A method and system for robust attitude control of a morphing aircraft

Abstract

The present application relates to a kind of morphing aircraft robust attitude control method and control system, the control method includes: S1: offline training aerodynamic inverse mapping reference neural network model;S2: construct online transfer learning sample and store;S3: based on the neural network transfer learning strategy of low rank self-adaption obtains adaptive aerodynamic inverse mapping neural network model;S4: construct hierarchical robust adaptive control architecture, decompose attitude control task, determine composite disturbance;S5: construct extended state observer, and composite disturbance is estimated on-line;S6: using composite disturbance estimated value carries out feedforward compensation;S7: using adaptive aerodynamic inverse mapping neural network model exports rudder deflection instruction.The control system uses the control method to carry out attitude control to morphing aircraft.The present application can improve the robustness of attitude control, and also can reduce the calculation complexity of controller and the calculation burden of on-board online learning, improve online adaptation efficiency and control security.

Description

Technical Field

[0001] This invention relates to a robust attitude control method and control system for a variable-configuration aircraft, belonging to the field of aircraft control technology. Background Technology

[0002] Modular aircraft can alter their aerodynamic shape during flight to meet mission requirements and adapt to the flight environment, thus achieving performance across a wide speed range, across multiple airspaces, and for multi-mission missions. For high-speed or hypersonic modular aircraft, configuration changes cause significant variations in the center of mass position, moment of inertia, aerodynamic parameters, and additional torques, resulting in highly nonlinear attitude dynamics and a wide range of parameter uncertainties. Achieving stable, safe, and robust attitude control under the combined effects of wide flight envelope, drastic configuration changes, and external disturbances is a key challenge in the design of modular aircraft control systems.

[0003] To address the highly nonlinear and wide-ranging parameter uncertainties inherent in the aerodynamic models of variable-structure aircraft, data-driven methods such as neural networks are employed for aerodynamic modeling, moment estimation, disturbance observation, and rudder deflection calculation in order to reduce reliance on precise aerodynamic models. However, in actual cross-domain flights, facing dramatic structural changes, aerodynamic pull, and sudden external airflow disturbances, relying solely on neural networks for learning can easily lead to "identification residuals" in the neural network aerodynamic model due to parameter update lags or incomplete coverage of boundaries by training samples. If these identification residuals are not eliminated in a timely manner, they will directly accumulate in the control loop, seriously threatening flight safety.

[0004] Meanwhile, under actual flight conditions, the online application of neural networks faces two key challenges. Firstly, limited by the onboard computer's computing power, traditional deep networks often cannot perform real-time online full-parameter updates. Secondly, during long-duration steady-state flight, the system lacks persistent excitation (PE), leading to highly homogenized online experience data. This makes the network prone to local overfitting and catastrophic forgetting of crucial historical information. If the aircraft executes sudden large maneuvers or step commands, the network may become unstable due to insufficient adaptability, thus threatening flight safety.

[0005] Furthermore, in attitude controller design, traditional backstepping control is often used to handle the strong nonlinear characteristics of the system. However, in multivariable design, backstepping requires repeated analytical differentiation of the virtual control variables of the previous stage. As the number of control loop layers increases, the system states become increasingly coupled, leading to a geometrical increase in the derivative terms, resulting in a severe "computational explosion" problem. This greatly consumes airborne computing resources and makes it difficult to meet the real-time requirements of high-speed flight.

[0006] Therefore, designing an efficient control architecture that can avoid backstepping calculation explosion, actively estimate and eliminate neural network aerodynamic model identification residuals and multi-source interference, and simultaneously complete neural network online migration is a technical challenge that urgently needs to be solved in the field of variable configuration aircraft control. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, this invention provides a robust attitude control method and control system for a variable-configuration aircraft, which can improve the robustness of attitude control, reduce the computational complexity of the controller and the computational burden of onboard online learning, and improve online adaptation efficiency and control safety.

[0008] The technical solution adopted in this invention is: a robust attitude control method for a variable-structure aircraft, comprising the following steps: S1: Collect wind tunnel test data, numerical simulation data and / or historical flight data of the modified aircraft, and train the aerodynamic inverse mapping benchmark neural network model offline based on the collected data; S2: During the online flight of the variable-structure aircraft, online transfer learning samples are constructed based on the error-driven adaptive experience replay strategy and real-time flight data, and stored in the experience replay pool for online training of the aerodynamic inverse mapping benchmark neural network model. S3: Freeze the weights of the aerodynamic inverse mapping baseline neural network model, introduce a low-rank adaptive parameter matrix into the network layer, and combine it with the error-driven adaptive empirical replay strategy to update only the low-rank parameters to obtain the adaptive aerodynamic inverse mapping neural network model. S4: Construct a hierarchical robust adaptive control architecture, decompose the attitude control task into top-level virtual torque generation and bottom-level aerodynamic inverse mapping calculation, and merge the bottom-level inverse mapping error, external disturbance, parameter uncertainty and unmodeled dynamics into a composite disturbance. S5: Extend the composite disturbance into a new state variable, construct an extended state observer, and perform online estimation of the composite disturbance; S6: Dynamic surface control is used to generate the desired virtual control torque, and feedforward compensation is performed using the composite disturbance estimate output by the extended state observer. S7: Convert the top-level desired virtual control torque into desired aerodynamic torque coefficient, input the current flight state, variant parameters and desired aerodynamic torque coefficient into the adaptive aerodynamic inverse mapping neural network model, and output three-channel rudder deflection commands to control the attitude of the variant aircraft.

[0009] Preferably, in step S1, an ideal nonlinear aerodynamic inverse mapping model and a loss function are defined. The nominal optimal weight of the offline-trained nonlinear aerodynamic inverse mapping model is obtained by minimizing the loss function and recorded as the benchmark weight, thus obtaining the aerodynamic inverse mapping benchmark neural network model.

[0010] Preferably, in S2, a sample admission criterion is constructed in the spatial dimension. When the state space distance between a new sample and the most recently pushed sample is less than a preset dead zone threshold, the information increment of the new sample is considered to be limited and it is not pushed onto the stack; otherwise, the new sample is pushed onto the stack.

[0011] Furthermore, a relative age is maintained for each sample added to the stack in the time dimension, and the age of all existing samples increases each time a new sample is allowed to be added to the stack. Define a dynamic forgetting sharpness factor driven by the current network prediction error. Calculate the probability of each sample being eliminated based on the dynamic forgetting sharpness factor, the relative age of the samples pushed into the stack, and the maximum age of the samples in the experience pool. When the experience pool reaches its maximum capacity limit, replace the eliminated samples with new samples according to the elimination probability. The age of the new samples is initialized.

[0012] Furthermore, when the network error is small, the experience pool is updated gradually to retain historically representative samples; when the network error suddenly increases, the forgetting sharpness rises rapidly, and the experience pool actively eliminates old configuration samples to make room for new working condition samples.

[0013] Preferably, in S3, for any fully connected layer in the aerodynamic inverse mapping baseline neural network model, its actual equivalent weight is represented as the sum of the ideal fixed weight and the low-rank correction term. The low-rank correction term is obtained by multiplying the dimension reduction matrix and the dimension increase matrix. During the online learning process, the offline fixed weight (backbone weight) is kept frozen, and only the dimension reduction matrix and the dimension increase matrix are updated, so that the number of parameters updated online in a single layer is reduced from the original full matrix scale to the low-rank scale, and finally the nonlinear aerodynamic inverse mapping neural network model is obtained.

[0014] Furthermore, during the initialization phase, the LoRA branch uses random initialization of the dimensionality reduction matrix and zero initialization of the dimensionality increase matrix to ensure that the low-rank correction term is zero at the start of online learning. Subsequently, the gradient of the offline solidified weights (backbone weights) is frozen, and the loss function is replayed based on experience. Only the gradient update with the correction term is performed on the low-rank matrix, which makes the number of parameters for single-layer online update drop sharply. Thus, without destroying the offline benchmark features, it can achieve fast online adaptation to local start-up drift.

[0015] Preferably, in step S4, a control-oriented dynamic model is established. Under the hierarchical decoupling architecture, the mapping deviation between the actual aerodynamic torque and the top-level expected virtual torque is defined as the network approximation residual of the bottom-level inverse calculation. This residual is then combined with external disturbances and model uncertainties to form a composite disturbance. The composite disturbance is substituted into the control-oriented dynamic model to obtain the equivalent control object faced by the top-level controller. This allows the bottom-level inverse mapping error to be uniformly absorbed into the top-level disturbance channel, providing a clear design interface for the upper-level robust controller.

[0016] Preferably, in step S5, an extended state system is constructed based on the extended state vector, the rate of change of the disturbance, and the top-level expected virtual torque. Then, an extended state observer is designed based on the extended state system, the observation error is defined, and the observer gain matrix is ​​constrained so that the observer error can converge exponentially to an adjustable small neighborhood near zero, thereby obtaining an accurate estimate of the composite disturbance.

[0017] Preferably, in step S6, the top-level controller adopts a dynamic surface control structure based on backstepping. First, the attitude angle tracking error is defined and combined with the attitude kinematics equations, based on... Lyapunov The kinematic desired virtual control law is designed based on stability theory. A first-order low-pass filter is introduced during the recursive process to smooth the desired virtual angular velocity. Then, the angular velocity tracking error is defined and combined with the dynamic equation containing composite disturbances. The composite disturbance estimate output by the extended state observer is used for feedforward compensation to obtain the final desired virtual control torque at the top level.

[0018] Preferably, in S7, the flight state includes angle of attack and sideslip angle, and the structural parameters include folding angle.

[0019] A robust attitude control system for a variable-configuration aircraft, employing any of the robust attitude control methods for variable-configuration aircraft disclosed in this invention to control the attitude of the variable-configuration aircraft, including: The benchmark model building module is used to collect wind tunnel test data, numerical simulation data and / or historical flight data of the modified aircraft, and to obtain the aerodynamic inverse mapping benchmark neural network model offline based on the collected data; Learning Sample Construction Module: Used to construct online transfer learning samples based on error-driven adaptive experience replay strategy and real-time flight data during the online flight of the modified aircraft, and store them in the experience replay pool; Adaptive model building module: used to freeze the weights of the aerodynamic inverse mapping baseline neural network model, introduce a low-rank adaptive parameter matrix into the network layer, and combine an error-driven adaptive empirical replay strategy to update only the low-rank parameters to obtain an adaptive aerodynamic inverse mapping neural network model. Attitude control task decomposition module: used to build a hierarchical robust adaptive control architecture, decompose the attitude control task into top-level virtual torque generation and bottom-level aerodynamic inverse mapping calculation, and merge the bottom-level inverse mapping error, external disturbance, parameter uncertainty and unmodeled dynamics into a composite disturbance. Composite disturbance estimation module: used to extend composite disturbances into new state variables, construct extended state observers, and perform online estimation of composite disturbances; Feedforward compensation module: used to generate the desired virtual control torque using dynamic surface control, and to perform feedforward compensation using the composite disturbance estimate output by the extended state observer; Control command generation module: It is used to convert the top-level desired virtual control torque into desired aerodynamic torque coefficient, input the current flight state, variable parameters and desired aerodynamic torque coefficient into the adaptive aerodynamic inverse mapping neural network model, and output three-channel rudder deflection commands.

[0020] The beneficial effects of this invention are: (1) This invention uses an extended state observer to uniformly estimate the composite disturbances formed by the bottom inverse mapping residual, external disturbances, parameter uncertainties and unmodeled dynamics, and performs feedforward compensation in the dynamic surface control law, which effectively improves the robustness of the system under aerodynamic pull, sudden disturbances and drastic structural changes. (2) The present invention introduces a low-pass filter through dynamic surface control to avoid repeated differentiation in the traditional backstepping method, which effectively reduces the computational complexity of the controller and improves the real-time performance of the project. (3) This invention effectively reduces the computational burden of airborne online learning by updating only a small number of low-rank parameters through low-rank adaptive online transfer learning; (4) This invention reduces the occupation of steady-state redundant samples in the experience pool by using an error-driven adaptive experience replay strategy, and quickly absorbs high-value samples during sudden instructions or drastic changes, which effectively improves online adaptation efficiency and control security. Attached Figure Description

[0021] Figure 1 This is a flowchart of the robust attitude control method for a variable-structure aircraft of the present invention; Figure 2 This invention provides a hierarchical robust attitude control architecture for the robust attitude control method of a variable-structure aircraft. Figure 3 This is a schematic diagram of the neural network online transfer learning mechanism of the robust attitude control method for variable-structure aircraft of the present invention; Figure 4 This is a schematic diagram of the offline neural network solution results in an embodiment of the present invention; Figure 5 This is a comparison chart of online transfer learning control results in embodiments of the present invention; Figure 6 This is a comparison chart of robust control results in embodiments of the present invention. Detailed Implementation

[0022] See Figures 1-3 This invention discloses a robust attitude control method for a variable-structure aircraft, comprising the following steps: S1: Collect wind tunnel test data, numerical simulation data and / or historical flight data of the modified aircraft, and train the aerodynamic inverse mapping benchmark neural network model offline based on the collected data; S2: During the online flight of the variable-structure aircraft, online transfer learning samples are constructed based on the error-driven adaptive experience replay strategy and real-time flight data, and stored in the experience replay pool for online training of the aerodynamic inverse mapping benchmark neural network model. S3: Freeze the weights of the aerodynamic inverse mapping baseline neural network model, introduce a low-rank adaptive parameter matrix into the network layer, and combine it with the error-driven adaptive empirical replay strategy to update only the low-rank parameters to obtain the adaptive aerodynamic inverse mapping neural network model. S4: Construct a hierarchical robust adaptive control architecture, decompose the attitude control task into top-level virtual torque generation and bottom-level aerodynamic inverse mapping calculation, and merge the bottom-level inverse mapping error, external disturbance, parameter uncertainty and unmodeled dynamics into a composite disturbance. S5: Extend the composite disturbance into a new state variable, construct an extended state observer, and perform online estimation of the composite disturbance; S6: Dynamic surface control is used to generate the desired virtual control torque, and feedforward compensation is performed using the composite disturbance estimate output by the extended state observer. S7: Based on the definition of aerodynamic moment, the top-level desired virtual control moment is converted into the desired aerodynamic moment coefficient. The current flight state, variant parameters and desired aerodynamic moment coefficient are input into the adaptive aerodynamic inverse mapping neural network model, and the three-channel rudder deflection command is output to control the attitude of the variant aircraft.

[0023] The specific steps of S1 preferably include: The ideal nonlinear aerodynamic inverse mapping model is defined as follows: ; in, The ideal network weight matrix; It is a vector of nonlinear functions consisting of the activation functions of the hidden layers; The minimum approximation residual inherent in the ideal network; The input parameters for the neural network include the required aerodynamic torque coefficient and the current state.

[0024] The loss function is defined as follows: ; in, This represents the number of offline training samples. This represents the actual rudder deflection in the sample; These are the current weight parameters of the neural network. The nominal optimal weights of the offline-trained nonlinear aerodynamic inverse mapping model are obtained by minimizing the loss function. And these are recorded as the baseline weights, finally yielding the aerodynamic inverse mapping baseline neural network model as follows: .

[0025] The specific steps of S2 preferably include: First, in the spatial dimension, based on the new sample Compared with the most recent push sample State space distance between Constructing sample admission criteria: , ; in, The preset dead zone threshold, This is a flag indicating whether to push the data onto the stack.

[0026] When new samples Compared with the most recent push sample The state space distance between them is less than the preset dead zone threshold. If the information increment of a new sample is considered limited, it will not be pushed onto the stack, thus suppressing the intrusion of steady-state redundant data on the diversity of the experience pool; otherwise, the new sample will be pushed onto the stack. Compared with the most recent push sample The state space distance between them is greater than or equal to a preset dead zone threshold. At that time, new samples are pushed onto the stack.

[0027] Secondly, in the time dimension, a relative age is maintained for each input sample. Each time a new sample is allowed to be pushed onto the stack, the age of all existing samples increases by 1.

[0028] Defined by the current network prediction error Driven dynamic forgetting sharpness factor : ; in, Based on forgetting sharpness, it represents the smooth elimination benchmark under steady state; The error-driven gain limits the upper limit of the accelerated elimination rate; This is an error sensitivity adjustment factor that controls the severity of the sharpness response to the residual.

[0029] Calculate the probability of each sample being eliminated. : ; in, This represents the maximum age of the samples in the current experience pool.

[0030] When the experience pool reaches its maximum capacity, the eliminated samples are replaced with new data (new samples) based on the elimination probability, and the age of the new data is initialized to 0. When the network error is small, the experience pool is updated in a relatively gradual manner to retain historically representative samples; when the network error suddenly increases, the forgetting sharpness rises rapidly, and the experience pool actively eliminates old configuration samples to make room for new operating condition samples, thereby accelerating the online adaptation of network aerodynamic mutations. This strategy maintains the state space dispersion and information abundance of the experience pool under conditions of insufficient continuous stimulation, providing effective sample support for the continuous online learning of the underlying neural network.

[0031] The specific steps of S3 preferably include: For any fully connected layer (i.e., for all connected layers) in the aerodynamic inverse mapping baseline neural network model, its actual equivalent weights are... Represented as ideal fixed weight With low-rank correction term sum: ; Among them, the low-rank correction term is derived from the dimension reduction matrix. and the increasing dimension matrix Multiplying them together, we get , is a very small hyperparameter. and These represent the number of neurons in adjacent connection layers. The matrix is ​​low-rank. During online learning, only these two low-rank matrices (the reduced-dimensional matrix and the increased-dimensional matrix) are updated, while the backbone weights (offline fixed weights) remain frozen. Through this design, the number of parameters updated online in a single layer can be significantly reduced from the original full matrix size to a low-rank size, thereby greatly reducing the online computational burden.

[0032] The final nonlinear aerodynamic inverse mapping neural network model is as follows: ; During the initialization phase, LoRA branches use a dimensionality reduction matrix. Random initialization, increasing the dimension of the matrix Zero initialization: ; in, The variance is used for random initialization of the dimension reduction matrix.

[0033] This initialization strategy ensures low-rank correction terms at the start of online learning. The network output is completely consistent with the offline baseline network. Subsequently, the backbone weight matrix is ​​frozen. The gradient, based on empirical replay of the loss function, is only applicable to low-rank matrices. and Execute with Gradient update: ; in, and These are the online adaptive learning rates; and To provide robust correction coefficients and prevent parameter divergence.

[0034] Through this low-rank adaptive mechanism, the number of parameters updated online in a single layer is reduced from the original... sudden drop This allows for rapid online adaptation to local aerodynamic drift without compromising offline baseline characteristics.

[0035] The specific steps of S4 preferably include: First, establish a control-oriented dynamic model: ; In the formula, ; in, , , These represent the angle of attack, sideslip angle, and roll angle, respectively. , , These represent the angular velocities of the roll, yaw, and pitch channels, respectively. to These are inertial parameters; , , These represent the total torque of the roll, yaw, and pitch channels, respectively. and These represent the attitude angle vector and the angular velocity vector, respectively. The total torque of the three channels is represented by the deformation-induced additional torque and the equivalent rudder deflection angle of the three axes. , , The composition of the aerodynamic torque generated by deflection; This represents a composite disturbance term consisting of external disturbances, parameter uncertainties, and unmodeled dynamics.

[0036] Constructing a hierarchical robust adaptive control architecture (such as...) Figure 2 As shown in the diagram, the attitude control task is decomposed into top-level virtual torque generation and bottom-level aerodynamic inverse mapping calculation. In this layered decoupled architecture, the top-level controller does not directly output rudder deflection commands, but instead outputs the desired virtual control torque. Due to the approximation error in the inverse calculation of the underlying neural network, the actual aerodynamic torque... With top-level expected virtual torque A mapping bias will occur between them. This bias is defined as the network approximation residual of the underlying inverse solution. This is combined with external disturbances and model uncertainties into a composite disturbance. Substituting this composite disturbance into the control-oriented dynamics model, we obtain the equivalent controlled object faced by the top-level controller: ; in, Through this process, the underlying inverse mapping error is uniformly absorbed into the top-level disturbance channel, thus providing a clear design interface for the upper-level robust controller.

[0037] The specific steps of S5 preferably include: To proactively compensate for the disturbances in the above composite process, they are extended into new state variables. And construct an extended state system: ; Among them, the expansion state vector , The rate of change of the disturbance has an upper bound. The system matrices are as follows: , ; Based on this system, an extended state observer is designed: ; in, For the state estimation vector, Let be the observer gain matrix. Define the observation error as... The observation error dynamics can be obtained as follows: ; Let the observer gain matrix be: ; in, For the observer bandwidth, thus making the matrix For a strict Hurwitz matrix, given that the rate of change of the perturbation is bounded, the observer error can converge exponentially to a small, adjustable neighborhood near zero, thus obtaining an accurate estimate of the composite perturbation. .

[0038] It is important to emphasize that the Extended State Observer (ESO) does not completely replace the task of inverse computation of the underlying neural network. Its main role is to provide critical disturbance rejection compensation and security support when the network has not yet completed online adaptation or when the system encounters sudden operating conditions. This design can both reduce the observation burden of the observer and significantly enhance the system's safety margin during the online learning transition phase.

[0039] The specific steps of S6 preferably include: The top-level controller employs a dynamic surface control structure based on backstepping. First, the attitude angle tracking error is defined. And by combining the attitude kinematics equations, we get: ; Treat angular velocity as For the virtual control input of the attitude angle subsystem, based on Lyapunov Stability theory design of kinematic expectation virtual control law : ; in, It is a positive definite diagonal parameter matrix used to adjust the convergence speed of attitude error.

[0040] To avoid the "computational explosion" caused by repeatedly differentiating virtual control instructions in the traditional backstepping method, a time constant is introduced in the recursive process. A first-order low-pass filter is used to smooth the desired virtual angular velocity: ; in, and These represent the smoothed virtual angular velocity command and its derivative, respectively. This represents the dynamic boundary error introduced by the filter.

[0041] Subsequently, the angular velocity tracking error was defined. By combining the dynamic equations that include complex perturbations, we obtain: ; Composite perturbation estimate using the output of the extended state observer By performing feedforward compensation, the final desired virtual control torque at the top level can be obtained. : ; in, It is a positive definite diagonal gain matrix.

[0042] At this point, the top-level controller design is complete. The controller outputs a set of pure torque commands that do not depend on an accurate aerodynamic model, which are then sent as cross-layer signals to the bottom-level inverse solver.

[0043] The specific steps of S7 preferably include: Considering that the top-level distribution is the total desired torque including the variable-structure additional torque, while the bottom-level network essentially learns the mapping relationship between the desired aerodynamic torque coefficient and the rudder deflection, it is necessary to first process the total desired torque... Convert to desired aerodynamic moment coefficient : ; in, Apply a torque to the known deformation. (Transformed) It will be used as the direct input instruction for the neural network.

[0044] Desired aerodynamic torque coefficient and the current state of attack angle Sideslip angle folding angle The input is fed into the corresponding neural network to obtain the rudder deflection command.

[0045] This invention also discloses a robust attitude control system for a variable-structure aircraft, which uses any of the robust attitude control methods for variable-structure aircraft disclosed in this invention to control the attitude of the variable-structure aircraft, including: The benchmark model building module is used to collect wind tunnel test data, numerical simulation data and / or historical flight data of the modified aircraft, and to obtain the aerodynamic inverse mapping benchmark neural network model offline based on the collected data; Learning Sample Construction Module: Used to construct online transfer learning samples based on error-driven adaptive experience replay strategy and real-time flight data during the online flight of the modified aircraft, and store them in the experience replay pool; Adaptive model building module: used to freeze the weights of the aerodynamic inverse mapping baseline neural network model, introduce a low-rank adaptive parameter matrix into the network layer, and combine an error-driven adaptive empirical replay strategy to update only the low-rank parameters to obtain an adaptive aerodynamic inverse mapping neural network model. Attitude control task decomposition module: used to build a hierarchical robust adaptive control architecture, decompose the attitude control task into top-level virtual torque generation and bottom-level aerodynamic inverse mapping calculation, and merge the bottom-level inverse mapping error, external disturbance, parameter uncertainty and unmodeled dynamics into a composite disturbance. Composite disturbance estimation module: used to extend composite disturbances into new state variables, construct extended state observers, and perform online estimation of composite disturbances; Feedforward compensation module: used to generate the desired virtual control torque using dynamic surface control, and to perform feedforward compensation using the composite disturbance estimate output by the extended state observer; Control command generation module: It is used to convert the top-level desired virtual control torque into desired aerodynamic torque coefficient, input the current flight state, variable parameters and desired aerodynamic torque coefficient into the adaptive aerodynamic inverse mapping neural network model, and output three-channel rudder deflection commands.

[0046] Example (verifying the technical effectiveness of the robust attitude control method for the modified aircraft of the present invention): (1) Experimental environment This embodiment first conducts simulation experiments using a digital model of the aircraft under nominal conditions, collecting simulation data to construct offline training and test sets. Then, using the method described in S1, the collected training set data is used to train an offline neural network (an ideal nonlinear aerodynamic inverse mapping model), and the test set data is used to compare and test it with traditional methods. Before the online simulation begins, the aircraft's aerodynamic parameters are randomly biased by 20% to simulate model uncertainty, thereby testing the effectiveness of the online transfer learning strategy. Finally, external disturbances are added to compare the performance of the proposed method with the traditional least squares method in complex environments.

[0047] (2) Experimental results and analysis The offline neural network solution results, online transfer learning control results, and robust control results are as follows: Figure 4 , Figure 5 and Figure 6 As shown. By Figure 4 As can be seen, on the test set, the error heatmaps of the method of this invention maintain a relatively uniform low error distribution in all three channels, with no obvious local distortion areas, and the offline identification error is smaller than that of the comparison method. Figure 5 It can be seen that the forgetting factor adaptively adjusts with the online identification error, thereby reducing the online identification error and further improving the control effect of the method of the present invention under model uncertainty. Figure 6 It can be seen that, under complex conditions of superimposed external interference and model uncertainty, the method of the present invention is not only superior to the comparative method in terms of tracking performance, but also better meets the actual flight control requirements in terms of the smoothness of control commands.

[0048] Unless otherwise specified or further limited to one preferred or optional technical means being another, the preferred and optional technical means disclosed in this invention can be arbitrarily combined to form several different technical solutions.

Claims

1. A robust attitude control method for a variable-structure aircraft, characterized in that... Includes the following steps: S1: Collect wind tunnel test data, numerical simulation data and / or historical flight data of the modified aircraft, and train the aerodynamic inverse mapping benchmark neural network model offline based on the collected data; S2: During the online flight of the variable-structure aircraft, online transfer learning samples are constructed based on the error-driven adaptive experience replay strategy and real-time flight data, and stored in the experience replay pool for online training of the aerodynamic inverse mapping benchmark neural network model. S3: Freeze the weights of the aerodynamic inverse mapping baseline neural network model, introduce a low-rank adaptive parameter matrix into the network layer, and combine it with the error-driven adaptive empirical replay strategy to update only the low-rank parameters to obtain the adaptive aerodynamic inverse mapping neural network model. S4: Construct a hierarchical robust adaptive control architecture, decompose the attitude control task into top-level virtual torque generation and bottom-level aerodynamic inverse mapping calculation, and merge the bottom-level inverse mapping error, external disturbance, parameter uncertainty and unmodeled dynamics into a composite disturbance. S5: Extend the composite disturbance into a new state variable, construct an extended state observer, and perform online estimation of the composite disturbance; S6: Dynamic surface control is used to generate the desired virtual control torque, and feedforward compensation is performed using the composite disturbance estimate output by the extended state observer. S7: Convert the top-level desired virtual control torque into desired aerodynamic torque coefficient, input the current flight state, variant parameters and desired aerodynamic torque coefficient into the adaptive aerodynamic inverse mapping neural network model, and output three-channel rudder deflection commands to control the attitude of the variant aircraft.

2. The robust attitude control method for a variable-structure aircraft according to claim 1, characterized in that... In S1, an ideal nonlinear aerodynamic inverse mapping model and a loss function are defined. The nominal optimal weight of the offline-trained nonlinear aerodynamic inverse mapping model is obtained by minimizing the loss function and recorded as the benchmark weight, thus obtaining the aerodynamic inverse mapping benchmark neural network model.

3. The robust attitude control method for a variable-configuration aircraft according to claim 1, characterized in that... In S2, a sample admission criterion is constructed in the spatial dimension. When the state space distance between a new sample and the most recently pushed sample is less than a preset dead zone threshold, the information increment of the new sample is considered to be limited and it is not pushed onto the stack. Otherwise, the new sample is pushed onto the stack.

4. The robust attitude control method for a variable-configuration aircraft according to claim 3, characterized in that... In the time dimension, maintain a relative age for each sample added to the stack. Each time a new sample is allowed to be added to the stack, the age of all existing samples increases. Define a dynamic forgetting sharpness factor driven by the current network prediction error. Calculate the probability of each sample being eliminated based on the dynamic forgetting sharpness factor, the relative age of the samples pushed into the stack, and the maximum age of the samples in the experience pool. When the experience pool reaches its maximum capacity limit, replace the eliminated samples with new samples according to the elimination probability. The age of the new samples is initialized.

5. The robust attitude control method for a variable-configuration aircraft according to claim 1, characterized in that... In S3, for any fully connected layer in the aerodynamic inverse mapping baseline neural network model, its actual equivalent weight is represented as the sum of the ideal fixed weight and the low-rank correction term. The low-rank correction term is obtained by multiplying the dimension reduction matrix and the dimension increase matrix. During the online learning process, the offline fixed weight is kept frozen, and only the dimension reduction matrix and the dimension increase matrix are updated, so that the number of parameters updated online in a single layer is reduced from the original full matrix scale to the low-rank scale, and finally the nonlinear aerodynamic inverse mapping neural network model is obtained.

6. The robust attitude control method for a variable-configuration aircraft according to claim 1, characterized in that... In S4, a control-oriented dynamic model is established. Under the hierarchical decoupling architecture, the mapping deviation between the actual aerodynamic torque and the top-level expected virtual torque is defined as the network approximation residual of the bottom-level inverse solution. This residual is then combined with external disturbances and model uncertainties to form a composite disturbance. The composite disturbance is substituted into the control-oriented dynamic model to obtain the equivalent control object faced by the top-level controller. This allows the bottom-level inverse mapping error to be uniformly absorbed into the top-level disturbance channel, providing a clear design interface for the upper-level robust controller.

7. The robust attitude control method for a variable-configuration aircraft according to claim 1, characterized in that... In S5, an extended state system is constructed based on the extended state vector, the rate of change of the disturbance, and the top-level expected virtual torque. Then, an extended state observer is designed based on the extended state system. The observation error is defined and the observer gain matrix is ​​constrained so that the observer error can converge exponentially to an adjustable small neighborhood near zero, thereby obtaining an accurate estimate of the composite disturbance.

8. The robust attitude control method for a variable-configuration aircraft according to claim 1, characterized in that... In S6, the top-level controller adopts a dynamic surface control structure based on backstepping. First, the attitude angle tracking error is defined and combined with the attitude kinematics equations, based on... Lyapunov The kinematic desired virtual control law is designed based on stability theory. A first-order low-pass filter is introduced during the recursive process to smooth the desired virtual angular velocity. Then, the angular velocity tracking error is defined and combined with the dynamic equation containing composite disturbances. The composite disturbance estimate output by the extended state observer is used for feedforward compensation to obtain the final desired virtual control torque at the top level.

9. The robust attitude control method for a variable-configuration aircraft according to claim 1, characterized in that... In the S7, the flight state includes the angle of attack and the sideslip angle, and the structural parameters include the folding angle.

10. A robust attitude control system for a variable-configuration aircraft, characterized in that... The attitude control of a variable-structure aircraft is performed using the robust attitude control method for variable-structure aircraft as described in any one of claims 1-9, including: The benchmark model building module is used to collect wind tunnel test data, numerical simulation data and / or historical flight data of the modified aircraft, and to obtain the aerodynamic inverse mapping benchmark neural network model offline based on the collected data; Learning Sample Construction Module: Used to construct online transfer learning samples based on error-driven adaptive experience replay strategy and real-time flight data during the online flight of the modified aircraft, and store them in the experience replay pool; Adaptive model building module: used to freeze the weights of the aerodynamic inverse mapping baseline neural network model, introduce a low-rank adaptive parameter matrix into the network layer, and combine an error-driven adaptive empirical replay strategy to update only the low-rank parameters to obtain an adaptive aerodynamic inverse mapping neural network model. Attitude control task decomposition module: used to build a hierarchical robust adaptive control architecture, decompose the attitude control task into top-level virtual torque generation and bottom-level aerodynamic inverse mapping calculation, and merge the bottom-level inverse mapping error, external disturbance, parameter uncertainty and unmodeled dynamics into a composite disturbance. Composite disturbance estimation module: used to extend composite disturbances into new state variables, construct extended state observers, and perform online estimation of composite disturbances; Feedforward compensation module: used to generate the desired virtual control torque using dynamic surface control, and to perform feedforward compensation using the composite disturbance estimate output by the extended state observer; Control command generation module: It is used to convert the top-level desired virtual control torque into desired aerodynamic torque coefficient, input the current flight state, variable parameters and desired aerodynamic torque coefficient into the adaptive aerodynamic inverse mapping neural network model, and output three-channel rudder deflection commands.

Images