space to the control parameter space; by maintaining a predefined set of weight anchor points , these anchor points form a convex hull in the weight space through the convex combination theory, and its mathematical expression is: (7), where is the convex combination coefficient of the i-th weight anchor point, and N is the total number of anchor points; When the large language model generates an arbitrary task weight vector the weight mapping layer first determines whether the vector lies within a predefined convex hull; if it does, the system determines a set of convex combination coefficients by solving a convex optimization problem such that the weight vector can be expressed as a linear combination of the anchor points: (8), Design a weight projection mechanism: For weight vectors not within the convex hull, the system adopts an optimal projection strategy to project them onto the nearest convex hull boundary point. The projection calculation formula is as follows: (9), where is the projected weight vector, represents the two-norm, is the minimum value in (3) Construct an offline optimization database to provide full-coverage optimal control parameters Design a three-dimensional database architecture: The weight dimension of the database corresponds to N predefined weight anchors , each anchor representing a specific combination of control performance preferences; the operating condition dimension only includes height and speed parameters; the optimization pole dimension corresponds to the optimal pole configuration for each weight anchor-operating condition combination; Establish a database index structure: The database index structure is represented in the form of a two-dimensional mapping as: (10), where represents the i-th weighted anchor point, represents the j-th operating condition point, represents the corresponding optimal pole placement; Implement offline multi-objective optimization: The offline optimization process adopts a multi-objective optimization method, and its cost function is designed as: (11), where is the pole placement vector, is the overshoot cost, is the settling time cost, is the steady-state error cost, is the control surface change smoothness cost, is the robust margin cost; (4) Develop an interpolation calculation layer to achieve the optimal pole configuration of double-domain interpolation The interpolation calculation layer implements a double-domain interpolation algorithm, 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 progressive interpolation strategy, first performing interpolation matching in the weight domain, then performing refined interpolation in the operating condition domain, and finally realizing the organic combination of the two domains; Implement the weight domain interpolation algorithm: In the weight domain interpolation stage, the system performs interpolation calculations on predefined weight anchor points using the convex combination coefficients transmitted by the weight mapping layer; specifically, for any task weight vector output by the large language model , the system determines the optimal convex combination coefficients by solving a least squares optimization problem: (12) This process needs to satisfy the normalization constraint and the non-negativity constraint ; Design the operating condition domain interpolation mechanism: During the working condition domain interpolation stage, the system targets the current flight state and each weight anchor point , selects K nearest neighbor working condition points from the corresponding pole database, and uses the radial basis function interpolation method to obtain the local optimal pole configuration; the interpolation calculation formula is: (13) Among them, is the locally optimal pole obtained by interpolation in the operating condition domain; is the interpolation coefficient based on distance weighting, is the bandwidth parameter of the radial basis function, which is used to control the smoothness and locality of interpolation, is the j-th nearest neighbor operating condition point; Execute the final pole synthesis: The final pole synthesis stage organically integrates the interpolation results of the weight domain and the operating condition domain; the system calculates the final optimal pole configuration through a formula: (14) Among them, is the convex combination coefficient obtained by weight domain interpolation; the final pole placement is passed to the control execution layer; (5) Establish a control execution layer to achieve precise flight control The control execution layer includes an outer loop control module and an inner loop control module, which are 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 trajectory generation module of the intelligent decision-making layer, and uses the dynamic inversion method to establish the mapping relationship from the desired trajectory to the inner loop instruction: Among them, is the angle of attack command, is the equivalence ratio command, is the altitude rate of change command, is the speed rate of change command, and are the aircraft dynamics functions; Construct the inner loop control module: The inner loop control module receives the angle of attack and equivalence ratio commands output by the outer loop control module, and based on the adaptive pole placement method, uses the optimal pole placement provided by the interpolation calculation layer to achieve precise attitude and propulsion system control; The pole scheduling adopts a progressive adjustment method: (16) Among them, is an adjustment coefficient that controls the pole update speed, is the control step number.

2. The intelligent control system for hypersonic vehicle based on large language model according to claim 1, characterized in that For the instruction input of the instruction preprocessing module, the flight instruction format is: "Adjust the altitude to H kilometers and the speed to M Mach within T seconds, and the required performance preference is 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 intelligent control system for hypersonic vehicles based on large language models according to claim 1, characterized in that, The process by which the large language model identifies the task preferences of the operator is as follows: For time - priority tasks, when the instruction contains keywords such as "as soon as possible", "quickly", "urgently", the system increases the value of the adjusted time weight. For smooth - priority tasks, when keywords such as "smooth", "gentle", "comfortable" are detected, the system increases the value of the overshoot weight. For accuracy - priority tasks, when keywords such as "accurate", "precise", "error - free" are present, the system increases the steady - state error weight. For safety - priority tasks, when keywords such as "safe", "stable", "conservative" are detected, the system increases the robust margin weight. For comfort - priority tasks, when keywords such as "comfortable", "smooth", "load - reduction" are present, the system increases the value of the rudder - surface change smoothness weight. The value.

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

Images