Intelligent control method of hypersonic shape-shifting aircraft driven by knowledge and data hybridization
Through the intelligent control method driven by knowledge and data, combined with flight action mechanics model and neural network compensation, the control problem of hypersonic deformed aircraft under complex interference is solved, the system stability and high-precision control are achieved, and the aircraft's stability is improved. Robustness and adaptability.
Patent Information
- Application Number
- CN202411223841.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-03
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-09-03
AI Technical Summary
The control problems of hypersonic deformed aircraft under complex interference are difficult to achieve system stability and high-precision control in the existing technology. Methods that rely solely on model knowledge or experimental data cannot fully meet the design needs.
Using a hybrid intelligent control method driven by knowledge and data, establish a knowledge-based flight action mechanics model, and combine neural network compensation and meta-learning algorithms to design a benchmark control law based on offline data and an online intelligent controller, and adjust the parameters by data-driven and compensators to enhance the robustness and adaptability of the control system.
The stability and reliability of hypersonic deformed aircraft under complex interference has been achieved, the intelligence level and engineering application value of the control system have been significantly improved, and the development of hypersonic flight technology has been promoted.
Smart Images

Figure CN118732589B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of hypersonic aircraft control, and relates to an intelligent control method for a hypersonic deformable aircraft driven by a hybrid of knowledge and data. Background Art
[0002] With the increasing diversity of aircraft missions, high-speed aircraft with wide speed range and cross-airspace characteristics have become one of the frontier research hotspots in the field of aerospace today. Compared with traditional aircraft, high-speed aircraft show new technical features such as wide flight envelope, variant layout and demand for multi-element hybrid control effects. Hypersonic deformable aircraft can obtain better aerodynamic shape and more efficient energy utilization by changing its aerodynamic configuration, thereby achieving stronger environmental adaptability and flight mission adjustment capabilities. However, the deformability of the aircraft not only improves its aerodynamic characteristics, but also brings great challenges to the design of the control system. On the one hand, it is difficult to model the deformation process and it is difficult to obtain an accurate aircraft system model, which makes the unmodeled dynamics that need to be considered in the control process. On the other hand, the aircraft system exhibits strong coupling, strong nonlinearity and multi-modal characteristics during the deformation process. Internal and external disturbances during the deformation process are easy to cause system uncertainty, which may cause the closed-loop system to diverge, greatly increasing the difficulty of precise control.
[0003] In "An Intelligent Deformation Decision Method for High-Speed Deformable Aircraft" (see "Journal of Astronautics", December 2022, Vol. 43, No. 12, pp. 1665-1675), Zhang Yuan, Huang Wanwei, Nie Ying, etc. proposed an intelligent deformation decision method based on a deep deterministic strategy algorithm for the deformation decision problem of a type of high-speed deformable aircraft with continuously variable sweep angles. Simulations have proved that the intelligent deformation strategy can meet comprehensive performance indicators including aerodynamic performance and control error. However, the designed intelligent deformation scheme relies on the effect of the reinforcement learning algorithm, and it is necessary to ensure that the algorithm has strong generalization. In actual engineering applications, it may be difficult for the system to converge with changes in complex working conditions, which makes the system unstable.
[0004] In "Event-triggered finite-time control of hypersonic shape-shifting aircraft" (see Acta Aeronautica Sinica, Vol. 44, No. 15, August 2023, p. 528494-1-14), Zhang Hao, Wang Peng, Tang Guojian, etc. proposed a new event-triggered fast finite-time control method to address the problem of high-performance control and low resource consumption in the deformation flight phase of hypersonic shape-shifting aircraft. Although this method makes full use of dynamic models and advanced control strategies as knowledge-driven solutions, its algorithm design is complex and the control effect depends on the results of parameter tuning, which limits the use of this method in practical engineering applications.
[0005] In "A neural network-based time-scale separation aircraft elastic body intelligent control method" (CN108333939B), Zhang Qi decomposed the fast and slow time scales of the hypersonic aircraft, designed an intelligent control strategy, and designed a sliding mode control method for the fast-changing time scale part that characterizes the elastic mode of the system to solve the problem that the existing elastic hypersonic aircraft control method cannot achieve high-precision control under the influence of uncertain information and elastic vibration. However, this method designs complex control laws, update laws, and switching rules when dealing with flight control. For flight missions under the large package line, a lot of control parameter adjustments are required to ensure the improvement of control performance.
[0006] Wang Peng and Tang Guojian, in "A hypersonic aircraft expert intelligent control method and aircraft" (CN110244752A), determined the real-time control parameters of the PID controller of the hypersonic aircraft according to the real-time deformation state of the hypersonic aircraft by building an expert system, realized the intelligent adjustment of the control parameters, improved the adaptability of the control parameters, and effectively solved the problem of the difficulty in adjusting the control parameters of the hypersonic aircraft controller. This method has strong robustness, but because it relies more on the parameter setting of the PID controller and ignores the importance of prior knowledge such as the dynamic model, it limits the implementation of this method in complex flight missions outside the scope of the expert database to a certain extent, and it is difficult to grasp the improvement of control quality.
[0007] In summary, although there have been many research results on the control system design of hypersonic aircraft, the method that relies solely on model knowledge or experimental data cannot fully meet the design requirements of hypersonic deformable aircraft. Therefore, designing a controller that combines knowledge and data-driven to ensure the attitude stability of the aircraft system has become a key task.
[0008] At present, the control method driven by knowledge and data hybrid has made some progress in the field of high-speed flight control, but most of the research is completed under simplified conditions and assumptions, and there is still a gap with actual engineering applications. In order to further promote the engineering application of this method, the present invention proposes an intelligent control method for hypersonic deformable aircraft based on knowledge and data hybrid drive. This method fully combines the knowledge of dynamic models, offline flight simulation data and multi-source disturbance information processing, and uses data-driven parameter adjusters, compensators, model learners and controllers to enhance the robustness and adaptability of the control system through different mechanisms, and finally realizes online intelligent control design based on meta-learning. This method aims to promote the development of hypersonic deformable aircraft technology towards higher reliability and better stability. Summary of the invention
[0009] Aiming at the control problem of hypersonic deformable aircraft under complex interference, the present invention proposes a knowledge- and data-driven intelligent control technology. First, a knowledge-based flight dynamics model is established and converted into a model suitable for control. This step ensures the accuracy and operability of the model and lays a solid foundation for subsequent control design. Next, a benchmark control law based on neural network compensation is designed to provide a basic framework for offline data training and online intelligent control. This benchmark control law improves the robustness and adaptability of the system through the neural network compensation mechanism and can cope with various flight conditions. Furthermore, the impact of multi-source complex interference on aircraft control is comprehensively considered. The aerodynamic force and torque modeling error is regarded as model uncertainty interference, the additional torque caused by deformation is regarded as internal interference, and the impact of the external environment (such as wind, temperature change, airflow disturbance, etc.) is regarded as external interference. In the offline training stage, these offline data are used to carry out training under the benchmark control framework for different deformation modes, model uncertainty and multi-source disturbance effects to generate high-quality data samples. Finally, an online controller based on meta-learning is designed. On the basis of making full use of offline training data, a three-step online control design scheme is proposed: first, the online tuning of control parameters is realized through meta-learning, so that the control system can quickly adapt to changing flight conditions; second, real-time compensation of unknown interference is performed based on prior knowledge to improve the anti-interference ability of the system; finally, the online adaptive control of the aircraft is realized by using the meta-learning intelligent control method to ensure the stability and reliability of the aircraft in various complex environments. This method of combining knowledge and data can not only solve the control problem of hypersonic deformable aircraft under complex interference, but also significantly improve the intelligence level and engineering application value of the control system, and promote the development of hypersonic flight technology.
[0010] The technical solution of the present invention:
[0011] The method for intelligent control of a hypersonic shape-shifting aircraft driven by hybrid knowledge and data includes the following steps:
[0012] Step 1: Construction of knowledge-based aircraft dynamics model
[0013] (1.1) Kinetic model
[0014] Establishing the dynamic model of hypersonic deformable aircraft:
[0015] (1)
[0016] Where: α, β and They represent the angle of attack, sideslip angle and bank angle respectively, θ represents the flight path angle; ω x ,ω y and ω zrepresents the roll, yaw and pitch angular velocity respectively; L, D and N represent the lift, drag and lateral force respectively; V, m s and g are the flight speed, aircraft mass and gravitational acceleration; M x 、M y and M z Represent rolling, yaw and pitch moments respectively; M xs 、M ys and M sz I represents the additional deformation moment caused by the deformation of the three degrees of freedom: roll, yaw and pitch; x ,I y and I z represent the moments of inertia about the x, y, and z axes respectively.
[0017] Force and moment are expressed as
[0018] (2)
[0019] Where: q represents the dynamic pressure, κ is the deformation rate of the wingspan and sweep angle, S(κ) is the actual reference area when there is deformation, C L (κ), C D (κ), C N (κ) is the equivalent lift, lateral force and drag coefficient when there is deformation, m x (κ), m y (κ), m z (κ) is the equivalent rolling, yaw and pitch moment coefficients with deformation, L ref is the average geometric chord length.
[0020] (1.2) Control-oriented model transformation
[0021] Formula (1) is described as a control-oriented mathematical model:
[0022] (3)
[0023] In the formula: the state quantity is expressed as x ˙ 1 = [ α , β , γ V ] , x ˙ 2 = [ ω x , ω y , ω z ] , control amount u = [ δ x , δ y , δ z ] , Expressed as the rudder angles in pitch, roll and yaw. represents the known nonlinear function of the aircraft in the nominal case, represents the known aircraft rudder efficiency coefficient under nominal conditions, Represents uncertainties, including additional moments caused by deformation, errors in aerodynamic force and moment modeling, and the effects of external disturbances.
[0024] Step 2: Design of reference control law based on neural network compensation
[0025] Assume that the composite perturbation can be approximated by a constant-weight neural network with a sufficient number of neurons, namely: ,in represents the weight matrix, express dimensional space of size, where Indicates the dimension of the child element. represents the corresponding basis function matrix, express The dimensional space of size, represents the approximation error of the neural network, express More specifically,
[0026] (4)
[0027] In the formula, represents the weight vector, represents the corresponding basis function vector.
[0028] For the mathematical model oriented to control, the attitude angle tracking error is first defined: ,in is the corresponding attitude angle reference instruction.
[0029] Design the virtual control quantity as follows:
[0030] (5)
[0031] Where: k 1 = diag [ k 11 , k 12 , k 13 ] represents the positive parameter to be designed, The optimal weight The estimated value of Represents the derivative of a reference instruction.
[0032] The actual control quantity of the design is as follows:
[0033] (6)
[0034] Where: k 2 = diag [ k 21 , k 22 , k 23 ] represents the positive parameter to be designed, The optimal weight The estimated value of .
[0035] Given the update law for the neural network weights:
[0036] (7)
[0037] Where: i = 1,2, and All represent the positive definite parameter matrices to be designed. The optimal weight The estimated value of .
[0038] Step 3: Knowledge acquisition and network training based on offline data
[0039] The aerodynamic force and torque modeling error of formula (2) is regarded as the model uncertainty interference, the additional torque caused by deformation is regarded as internal interference, and the influence of the external environment is regarded as external interference. The offline part of the training considers the data under different deformation modes, different model uncertainties, and multi-source disturbances. Under the action of the benchmark control law formula (6) in step (2), the network weight update training of formula (7) is carried out as a flight test data sample. During the training process, the control command u and the neural network weights are established from the flight mission and flight state. The obtained neural network weights provide weight information for the intelligent online control scheme based on meta-learning in step (4), thereby generating online control instructions.
[0040] Step 4: Design of online intelligent control scheme based on meta-learning
[0041] The meta-learning based online controller design scheme includes three parts: online tuning of control parameters based on meta-learning, unknown disturbance compensation based on prior knowledge, and online application of meta-learning based intelligent control.
[0042] The online intelligent controller is designed as follows:
[0043] (8)
[0044] Where: k 1 = diag [ k 11 , k 12 , k 13 ] and k 2 = diag [ k 21 , k 22 , k 23 ] represents the control parameters to be designed, and Represents the neural network weights obtained through offline learning.
[0045] Based on formula (8), the design process of the online intelligent control scheme based on meta-learning is as follows:
[0046] (4.1) Online tuning of control parameters based on meta-learning: Online tuning of control parameters based on meta-learning aims to adjust controller parameters in real time through meta-learning algorithms to adapt to changes in the dynamic environment. The specific steps are as follows: First, a meta-model that can quickly adapt to environmental changes is established through historical data and prior knowledge. The meta-model is generated through the meta-learning process and is mainly in the form of a neural network. The input of the meta-model is historical data, current flight status, and control input, and the output is the main parameters of the online intelligent controller in formula (8): k 1 = diag [ k 11 , k 12 , k 13 ] and k 2 = diag [ k 21 , k 22 , k 23 ] . Among them, historical data includes previous flight data of the aircraft under different environments and conditions, controller response data, and the relationship between changes in environmental parameters and controller performance, such as angle of attack, sideslip angle, bank angle, flight path angle, roll, yaw and pitch angular velocity, lift, drag and lateral force, flight speed, roll, yaw and pitch moment. Collect and organize historical data and prior knowledge of the aircraft in different environments. These data reflect the stability and performance of the aircraft further maintained by adjusting parameters through data-driven controllers under various working conditions. The data needs to go through preprocessing steps such as cleaning, denoising and normalization to ensure its quality and consistency. Extract key features from the processed data, which should be able to reflect the relationship between aircraft performance and environmental parameters, such as speed, angle, acceleration, etc., as well as the influence of external disturbances. Divide the historical data into training set and test set, and adjust the model parameters in the training set so that it can accurately predict the optimal settings of control parameters. Verify the performance of the meta-model on the test set, and further optimize the model to ensure its robustness and stability in practical applications. The trained meta-model can continuously receive new data and update the model parameters in real time through online learning, so that it can continuously adapt to new environmental changes. Using the above historical data and prior knowledge, the controller parameters can be optimized online, providing strong support for the stable flight of hypersonic shape-shifting aircraft in complex environments. In this design, the controller parameters to be optimized are defined as ,The goal of meta-learning is to find a set of optimal controller parameters through an offline optimization process, so that the adaptive controller can perform well under different trajectory and disturbance conditions.
[0047] The optimal target for training parameters is:
[0048] (9)
[0049] in, Represents different nominal trajectories, each of which corresponds to an independent task. represents the trajectory data containing different disturbance information under each reference trajectory, represents the time interval of the nominal trajectory. i and j represent the dimension and number of columns of the aircraft data samples. represents the flight data 1 of the i-th dimension and j-th column, Indicates the corresponding reference value. represents the flight data 2 of the i-th dimension and j-th column, Indicates the corresponding reference value. represents the sum of squared errors of the information performance index of the data in the i-th dimension and j-th column, , , All are positive parameters to be designed. Design the optimal target of the above training parameters, and obtain the optimal control parameters after optimization , so that good control performance can be achieved under each training trajectory and disturbance. This method can not only ensure that the controller performs well on the training data samples, but also can be generalized to flight missions with unknown interference conditions.
[0050] (4.2) Compensation for unknown interference based on prior knowledge: Compensation for unknown interference based on prior knowledge aims to use known system characteristics and historical data to compensate for unknown interference in the system. Because hypersonic deformable aircraft will encounter a variety of complex interference factors during flight, such as airflow disturbances, structural vibrations, ambient temperature changes, etc., these interferences may be nonlinear, strongly coupled, and difficult to model in advance. Therefore, a neural network is used to compensate for complex interference, which is specifically expressed as the neural network form based on prior knowledge in equation (8): and , where the network weight and After offline training, stable and effective neural network weight information is extracted and stored as a priori knowledge base.
[0051] (4.3) Online application based on meta-learning intelligent control: Online application based on meta-learning intelligent control aims to use meta-learning algorithms to achieve intelligent control of complex systems. Equation (8) is an online control design scheme, where is the virtual control quantity, The actual control quantity. , , and is the nominal information of the hypersonic shape-shifting vehicle, and Represents the first-order derivative of the reference command and the virtual control quantity, combined with the unknown disturbance compensation based on prior knowledge, to obtain the neural network form based on prior knowledge and , further carry out online tuning of control parameters based on meta-learning, adjust controller parameters in real time through meta-learning algorithm to adapt to changes in dynamic environment, and complete the online deployment of intelligent control scheme.
[0052] Beneficial effects of the present invention:
[0053] The present invention fully coordinates the characteristics of knowledge-driven and data-driven strategies, making the advantages of the two methods complementary, showing superiority in processing control parameter setting, neural network compensation, and model uncertainty learning, and is an important idea for promoting the development of hypersonic deformable aircraft control technology. Among them, the knowledge-driven part covers the dynamic model of the hypersonic deformable aircraft, the aerodynamic information and control theory knowledge mastered before performing a specific flight mission; the data-driven part includes the original data obtained from various ground experiments, historical flight data, and data obtained through various real-time measurements and processed and used online during the execution of a specific flight mission. Integrating the knowledge of the dynamic model, training an online intelligent control scheme with learning ability through the training set avoids the limitations of the knowledge-driven methods commonly used in engineering and the data-driven methods in the field of artificial intelligence. In addition, the scheme designed by the present invention has a theoretical stability guarantee, fully considers the control difficulties of the hypersonic deformable aircraft, and can provide a knowledge-data hybrid driving strategy with stronger pertinence, better stability and higher reliability. This method is a high-quality intelligent control method driven by a hybrid of knowledge and data, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 It is an overall flow chart of the intelligent control method of a hypersonic shape-shifting aircraft;
[0055] Figure 2 It is a framework for intelligent control of hypersonic shape-shifting aircraft driven by knowledge and data hybrid;
[0056] Figure 3 is the curve of attitude angle variation of hypersonic deformable aircraft over time;
[0057] Figure 4 is the curve of attitude angular velocity of hypersonic deformable aircraft changing with time;
[0058] Figure 5 It is the curve of the rudder deflection command of the hypersonic deformable aircraft changing with time. DETAILED DESCRIPTION
[0059] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.
[0060] like Figure 1 and Figure 2As shown, the knowledge and data hybrid driving hypersonic deformation aircraft intelligent control method of the present invention is as follows:
[0061] Step 1: Construction of knowledge-based aircraft dynamics model
[0062] (1.1) Kinetic model
[0063] Establishing the dynamic model of hypersonic deformable aircraft:
[0064] (10)
[0065] Where: α, β and They represent the angle of attack, sideslip angle and bank angle respectively, θ represents the flight path angle; ω x ,ω y and ω z represents the roll, yaw and pitch angular velocity respectively; L, D and N represent the lift, drag and lateral force respectively; V, m s and g are the flight speed, aircraft mass and gravitational acceleration; M x 、M y and M z Represent rolling, yaw and pitch moments respectively; M xs 、M ys and M sz I represents the additional deformation moment caused by the deformation of the three degrees of freedom: roll, yaw and pitch; x ,I y and I z represent the moments of inertia about the x, y, and z axes respectively.
[0066] Force and moment are expressed as
[0067] (11)
[0068] Where: q represents the dynamic pressure, κ is the deformation rate of the wingspan and sweep angle, S(κ) is the actual reference area when there is deformation, C L (κ), C D (κ), C N (κ) is the equivalent lift, lateral force and drag coefficient when there is deformation, m x (κ), m y (κ), m z (κ) is the equivalent rolling, yaw and pitch moment coefficients with deformation, L ref is the average geometric chord length.
[0069] (1.2) Control-oriented model transformation
[0070] Formula (1) is described as a control-oriented mathematical model:
[0071] (12)
[0072] In the formula: the state quantity is expressed as x ˙ 1 = [ α , β , γ V ] , x ˙ 2 = [ ω x , ω y , ω z ] , control amount u = [ δ x , δ y , δ z ] , Expressed as the rudder angles in pitch, roll and yaw. represents the known nonlinear function of the aircraft in the nominal case, represents the known aircraft rudder efficiency coefficient under nominal conditions, Represents uncertainties, including additional moments caused by deformation, errors in aerodynamic force and moment modeling, and the effects of external disturbances.
[0073] Step 2: Design of reference control law based on neural network compensation
[0074] Assume that the composite perturbation can be approximated by a constant-weight neural network with a sufficient number of neurons, namely: ,in represents the optimal weight matrix, express dimensional space of size, where Indicates the dimension of the child element. represents the corresponding basis function matrix, express The dimensional space of size, represents the approximation error of the neural network, express More specifically,
[0075] (13)
[0076] In the formula, represents the optimal weight vector, represents the corresponding basis function vector.
[0077] For the mathematical model oriented to control, the attitude angle tracking error is first defined: ,in is the corresponding attitude angle reference command. Then the attitude angle tracking error dynamics can be expressed as:
[0078] (14)
[0079] For this purpose, the virtual control quantity is designed as follows:
[0080] (15)
[0081] Where: k 1 = diag [ k 11 , k 12 , k 13 ] represents the positive parameter to be designed, The optimal weight The estimated value of Represents the derivative of a reference instruction.
[0082] Define the attitude angular velocity tracking error: , then the attitude angular velocity tracking error dynamics can be expressed as:
[0083] (16)
[0084] In this regard, the actual control quantity is designed as follows:
[0085] (17)
[0086] Where: k 2 = diag [ k 21 , k 22 , k 23 ] represents the positive parameter matrix to be designed, The optimal weight The estimated value of .
[0087] Furthermore, the update law of the neural network weights is given:
[0088] (18)
[0089] Where: i = 1,2, and represents the positive definite parameter matrix to be designed, The optimal weight The estimated value of .
[0090] Define the weight matrix estimation error . Construct Lyapunov function, , and find its derivative
[0091] (19)
[0092] By choosing appropriate parameters , available
[0093] (20)
[0094] In the formula, the state quantity is expressed as , .
[0095] Further from the boundedness lemma, we can know the following inequality relationship:
[0096] (twenty one)
[0097] The above system is eventually consistent, bounded and stable and satisfies:
[0098] (twenty two)
[0099] It can be seen from the above formula that by adjusting the control parameters, the final bound of the attitude angle tracking error can be ensured to be less than , achieving high-performance tracking control.
[0100] Step 3: Knowledge acquisition and network training based on offline data
[0101] Realizing autonomous and precise control of aircraft under complex interference from multiple sources and endowing the aircraft with intelligent behavior capabilities such as strong autonomy, strong adaptability, and strong survival are the core technical issues that need to be solved in the development of hypersonic deformable aircraft. This scheme regards the aerodynamic force and torque modeling errors as model uncertainty interference, the additional torque caused by deformation as internal interference, and the influence of the external environment (such as wind, temperature changes, airflow disturbances, etc.) as external interference. The offline part of the training is used to consider offline data under different deformation modes, different model uncertainties, and multi-source disturbances as data samples for training under the benchmark control framework (benchmark control law formulas (6) and (7) in step (2)).
[0102] (3.1) Dataset generation: This step is to generate reliable and representative data that can cover the complex flight conditions of hypersonic vehicles, so as to train a neural network prior knowledge model with high accuracy and high generalization. First, it is necessary to collect and generate data covering multiple deformation modes, model uncertainties and multi-source disturbance effects. These data need to simulate the performance of actual aircraft in various complex environments as comprehensively as possible to ensure the adequacy and generalization ability of model training. Then, simulation data generation is performed to simulate various possible flight conditions and interference situations to generate a large amount of training data. Specifically: First, by changing the aerodynamic and torque model parameters, data containing model errors are generated to simulate model uncertainty. Second, by modeling the structural deformation of the aircraft, the corresponding torque additional data is generated to simulate the deformation mode. Third, by simulating different environmental conditions (such as wind speed, airflow changes, etc.), external interference data is generated to simulate environmental interference. Finally, the data set is divided into training set, validation set and test set with an allocation ratio of 8:1:1 for subsequent design verification.
[0103] (3.2) Training and verification: After the flight test data set is generated, this data is used for training under the benchmark control framework. During the training process, the model parameters and control strategies are continuously adjusted to enable it to effectively cope with various model uncertainties and multi-source disturbances. Through this design, the unknown dynamics in the neural adaptive control process can be converted into weight knowledge, thereby reflecting the true learning ability of the neural network. The above work acquires and stores empirical knowledge in the form of constant neural networks, and then uses the stored neural network weight information to design an online controller based on empirical knowledge to obtain more excellent control performance. While training, verification work is carried out to evaluate the performance and stability of the model to ensure its reliability and robustness when dealing with complex interference.
[0104] (3.3) Testing and application: After sufficient training and verification, the model is applied to the control system of a hypersonic deformable aircraft. The cross-validation method is used to evaluate the performance of the model on different data sets to prevent overfitting. The robustness of the model is tested by simulating interference conditions in complex environments, and the adaptability of the model and the actual control effect are further verified. According to the test results, the model is adjusted and optimized as necessary, and then the model is continuously optimized in the offline stage to gradually improve the performance and stability of the control system, ensuring that it can achieve high-precision and high-reliability autonomous control in actual applications.
[0105] Step 4: Design of online intelligent control scheme based on meta-learning
[0106] This section presents an online controller design scheme based on meta-learning, which includes three parts: online tuning of control parameters based on meta-learning, compensation for unknown disturbances based on prior knowledge, and online application of intelligent control based on meta-learning.
[0107] The online intelligent controller is designed as follows:
[0108] (twenty three)
[0109] Where: k 1 = diag [ k 11 , k 12 , k 13 ] and k 2 = diag [ k 21 , k 22 , k 23 ] represents the control parameters to be designed, and Represents the neural network weights obtained through offline learning.
[0110] Based on formula (23), the design process of the online intelligent control scheme based on meta-learning is as follows:
[0111] (4.1) Meta-learning-based online tuning of control parameters: Meta-learning-based online tuning of control parameters aims to adjust controller parameters in real time through meta-learning algorithms to adapt to changes in dynamic environments. The specific steps are as follows: First, a meta-model that can quickly adapt to environmental changes is established through historical data and prior knowledge. The historical data includes the flight data of the aircraft in previous environments and conditions, controller response data, and the relationship between environmental parameter changes and controller performance, such as angle of attack, sideslip angle, bank angle, flight path angle, roll, yaw and pitch angular velocity, lift, drag and lateral force, flight speed, roll, yaw and pitch moment. Collect and organize the historical data and prior knowledge of the aircraft in different environments. These data reflect the stability and performance of the aircraft under various working conditions by adjusting the controller parameters through data-driven control. The data needs to undergo pre-processing steps such as cleaning, denoising and normalization to ensure its quality and consistency. Extract key features from the processed data. These features should be able to reflect the relationship between the performance of the aircraft and environmental parameters, such as speed, angle, acceleration, etc., as well as the influence of external disturbances. The historical data is divided into a training set and a test set. The model parameters are adjusted in the training set so that it can accurately predict the optimal settings of the control parameters. The performance of the metamodel is verified on the test set, and the model is further optimized to ensure its robustness and stability in practical applications. The trained metamodel can continuously receive new data and update the model parameters in real time through online learning, so that it can continuously adapt to new environmental changes. Using the above historical data and prior knowledge, the controller parameters are optimized online, providing strong support for the stable flight of hypersonic deformable aircraft in complex environments. In this design, the controller parameters to be optimized are defined. ,The goal of meta-learning is to find a set of optimal controller parameters through an offline optimization process, so that the adaptive controller can perform well under different trajectory and disturbance conditions.
[0112] The optimal target for training parameters is:
[0113] (twenty four)
[0114] in, Represents different nominal trajectories, each of which corresponds to an independent task. represents the trajectory data containing different disturbance information under each reference trajectory, represents the time interval of the nominal trajectory. i and j represent the dimension and number of columns of the aircraft data samples. represents the flight data 1 of the i-th dimension and j-th column, Indicates the corresponding reference value. represents the flight data 2 of the i-th dimension and j-th column, Indicates the corresponding reference value. represents the sum of squared errors of the information performance index of the data in the i-th dimension and j-th column, , , All are positive parameters to be designed. Design the optimal target of the above training parameters, and obtain the optimal control parameters after optimization , so that good control performance can be achieved under each training trajectory and disturbance. This method can not only ensure that the controller performs well on the training data samples, but also can be generalized to flight missions with unknown interference conditions.
[0115] (4.2) Compensation for unknown disturbances based on prior knowledge: Compensation for unknown disturbances based on prior knowledge aims to use known system characteristics and historical data to compensate for unknown disturbances that occur in the system. Because hypersonic deformable aircraft will encounter a variety of complex disturbances during flight, such as airflow disturbances, structural vibrations, and ambient temperature changes, these disturbances may be nonlinear, strongly coupled, and difficult to model in advance. The specific steps are as follows: First, the hypersonic deformable aircraft is equipped with multiple sensors to monitor the changes of key parameters during flight in real time, such as flight speed, attitude angle, angular velocity, and external environmental parameters such as airflow and temperature. Secondly, by analyzing these response data under different interference conditions, the neural network model is trained offline to extract stable and effective neural network weight information and store it as a prior knowledge base. The knowledge base contains known interference patterns (such as certain airflow disturbances) and corresponding controller gains. Finally, when a certain interference pattern is identified by online monitoring, the system will extract the compensation strategy corresponding to the interference pattern from the prior knowledge base. The compensation strategy is based on the neural network prior weight information obtained by offline training to calculate the control adjustment scheme that adapts to the current environment. At the same time, the new data collected during the flight can further optimize the prior knowledge base and improve the future compensation effect. In the control of hypersonic deformable aircraft, this interference compensation method based on prior knowledge can not only significantly improve the robustness of the aircraft, but also reduce the sensitivity to changes in the external environment, thereby enhancing the adaptability of the aircraft in complex and changing environments.
[0116] (4.3) Online application of intelligent control based on meta-learning: Online application of intelligent control based on meta-learning aims to realize intelligent control of complex systems by using meta-learning algorithms. The specific steps are as follows: First, design an intelligent controller based on meta-learning, which can automatically adjust the control strategy according to environmental changes and system status. Then, use the meta-learning algorithm to train the intelligent controller so that it can make optimal control decisions under different operating conditions. Finally, in practical applications, the intelligent controller monitors the system status in real time and adjusts the control strategy according to the output of the meta-learning algorithm to realize intelligent management of the control system. First, develop an intelligent control scheme for hypersonic deformable aircraft based on meta-learning. The core of this control scheme lies in its adaptive ability, that is, it can automatically adjust the control strategy according to the current state of the aircraft (such as speed, attitude, angular velocity, etc.) and environmental changes (such as airflow, temperature, etc.). This control scheme adopts a meta-learning algorithm, which enables it to quickly learn from limited data to cope with unknown or changing flight environments. After the baseline controller design is completed, it needs to be trained. The purpose of the meta-learning algorithm is to enable the intelligent control scheme to make optimal control decisions under a variety of operating conditions. The training process can use simulation environments or historical data to simulate various possible flight conditions and interference situations. Through training, the controller can learn the best flight control strategy in different environments and quickly adapt to new conditions in actual applications. In actual applications, the intelligent control scheme monitors the system status in real time through sensors, including the aircraft's attitude, speed, environmental changes, etc. When the environment changes or the system status deviates from expectations, the controller dynamically adjusts the control strategy based on the output of the meta-learning algorithm. For example, if the aircraft encounters sudden airflow, the controller can immediately adjust the rudder angle, thrust, and deformation rate to ensure flight stability. In addition, during the flight, the controller not only performs control tasks, but also continuously collects new data for online learning and adjustment. In this way, the controller can gradually optimize its control strategy and improve the overall performance of the system. For hypersonic deformable aircraft, this meta-learning-based intelligent controller can significantly improve the aircraft's adaptability in complex environments, ensuring that the aircraft can maintain highly stable and precise control when facing various interferences and changes. At the same time, the application of the meta-learning algorithm enables the controller to respond quickly with limited data, avoiding the delays and inaccuracies that may exist in traditional methods.
[0117] The above-mentioned meta-learning-based online controller design solution includes hardware configuration, software environment, data acquisition and preprocessing, control algorithm deployment, system integration and testing, and maintenance and update. In terms of hardware configuration, sufficient computing resources, sensors, actuators, and communication equipment are required; in terms of software environment, a suitable operating system and programming environment are selected, and necessary libraries and dependencies are installed. In the process of data acquisition and preprocessing, data is collected and cleaned in real time, and feature extraction is performed. The control algorithm optimizes the model through offline training and online training, and deploys the model in the control system to ensure its efficient operation. System integration includes software and hardware integration and data communication configuration, and system testing covers functions, performance, and security. In the process of algorithm implementation, the meta-learning algorithm needs to be encapsulated so that it can run efficiently in embedded systems. Programming languages such as C / C++ and Python can be used to develop intelligent control algorithms, and hardware acceleration (such as GPU and FPGA) can be used to improve computing efficiency.
[0118] In order to verify the intelligent control method of hypersonic deformable aircraft based on knowledge and data hybrid drive, this simulation experiment scheme includes the following steps. First, the aircraft dynamics model is established, and the parameters include wingspan of 2.5 meters, wing area of 1.5 square meters, average aerodynamic chord length of 0.5 meters and fuselage length of 5 meters. Aerodynamic parameters such as lift coefficient, drag coefficient and moment coefficient are set based on angle of attack and Mach number, with a mass of 500 kg and moments of inertia of 200, 180 and 220 kg square meters respectively. Subsequently, typical flight missions such as horizontal flight, climbing, turning and descending are simulated to generate trajectory data of hypersonic aircraft: 1) Horizontal flight, the flight altitude is set to 10,000 meters, and the flight speed is maintained at Mach 6. 2) Climbing mission, the aircraft climbs from 8,000 meters to 12,000 meters, and the speed is maintained at Mach 5 during the climb. 3) Turning mission, the aircraft turns at an angle of 45 degrees, the flight speed is Mach 4, and the flight altitude during the turn is 9,000 meters. 3) Descent mission: The aircraft descends from an altitude of 12,000 meters to 6,000 meters, and the speed is maintained at 5.5 Mach during the descent. The wind speed varies randomly between 0 and 30 meters per second, and Gaussian noise is used to simulate airflow disturbances. The noise intensity varies randomly between 0 and 10. The simulation time step is set to 0.01 seconds, the total time is 1000 seconds, and the position, speed, attitude angle, and control input data are recorded. In the offline training phase, the data is normalized, 80% is used for training, and 20% is used for verification. The benchmark control law uses a 3-layer 64-neuron neural network. The training algorithm uses the Adam optimizer, the learning rate is 0.001, and the number of training rounds is 1000.
[0119] In the online control design stage, a meta-learning algorithm is used, the initial control parameters are learned through historical data, and online tuning is performed using real-time data. Figures 3 to 5It is the attitude control result of a set of optimal control parameters after online tuning for the hypersonic aircraft. As can be seen from the figure, after the initial adjustment, the attitude angle and angular velocity errors of the aircraft gradually stabilize, indicating that the control system can effectively stabilize the attitude of the aircraft under the optimal parameters.
Claims
1. A knowledge and data hybrid driven hypersonic shape-shifting aircraft intelligent control method, characterized in that: The steps include: Step 1: Construction of knowledge-based aircraft dynamics model (1.1) Kinetic model Establishing the dynamic model of hypersonic deformable aircraft: Where: α, β and γ V They represent the angle of attack, sideslip angle and bank angle respectively, θ represents the flight path angle; ω x ,ω y and ω z They represent roll, yaw and pitch angular velocity respectively; L and N represent lift and lateral force respectively; V, m s and g are the flight speed, aircraft mass and gravitational acceleration; M x 、M y and M z Represent rolling, yaw and pitch moments respectively; M sx 、M sy and M sz They represent the additional deformation moments caused by the deformed shapes of the three degrees of freedom: roll, yaw and pitch; I x ,I y and I z denote the moments of inertia about the x, y, and z axes respectively; Force and moment are expressed as Where: D is the drag, q is the dynamic pressure, κ is the deformation rate of the wingspan and sweep angle, S(κ) is the actual reference area when there is deformation, C L (κ), C D (κ), C N (κ) is the equivalent lift, drag and lateral force coefficient with deformation, m x (κ), m y (κ), m z (κ) is the equivalent rolling, yaw and pitch moment coefficients with deformation, L ref is the average geometric chord length; (1.2) Control-oriented model transformation Formula (1) is described as a control-oriented mathematical model: In the formula: the state quantity is expressed as Control quantity u=[δ x ,δ y ,δ z ], δ x ,δ y ,δ z Expressed as the rudder angle in pitch, roll and yaw directions; f i represents the known nonlinear function of the aircraft under nominal conditions, g i represents the known aircraft rudder efficiency coefficient under nominal conditions, Δ i represents the uncertainty, including the additional moment caused by deformation, aerodynamic force and moment modeling errors, and external disturbance effects, where i = 1, 2; Step 2: Design of reference control law based on neural network compensation Assume that the composite perturbation can be approximated by a constant-weight neural network with a sufficient number of neurons, namely: in represents the weight matrix, Represents 3Σ i ×3-dimensional space, where Σ i Indicates the dimension of the sub-element; represents the corresponding basis function matrix, Represents 3Σ i ×1 dimensional space, represents the approximation error of the neural network, represents a 3×1 dimensional space; more specifically, In the formula, represents the weight vector, represents the corresponding basis function vector; For the mathematical model oriented to control, we first define the attitude angle tracking error: e1 = x1-x d , where x d is the corresponding attitude angle reference instruction; Design the virtual control quantity as follows: Where: k1 = diag[k 11 ,k 12 ,k 13 ] represents the positive parameter to be designed, is the estimated value of the optimal weight ω1, represents the derivative of the reference instruction; The actual control quantity of the design is as follows: Where: k2 = diag[k 21 ,k 22 ,k 23 ] represents the positive parameter to be designed, is the estimated value of the optimal weight ω2; Given the update law for the neural network weights: Where: i = 1, 2, and All represent the positive definite parameter matrices to be designed. is the optimal weight ω i An estimated value of Step 3: Knowledge acquisition and network training based on offline data The aerodynamic force and torque modeling error of formula (2) is regarded as the model uncertainty interference, the additional torque caused by deformation is regarded as internal interference, and the influence of the external environment is regarded as external interference; The offline training considers data under different deformation modes, different model uncertainties, and multi-source disturbances. Under the action of the benchmark control law formula (6) in step (2), the network weight update training of formula (7) is carried out as a flight test data sample. During the training process, the control command u and the neural network weights are established from the flight mission and flight state. The mapping relationship, i = 1, 2, the obtained neural network weights provide weight information for the intelligent online control scheme based on meta-learning in step (4), and then generate online control instructions; Step 4: Design of online intelligent control scheme based on meta-learning The online intelligent controller is designed as follows: Where: k1 = diag[k 11 ,k 12 ,k 13 ] and k2=diag[k 21 ,k 22 ,k 23 ] represents the control parameters to be designed, and Represents the weights of the neural network obtained through offline learning; Based on formula (8), the design process of the online intelligent control scheme based on meta-learning is as follows: (4.1) Online tuning of control parameters based on meta-learning: First, a meta-model that can quickly adapt to environmental changes is established through historical data and prior knowledge. The meta-model is generated through the meta-learning process and is expressed in the form of a neural network. The input of the meta-model is historical data, current flight status, and control input, and the output is the main parameter k1=diag[k 11 ,k 12 ,k 13 ] and k2=diag[k 21 ,k 22 ,k 23 ]; where historical data includes previous flight data of the aircraft under different environments and conditions, controller response data, and the relationship between environmental parameter changes and controller performance, including angle of attack, sideslip angle, bank angle, flight track angle, roll, yaw and pitch angular velocity, lift, drag and lateral force, flight speed, roll, yaw and pitch moment; collect and organize historical data and prior knowledge of the aircraft under different environments, and adjust the parameters of the controller through data-driven control to further maintain the stability and performance of the aircraft; use the above historical data and prior knowledge to achieve online optimization of controller parameters; define the controller parameters to be optimized K = (k1, k2), the goal of meta-learning is to find a set of optimal controller parameters through an offline optimization process, so that the adaptive controller can perform well under different trajectory and disturbance conditions; The optimal target for training parameters is: in, Represents different nominal trajectories, each of which corresponds to an independent task. represents the trajectory data containing different disturbance information under each reference trajectory, represents the time interval of the nominal trajectory; i and j represent the dimension and number of columns of the aircraft data sample; x 1ij represents the flight data 1 of the i-th dimension and j-th column, x 1dij Indicates the corresponding reference value; x 2ij represents the flight data 2 of the i-th dimension and j-th column, x 2dij Indicates the corresponding reference value; represents the error sum of squares of the data information performance index of the i-th dimension and j-th column, μ mean , μ ctr1 , μ ctr2 are all positive parameters to be designed; design the optimal target of the above training parameters, and obtain the optimal control parameter K after optimization; (4.2) Compensation for unknown interference based on prior knowledge: The neural network is used to compensate for complex interference, which is specifically expressed as the neural network based on prior knowledge in equation (8): and The network weight and After offline training, stable and effective neural network weight information is extracted and stored as a priori knowledge base; (4.3) Online application based on meta-learning intelligent control: Equation (8) is the online control design scheme, where is the virtual control quantity, is the actual control quantity; g1, g2, f1 and f2 are the nominal information of the hypersonic deformation vehicle, and Represents the first-order derivative of the reference command and the virtual control quantity, combined with the unknown disturbance compensation based on prior knowledge, to obtain the neural network form based on prior knowledge and Further carry out online tuning of control parameters based on meta-learning, adjust controller parameters in real time through meta-learning algorithm to adapt to changes in dynamic environment, and complete the online deployment of intelligent control solutions.
