A meta-learning-based adaptive preset performance control method for spacecraft

Through the adaptive preset performance control method based on meta-learning, the problems of insufficient robustness and high computational complexity in complex environments are solved, flexible preset performance control and efficient adaptability are achieved, and the Control performance and robustness of space vehicles.

CN118732511BActive Publication Date: 2024-11-22DALIAN UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411230031.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-04
Publication Date
2024-11-22
Estimated Expiration
2044-09-04

AI Technical Summary

Technical Problem

The existing technology is not robust enough when facing complex and dynamic environments, Koopman operator technology has high computational complexity, and the iterative learning method of feedforward adjustment is sensitive to noise. Heuristic dynamic programming may lead to Optimal control strategy limits the effect of adaptive preset performance control of space aircraft.

Method used

Adaptive preset performance control method based on meta-learning is adopted to establish a second-order nonlinear dynamic model containing perturbation terms, design a flexible preset performance function and error conversion system, introduce auxiliary variables to process input saturation, and use neural network model to online Training adaptive parameters to optimize controller performance.

Benefits of technology

It realizes that while meeting preset performance indicators, it effectively responds to input saturation, improves the robustness and adaptability of the system, reduces the demand for computing resources, and improves control accuracy and generalization capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118732511B_ABST
    Figure CN118732511B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of intelligent control technology, and proposes a spacecraft adaptive preset performance control method based on meta-learning. The method fully considers the input saturation problem of the system, and designs a flexible preset performance controller, so that the system can effectively cope with the limitation of input saturation while meeting the preset performance indicators. For the nominal controller, the meta-learning method is used to train the dynamic model offline, and the dynamic characteristics of the actual system are approximated through a variety of simulation scenarios to improve the accuracy and generalization ability of the model. According to the offline model, the control parameters are optimized under different simulation scenarios to obtain a set of optimal parameters for the nominal controller, thereby enhancing the adaptability of the control system to various complex environments and achieving more efficient and robust control performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of intelligent control technology and relates to a spacecraft adaptive preset performance control method based on meta-learning. Background Art

[0002] In the field of space vehicle control, control performance is often affected by the dynamic system composed of the vehicle itself and its interaction with the complex space environment. Since the environment of the space vehicle is extremely complex and changeable, the dynamic characteristics of the vehicle's attitude may change unpredictably, which makes the control system of the space vehicle face severe challenges. Considering the uncertainty of the mathematical model of the spacecraft, the nonlinear control method can achieve compensation by estimating the parameters online and approximating the uncertainty of the system, so as to achieve the expected control effect. In the face of uncertain disturbance factors in the space environment, the control system of the space vehicle can adopt an adaptive control method to dynamically respond to these external disturbances and internal uncertainties by adjusting the controller parameters in real time, and can be combined with other control methods to further improve the performance of the control system.

[0003] The preset performance control method introduces performance functions and error transformations to achieve preset performance for convergence speed, overshoot, and tracking error, so that the transient performance and steady-state performance of the aircraft control system have clear upper bounds. In recent years, the preset performance method has been widely used in the fields of flexible joint robot systems, servo systems, air-breathing hypersonic aircraft systems, and random systems. In actual systems, aerospace vehicles often experience input saturation and failure problems. In order to ensure the preset performance in the presence of these uncertainties and nonlinearities, a non-negative correction signal can be introduced to dynamically adjust the performance boundary, so that the performance constraints can be relaxed when input saturation occurs, and the preset performance boundary can be restored after the saturation disappears. This can effectively solve the vulnerability problem of preset performance control when the actuator is saturated, and ensure the stability and robustness of system performance.

[0004] In recent years, research on meta-learning in the field of control has also increased significantly. Meta-learning can adapt to new task requirements more quickly by learning from past data. Specifically, meta-learning collects input and output data on real systems and trains a parameterized adaptive controller that can adapt well to each model in the model set built from past input and output data. In the field of aircraft control, meta-learning methods can not only effectively deal with uncertainties in complex environments, but can also be combined with adaptive control to further improve the robustness and adaptability of the system.

[0005] Wan Lei, Zeng Jiangfeng, Li Yueming, etc., in "An Intelligent Learning Control Method for Tracking Unmanned Boat Tracks" (CN109283842B), adopted a fuzzy adaptive controller based on model online learning, and used a product inference engine to realize the reasoning of rule premises and rule conclusions, which can effectively shorten the adjustment time of the unmanned boat to reach the desired path, enhance the robustness of the control system, and reduce the dependence on the precise mathematical model of the unmanned boat, with a faster tracking control response speed and a smaller tracking error. Although the fuzzy adaptive controller in the patent adopts a model-based online learning method and uses a product inference engine for rule reasoning, its adaptability to complex and dynamic environments may be limited. However, in some extreme cases, such as scenarios where environmental parameters change dramatically or there is a high degree of uncertainty, fuzzy rules may not be sufficient to provide sufficient robustness, resulting in a decrease in the performance of the control system.

[0006] Zhang Xinglong, Lu Yang, Xu Xin, etc. used a data-driven method in “A data-driven robust learning predictive control method for robots” (CN110647042B) to abstract the nonlinear characteristics of the robot into a high-dimensional linear space to form a global linearization model. Based on the global linearization model and the state variables in the high-dimensional linear space, rolling optimization is performed to learn the approximate optimal predictive control sequence of the robot system at the current moment. This method converts complex nonlinear systems into processable linear models through data-driven and Koopman operator technology, thereby improving the description accuracy and control effect of the nonlinear characteristics of the system. However, the Koopman operator technology linearizes the state variables in a high-dimensional space, and this process usually requires a lot of computing resources. Especially in real-time control applications, this computational complexity may affect the response speed and real-time performance of the system, thereby limiting its application in resource-constrained environments.

[0007] In "Data-driven control method combining feedforward tuning with iterative learning" (see "Control Theory and Applications", June 2020, Issue 6, pp. 1376-1376), Bo Yumeng, Cao Mingsheng, and Gao Huibin introduced a parameterized feedforward controller and an input shaping filter, and used the gradient descent method to solve the optimal system feedforward controller. A hybrid data-driven control method that combines feedforward tuning with iterative learning control improves control accuracy and reduces dependence on system models. Although this method improves control accuracy, in actual applications, the feedforward controller may be more sensitive to system noise and external interference. If there are unmodeled dynamics or noise in the system, the effect of feedforward tuning may be affected, resulting in a decrease in controller performance.

[0008] In “Data-driven nonlinear approximate optimal regulation based on iterative neural dynamic programming” (Acta Automatica Sinica, March 2017, No. 3, pp. 366-375), Wang Ding, Mu Chaoxu, and Liu Derong utilized the adaptive and self-learning characteristics of neural networks to construct three neural networks (model network, judgment network, and execution network), and used heuristic dynamic programming technology to train the execution network so that it does not rely on system dynamic information and achieves the control effect of the approximate optimal control method. However, although heuristic dynamic programming helps to train the execution network to achieve approximate optimal control, it is essentially an approximate algorithm and may not always find the global optimal solution. In some cases, this approach may lead to suboptimal control strategies, thereby affecting the overall control performance. Summary of the invention

[0009] In order to solve the above problems, the present invention provides a spacecraft adaptive preset performance control method based on meta-learning.

[0010] A spacecraft adaptive preset performance control method based on meta-learning, comprising the following steps:

[0011] Step 1: Establish a second-order nonlinear dynamic model of the spacecraft with disturbance terms;

[0012] Step 2: Based on the second-order nonlinear dynamics model, a preset performance function with flexible boundaries is designed, and the corresponding error conversion system is constructed; while ensuring the system performance, an auxiliary variable η(t) is introduced to effectively deal with the common input saturation problem in actual control systems. Specifically, the auxiliary variable adaptively adjusts the performance boundary to avoid the situation in which the tracking error due to input saturation in the traditional method touches or exceeds the performance constraint boundary;

[0013] Step 3: Design the Lyapunov function based on the constructed error conversion system, and design the virtual control law and control signal based on the derivative of the Lyapunov function;

[0014] Step 4: Randomly generate u corresponding trajectories according to the designed control law and system dynamics model. The trajectory contains the state variables x1, x2, control law and and u data; by discretizing the trajectory data and selecting part of it as training data, the neural network model is trained to approximate the dynamic behavior of the system;

[0015] Step 5: Define the meta-training problem and obtain the optimal adaptive parameters based on the trained neural network model through online training.

[0016] Furthermore, in the first step, the second-order nonlinear dynamic model containing the disturbance term is established as follows:

[0017]

[0018] Among them, x1, x2 are the state variables of the system, f2 is an uncertain continuous nonlinear function, g2 is a constant and is not 0; d is the disturbance term of the nonlinear second-order system, v represents the ideal control input signal of the nonlinear second-order system, u(v) represents the corresponding actual control input signal, y represents the output of the nonlinear second-order system, and the control purpose is to design the actual controller input u(v) so that the system output y tracks the expected output signal y d Assume that d and y d are all bounded, y d The first-order derivative of is also bounded. In addition, the control saturation is expressed as follows:

[0019]

[0020] In the formula, u d represents the upper bound of input saturation, and sgn(·) represents the sign function.

[0021] The second step is as follows: define the error variable e1 = x1-y d , construct the preset performance function ρ(t):

[0022]

[0023] Among them, ρ0, ρ T ,T,α * are the positive parameters to be designed, which are the initial performance index value, the performance index value at stable state, the preset time, and the adjustment coefficient; t represents time, csch represents the hyperbolic tangent function, and the error conversion system ζ(t) is constructed:

[0024]

[0025] Among them, s(t) represents the auxiliary variable, E l (t) and E h (t) represents the lower and upper bounds of the error, defined as [E l (t),E h (t)] T =[e l (t),e h (t)] T +Λ1tanh(η(t)),e l (t) and e h (t) represent the upper and lower bounds of the error signal, respectively, e1(0) represents the initial value of the error, I2 represents the unit column vector, defined as I2 = [1 1] T , Λ k Represents the matrix or vector used to adjust the dynamic response of the system, defined as Λ p=[-λ p λ p ] T ,p=1,2,λ p represents the system adjustment parameter. The update law of the auxiliary variable η(t) is designed as:

[0026]

[0027] Where, σ1(t)=[sgn(vu d )+1](vu d ),σ2(t)=[sgn(v+u d )-1](v+u d ), m1 and m2 are positive design parameters used to adjust the dynamic response of the auxiliary system.

[0028] The third step is to define the error variable according to the nonlinear second-order system with disturbance term in, Represents the virtual control quantity to be designed. According to the error ζ(t) and e2, the Lyapunov function V is constructed:

[0029]

[0030] Taking the derivative of the Lyapunov function, we get

[0031]

[0032] in,

[0033]

[0034] The control law designed based on the Lyapunov function is expressed as:

[0035]

[0036] Among them, k1 and k2 represent positive parameters to be designed.

[0037] In the fourth step, the control law obtained in the third step is used as the nominal controller, and a smooth trajectory of m seconds is generated according to the nominal controller. In order to introduce different types of disturbances, j disturbance schemes are set, and a total of The trajectory dataset under the jth disturbance is defined as This collection contains a series of moments State variables parameter Control Input And the updated state variables k represents the discrete time series. By discretizing, we select As the offline training data set, the training set is defined as Training a deep neural network model Enable them to learn The input-output data of the trajectory accurately fits and approximates the nonlinear dynamic characteristics of the system, where Ψ j represents the set of parameters of a deep neural network. The goal of this training process is defined as follows:

[0038]

[0039] in, represents the jth track in the reference trajectory, Respectively represent the state quantity The estimated value of .

[0040] The fifth step is to define the adaptive controller to be optimized parameter θ = (k1, k2, θ Π ,Γ), where θ Π , Γ represent the parameters in the update law of the adaptive law and the adaptive law gain respectively. The goal of meta-learning is to find a set of optimal controller parameters θ through an offline optimization process. * , so that the adaptive controller can perform well under different trajectory and disturbance conditions. The optimal goal of training parameters is:

[0041]

[0042] in, Represents different nominal trajectories, each of which corresponds to an independent task. Represents the trajectory data under different disturbance conditions collected under each reference trajectory. Represents the time interval of the nominal trajectory. ctrl represents the regularization coefficient of the meta-learning loss function, μ meta represents the regularization coefficient, a ij represents the adaptive term, Π represents the adaptive update function, and π represents the actual control law function after adding the adaptive term. After optimization, a set of optimal control parameters θ is obtained. * , so that good control performance can be achieved under each training trajectory and disturbance. This approach ensures that the controller not only performs well on the training data, but also generalizes to unseen trajectories and disturbances.

[0043] Beneficial effects of the present invention:

[0044] 1. A flexible preset performance controller is designed to enable the system to effectively cope with input saturation constraints while meeting preset performance indicators. This feature enables the control system to maintain good performance when facing input constraints, while existing data-driven control methods do not consider state constraints and input constraints.

[0045] 2. The meta-learning method is used to train the dynamic model offline, and the dynamic characteristics of the actual system are approximated through a variety of simulation scenarios, which improves the accuracy and generalization ability of the model. This has better adaptability and robustness than the fuzzy adaptive controller based on preset rules and fuzzy sets and the neural dynamic programming method based on discrete models.

[0046] 3. Optimize the control parameters under different simulation scenarios to obtain a set of optimal parameters for the nominal controller, which enhances the control system's ability to adapt to various complex environments. In contrast, the data-driven control method combined with feedforward tuning and iterative learning requires retraining when facing new tasks or environmental changes, and has poor adaptability.

[0047] 4. Data-driven control methods often require the training of multiple neural networks, with high algorithm complexity, large computing resource requirements, and long training time. The method of the present invention reduces the demand for computing resources and improves training efficiency by combining meta-learning offline training with flexible preset performance control. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the basic flow chart of the present invention;

[0049] Figure 2 is the spacecraft attitude angle variation curve;

[0050] Figure 3 is the spacecraft attitude angular velocity change curve;

[0051] Figure 4 is the spacecraft control torque variation curve;

[0052] Figure 5 is the spacecraft attitude angle error variation curve;

[0053] Figure 6 is the spacecraft attitude angular velocity error variation curve;

[0054] Figure 7 is the attitude angle change curve of the hypersonic vehicle;

[0055] Figure 8 is the attitude angular velocity change curve of the hypersonic aircraft;

[0056] Fig. 9 It is the rudder deflection command change curve of the hypersonic aircraft. DETAILED DESCRIPTION

[0057] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0058] The basic process of the method of the present invention is as follows Figure 1 As shown, two specific examples are provided below for further explanation.

[0059] Example 1:

[0060] Taking the spacecraft attitude control problem as an example, the following solution steps are given:

[0061] Step 1: Establish a spacecraft attitude dynamics model. Consider the following attitude dynamics model of a flexible spacecraft with solar panels:

[0062]

[0063] Among them: represents the angular velocity of the star relative to the inertial coordinate system, Represents three-dimensional space. d Represents the external disturbance torque from the environment. represents the satellite attitude angle vector (roll, pitch and yaw angles respectively), and

[0064]

[0065] In the formula is the inertia matrix of the entire star, is the control torque vector, ω0 is the orbital angular velocity, is the external disturbance torque, represents the angular momentum vector of the flywheel, η is the vibration modal coordinate vector of the flexible attachment, δ is the coupling coefficient matrix between the satellite body and the solar panel, and N is the number of solar panel modes considered. ξ and Λ represent the damping and stiffness matrix vectors.

[0066] Under the small angle assumption, the attitude dynamics of a flexible spacecraft with solar panels can be expressed in a unified nonlinear form:

[0067]

[0068] in, v n represents the system control input, and f 2,n and d n They are nonlinear continuous function and external interference respectively.

[0069] Step 2: Define the reference signal y based on the spacecraft attitude dynamics model d,n and the error variable e 1,n =x1,n -y d,n , and then construct the preset performance function:

[0070]

[0071] Among them, ρ 0,n represents the initial performance index, T n Indicates the preset time. Represents the adjustment coefficient, which is used to adjust the rate of change of performance indicators, ρ T,n Represents the performance index when stable. And constructs the error transformation:

[0072]

[0073] Among them, s n (t) represents auxiliary variable, E l,n (t) and E h,n (t) represents the lower and upper bounds of the error, defined as [E l,n (t),E h,n (t)] T =[e l,n (t),e h,n (t)] T +Λ 1,n tanh(η n (t)), e l,n and e h,n They represent the upper and lower bounds of the error signal, respectively, and are defined as [e l,n (t),e h,n (t)] T =sgn(e1(0))(ρ n (t)-ρ T,n )I 2,n +Λ 2,n ρ n (t), I2 represents the unit column vector, which is defined as I2 = [1 1] T , Λ k,n Represents the matrix or vector used to adjust the dynamic response of the system, defined as Λ p,n =[-λ p,n λ p,n ] T ,p=1,2. The update law of the auxiliary variable η is designed as:

[0074]

[0075] Among them, σ 1,n (t) = [sgn(v n -u d,n )+1](v n -u d,n ), σ2,n (t) = [sgn(v n +u d,n )-1](v n +u d,n ), m 1,n and m 2,n is a positive design parameter used to adjust the dynamic response of the auxiliary system.

[0076] Step 3: Define the error variable based on the second-order system dynamics model of the spacecraft in, Is the virtual control instruction of the system. Construct Lyapunov function V:

[0077]

[0078] Taking the derivative of the Lyapunov function, we get

[0079]

[0080] in,

[0081]

[0082] The control law designed based on the Lyapunov function is expressed as:

[0083]

[0084] Among them, k 1,n , k 2,n is represented as a positive parameter to be designed.

[0085] Step 4: Use the above control law as the nominal controller and combine it with the spacecraft dynamics model for each attitude angle. In order to introduce different types of disturbances, j disturbance schemes are set, and a total of The trajectory set under the j-th disturbance is defined as This collection contains a series of moments State variables parameter Control Input And the updated state variables Choose from all perturbation options As the offline training data set, the training set is defined as Training a deep neural network model Enable them to learn The input-output data of the trajectory accurately fits and approximates the nonlinear dynamic characteristics of the system, where Ψ jrepresents the set of parameters of a deep neural network. The goal of this training process is defined as follows:

[0086]

[0087] Step 5: Define the adaptive controller parameter to be optimized θ = (k 1,n ,k 2,n ,θ Π,n ,Γ n ), where θ Π , Γ represents the parameters in the update law of the adaptive law and the adaptive law gain. The goal of meta-learning is to find a set of optimal controller parameters θ through an offline optimization process. * , so that the adaptive controller can perform well under different trajectory and disturbance conditions. The optimal goal of training parameters is:

[0088]

[0089] in, Represents different nominal trajectories, each of which corresponds to an independent task. Represents the trajectory data under different disturbance conditions collected under each reference trajectory. Represents the time interval of the nominal trajectory. ctrl represents the regularization coefficient of the meta-learning loss function, μ meta represents the regularization coefficient, a n Represents the adaptive parameters in this process, and Π represents the update rule of the adaptive law. After optimization, a set of optimal control parameters θ is obtained, which can achieve good control performance under each training trajectory and disturbance. This method ensures that the controller not only performs well on the training data, but also can be generalized to unseen trajectories and disturbances.

[0090] In order to verify the effectiveness of the proposed control technology, the following system parameters are selected to simulate the spacecraft attitude tracking control task. The nominal value of the moment of inertia J is

[0091]

[0092] External disturbance torque T from the environment d

[0093]

[0094] The initial value of the spacecraft attitude variable is Θ = [0.25 -0.21 0.1] T , the initial attitude angular velocity of the spacecraft is ω=[0 0 0] T deg / s. The desired attitude angle of the spacecraft is set to Θ d =[0 0 0]T .

[0095] The training data of the meta-learning module is generated as follows:

[0096] The initial three degrees of freedom angles of the spacecraft are randomly generated in the range of [-5,5], and the angular velocity values ​​are randomly generated in the range of [-10,10]. When using dynamics to fit the trajectory of the spacecraft's attitude change, the trajectory generation time range is 0 to 300 seconds, and the sampling frequency is 0.1s. The disturbance torque on the spacecraft is set to T d +δT d , where δT d Represented as Gaussian noise with mean 0 and variance 0.01. Regularization coefficient μ of meta-learning loss function ctrl Set to 0.01, the regularization coefficient μ meta Set to 0.001. The controller initialization parameters are set to diag(0.1,0.1), the adaptive parameters and adaptive update law are set to diag(1,1,1), diag(0.1,0.1,0.1). The Adam optimizer is used in the optimization process, the number of training rounds is 500, and the optimized controller parameters are k 1,n =18, k 2,n =0.8, adaptive update law Γ n =10, the adaptive parameter is θ Π,n =0.01. In addition, ρ 0,n =5, T n =10, ρ T,n = 1. The simulation curve obtained by optimizing the control parameters after training is as follows Figures 2 to 6 shown.

[0097] According to the simulation results, the response of attitude angle and attitude angular velocity has obvious oscillation in the initial stage (the first 50 seconds), especially the pitch angle and pitch angular velocity. However, the oscillation gradually weakened and stabilized after about 100 seconds, indicating that the system has good stability and the effectiveness of the controller design. Despite the large oscillation in the initial stage, the system was able to recover stability in a short time, showing good robustness and anti-interference ability.

[0098] Example 2:

[0099] Taking the attitude control problem of a hypersonic vehicle as an example, the following solution steps are given:

[0100] Step 1: Establish the attitude dynamics model of the hypersonic vehicle as follows:

[0101]

[0102] In the formula, represents the attitude angle, where σ is the roll angle, β is the sideslip angle, and α is the angle of attack; represents the attitude angular velocity, where p is the roll angular velocity, q is the yaw angular velocity, and r is the pitch angular velocity; Represents the control rudder deflection command, where δ a is the right elevator deflection command, δ e is the left elevator deflection command, δ r is the rudder deflection command; ψ is the conversion matrix between aerodynamic torque and aerodynamic control surface deflection, ζ represents the influence of uncertain aerodynamic force; the coordinate transformation matrix R, antisymmetric matrix Ω, and inertia matrix J are respectively

[0103]

[0104] Let x1 = γ and The model can be transformed into a control-oriented dynamics equation:

[0105]

[0106] Where: It is a nonlinear continuous function, g2=RJ -1 ψ is the nominal known quantity, u2(v) = δ is the control input quantity, d = RJ -1 ζ is external interference,

[0107] Step 2: Based on the control-oriented hypersonic vehicle dynamics model, define the reference signal y d , error variable e1 = x1-y d , and then construct the preset performance function:

[0108]

[0109] Among them, ρ0, ρ T ,T,α * are the positive parameters to be designed, which are the initial performance index value, the performance index value at stable state, the preset time, and the adjustment coefficient. t represents time, csch represents the hyperbolic tangent function, and the error conversion system ζ(t) is constructed:

[0110]

[0111] Among them, [E l (t),E h (t)] T =[e l (t),e h (t)] T + Λ1tanh(η(t)), and

[0112] [e l(t),e h (t)] T =sgn(e1(0))(ρ(t)-ρ T )I2+Λ2ρ(t),I2=[1 1] T , Λ p =[-λ p λ p ] T ,p=1,2. The update law of the auxiliary variable η(t) is designed as:

[0113]

[0114] Where, σ1(t)=[sgn(vu d )+1](vu d ),σ2(t)=[sgn(v+u d )-1](v+u d ), m1 and m2 are positive design parameters used to adjust the dynamic response of the auxiliary system.

[0115] Step 3: Define the error variable in, is the virtual control instruction of the system. The design control law expression is:

[0116]

[0117] Among them, k1 and k2 represent positive parameters to be designed.

[0118] Step 4: Use the above control law as the nominal controller and generate a smooth trajectory of m seconds according to the nominal controller. In order to introduce different types of disturbances, j disturbance schemes are set, and a total of The trajectory dataset under the jth disturbance is defined as This collection contains a series of moments State variables parameter Control Input And the updated state variables k represents the discrete time series. By discretizing, we select As the offline training data set, the training set is defined as Training a deep neural network model Enable them to learn The input-output data of the trajectory accurately fits and approximates the nonlinear dynamic characteristics of the system, where Ψ j represents the set of parameters of a deep neural network. The goal of this training process is defined as follows:

[0119]

[0120] Step 5: Define the adaptive controller parameters to be optimized θ = (k1, k2, θ Π ,Γ), where θ Π , Γ represents the parameters in the update law of the adaptive law and the adaptive law gain. The goal of meta-learning is to find a set of optimal controller parameters θ through an offline optimization process. * , so that the adaptive controller can perform well under different trajectory and disturbance conditions. The optimal goal of training parameters is:

[0121]

[0122] in, Represents different nominal trajectories, each of which corresponds to an independent task. Represents the trajectory data under different disturbance conditions collected under each reference trajectory. Represents the time interval of the nominal trajectory. ctrl represents the regularization coefficient of the meta-learning loss function, μ meta represents the regularization coefficient, a ij represents the adaptive term, Π represents the adaptive update function, and π represents the actual control law function after adding the adaptive term. After optimization, a set of optimal control parameters θ is obtained. * , so that good control performance can be achieved under each training trajectory and disturbance. This approach ensures that the controller not only performs well on the training data, but also generalizes to unseen trajectories and disturbances.

[0123] In order to verify the effectiveness of the proposed control technology, the following system parameters are selected to simulate the hypersonic vehicle attitude tracking control task. The nominal value of the moment of inertia J is

[0124]

[0125] The initial value of the hypersonic vehicle attitude is γ = [30 1 30] T , the initial attitude angular velocity of the hypersonic vehicle is ω=[0 0 0] T deg / s. The desired attitude angle of the hypersonic vehicle is set to Θ d =[5 0 70] T .

[0126] The training data of the meta-learning module is generated as follows:

[0127] The initial three degrees of freedom angle values ​​of the hypersonic aircraft are randomly generated in the range of [-50,50], and the angular velocity values ​​are randomly generated in the range of [-10,10]. When using dynamics to fit the hypersonic aircraft attitude change trajectory, the trajectory generation time range is 0 to 200 seconds, and the sampling frequency is 0.01s. The interference torque is set to Gaussian noise with a mean of 0 and a variance of 0.01. The regularization coefficient μ of the meta-learning loss function ctrl Set to 0.01, the regularization coefficient μ meta Set to 0.001. The controller initialization parameters are set to diag(0.1,0.1), the adaptive parameters and adaptive update law are set to diag(1,1,1), diag(0.1,0.1,0.1). The Adam optimizer is used in the optimization process, and the number of training rounds is 500. The number of training rounds is 500, and the optimized controller parameters are k1=[25,10,25], k2=[1.4,0.7,1.4], and the adaptive update law Γ n =10, the adaptive parameter is θ Π,n =0.01. In addition, ρ0=10, T n =10, ρ T,n =1.

[0128] The simulation results obtained by optimizing the control parameters after training are as follows: Figures 7 to 9 As shown. Among them, Figure 7 The attitude angle change curve of the hypersonic aircraft is given, including the roll angle, sideslip angle, and angle of attack. Figure 8 The time variation curves of roll angular velocity, yaw angular velocity, and pitch angular velocity are described. It can be seen that under the proposed meta-learning-based adaptive preset performance control method, the hypersonic aircraft exhibits a good attitude control effect. Furthermore, Fig. 9 The corresponding control input curves are given, including the left elevator command, the right elevator command and the rudder command, and the deflection angles of the control surfaces are all within a reasonable constraint range.

Claims

1. A spacecraft adaptive preset performance control method based on meta-learning, characterized in that: The following steps are involved: Step 1: Establish a second-order nonlinear dynamic model of the spacecraft with disturbance terms; Step 2: Based on the second-order nonlinear dynamics model, a preset performance function with flexible boundaries is designed and the corresponding error conversion system is constructed. While ensuring the system performance, an auxiliary variable η(t) is introduced to deal with the input saturation problem that is common in actual control systems. Step 3: Design the Lyapunov function based on the constructed error conversion system, and design the virtual control law and control signal based on the derivative of the Lyapunov function; Step 4: Randomly generate according to the designed control law and system dynamics model The corresponding trajectory contains the state variables x1, x2, and control law under different disturbances within m seconds. and u data; by discretizing the trajectory data and selecting part of it as training data, the neural network model is trained to approximate the dynamic behavior of the system; Step 5: Define the meta-training problem and obtain the optimal adaptive parameters based on the trained neural network model online training; Furthermore, in the first step, the second-order nonlinear dynamic model containing the disturbance term is established as follows: Among them, x1, x2 are the state variables of the system, f2 is an uncertain continuous nonlinear function, g2 is a constant and is not 0; d is the disturbance term of the nonlinear second-order system, v represents the ideal control input signal of the nonlinear second-order system, u(v) represents the corresponding actual control input signal, y represents the output of the nonlinear second-order system, and the control purpose is to design the actual controller input u(v) so that the system output y tracks the expected output signal y d ; Assume that d and y d are all bounded, y d The first-order derivative of is also bounded; in addition, the control saturation is expressed as follows: In the formula, u d represents the upper bound of input saturation, sgn(·) represents the sign function; The second step is as follows: define the error variable e1 = x1-y d , construct the preset performance function ρ(t): Among them, ρ0, ρ T ,T,α * are the positive parameters to be designed, which are the initial performance index value, the performance index value at stable state, the preset time, and the adjustment coefficient; t represents time, csch represents the hyperbolic tangent function, and the error conversion system ζ(t) is constructed: Among them, s(t) represents the auxiliary variable, E l (t) and E h (t) represents the lower and upper bounds of the error, defined as [E l (t),E h (t)] T =[e l (t),e h (t)] T +Λ1tanh(η(t)),e l (t) and e h (t) represent the upper and lower bounds of the error signal, respectively, [e l (t),e h (t)] T =sgn(e1(0))(ρ(t)-ρ T )I2+Λ2ρ(t), e1(0) represents the initial value of the error, I2 represents the unit column vector, and is defined as I2=[11] T , Λ k Represents the matrix or vector used to adjust the dynamic response of the system, defined as Λ p =[-λ p λ p ] T ,p=1,2,λ p represents the system adjustment parameter; the update law of the auxiliary variable η(t) is designed as: Where, σ1(t)=[sgn(vu d )+1](vu d ),σ2(t)=[sgn(v+u d )-1](v+u d ), m1 and m2 are positive design parameters used to adjust the dynamic response of the auxiliary system; The third step is to define the error variable according to the nonlinear second-order system with disturbance term in, Represents the virtual control quantity to be designed; construct the Lyapunov function V according to the error ζ(t) and e2: Taking the derivative of the Lyapunov function, we get in, The control law designed based on the Lyapunov function is expressed as: Among them, k1 and k2 represent the positive parameters to be designed; In the fourth step, the control law obtained in the third step is used as the nominal controller, and a smooth trajectory of m seconds is generated according to the nominal controller. The initial value of the state quantity of each trajectory is different; in order to introduce different types of disturbances, j disturbance schemes are set, and a total of reference trajectories; the trajectory dataset under the jth disturbance is defined as This collection contains a series of moments State variables parameter Control Input And the updated state variables k represents a discrete time series; it is selected from all perturbation schemes by discretization As the offline training data set, the training set is defined as Training a deep neural network model Enable them to learn The input-output data of the trajectory accurately fits and approximates the nonlinear dynamic characteristics of the system, where Ψ j represents the set of parameters of a deep neural network; the goal of this training process is defined as follows: in, represents the jth track in the reference trajectory, Respectively represent the state quantity An estimated value of The fifth step is to define the adaptive controller to be optimized parameter θ = (k1, k2, θ Π ,Γ), where θ Π , Γ represent the parameters in the update law of the adaptive law and the adaptive law gain respectively; the goal of meta-learning is to find a set of optimal controller parameters θ through an offline optimization process * , so that the adaptive controller can perform well under different trajectory and disturbance conditions; the optimal goal of the training parameters is: in, Represents different nominal trajectories, each of which corresponds to an independent task. represents the trajectory data under different disturbance conditions collected under each reference trajectory; represents the time interval of the nominal trajectory; μ ctrl represents the regularization coefficient of the meta-learning loss function, μ meta represents the regularization coefficient, a ij represents the adaptive term, Π represents the adaptive update function, and π represents the actual control law function after adding the adaptive term; after optimization, a set of optimal control parameters θ is obtained * .