Intelligent control design method for hypersonic variable configuration aircraft based on layered architecture

Abstract

The present invention belongs to the field of aerospace vehicle control, and relates to a hypersonic variable configuration aircraft intelligent control design method based on a hierarchical architecture. The present invention first establishes a hypersonic variable configuration aircraft dynamics model for the preparation of intelligent control. After that, an intelligent decision-making algorithm based on LSTM is designed to construct a basic framework of intelligent control, which can take into account the uncertainty and nonlinearity of the hypersonic variable configuration aircraft model. By designing an intelligent decision-making layer, the instructions of the guidance control loop are decided in real time according to the overall flight mission, current flight status, constraints, and optimization indicators. Furthermore, an HMV intelligent control algorithm is designed to reduce the influence of model uncertainty and nonlinearity caused by the variable configuration on hypersonic flight control, and at the same time solve the coupling problem of attitude dynamics and center of mass dynamics caused by the variable configuration.

Description

Technical Field

[0001] The present invention belongs to the intelligent guidance and control technology of hypersonic aircraft, and relates to an intelligent control design method for hypersonic variable configuration aircraft based on a hierarchical architecture. Background Art

[0002] Hypersonic vehicles have great economic, political and military value due to their high flight speed, complex flight profile, wide combat airspace, strong concealment performance, strong penetration capability, and flexible take-off methods. Hypersonic variable configuration vehicle (HMV) is a new type of aircraft that combines the advantages of hypersonic flight technology and variable configuration technology. Traditional hypersonic aircraft are mostly aimed at a single design point, which can ensure that the aircraft maintains a better performance at this design point, but cannot take into account the performance at non-design points and cannot meet the wide-range flight development needs of hypersonic aircraft. HMV can autonomously change its shape and structure according to the different operating environments encountered during hypersonic flight to obtain better aerodynamic characteristics and maneuverability, thereby improving flight performance. Therefore, the research on HMV technology has become an important trend in the development of future hypersonic aircraft.

[0003] During hypersonic flight, changes in structural shape increase the nonlinearity and uncertainty of the dynamic model, including: 1) The variable configuration will cause rapid changes in aerodynamic loads and dynamic models, resulting in considerable uncertainty and strong coupling in the control system; 2) The variable configuration process causes complex additional forces and torques, and there is a multivariate nonlinear mapping relationship between aerodynamic forces such as lift, drag, and pitch moment and the flight state and variable configuration control quantity, which increases the nonlinearity of the model; 3) The shape change causes perturbations in aerodynamic parameters, making it difficult to establish an accurate aerodynamic model through ground tests. The model still has mechanisms that are not fully analyzed, and uncertainty is needed to characterize the unmodeled dynamics. In the traditional trajectory design link, the changes in aerodynamic forces caused by structural variable configurations are usually considered to further change the results of trajectory optimization design. However, the guidance and control link pays more attention to the impact of torque changes caused by variable configurations. That is, the traditional design ideas are difficult to solve the guidance and control coupling problems caused by variable configurations, and cannot meet the high stability and robustness requirements. Through the idea of ​​a layered architecture, intelligent decision-making is introduced before traditional guidance and control. Through deep learning, the instructions required for guidance and control are decided based on the flight mission and current status, solving the guidance and control coupling problem caused by variable configuration.

[0004] Wang Zijian, Zhang Shuyu, and Hou Mingzhe designed an online time-varying parameter identification algorithm in "Configuration-variable aircraft control based on online parameter identification" (see Journal of Ordnance Equipment Engineering, October 2022, Issue 10, Pages 60-65) to determine the influence of factors such as the uncertainty of aerodynamic coefficients caused by shape changes and external disturbances on the aircraft, and ensure the maneuverability of the aircraft through a state feedback control strategy with fixed time convergence. This method distributes the control of configuration-variable instructions by controlling the input to change as little as possible near the working point. It lacks consideration of the coupling of attitude dynamics and particle dynamics caused by configuration-variable, which may lead to the contradiction between the attitude control method and the actual flight mission requirements.

[0005] Yang Bo, Zhu Yichuan, Wei Yanming, and Fan Zichen proposed a trajectory optimization and control analysis of foldable variant aircraft in "Trajectory Optimization and Control Analysis of Foldable Variant Aircraft" (see "China Space Science and Technology", June 2020, Vol. 40, No. 3, pp. 64-75), and based on the characteristics of near-space gliding hypersonic, Gaussian pseudo-spectral method was used to optimize the trajectory of fixed-wing aircraft and foldable wing aircraft. By comparing the foldable wing aircraft with the traditional fixed-wing aircraft in terms of range capability and heat flow avoidance capability, a trajectory optimization idea with comprehensive objectives was proposed. However, this method does not consider the influence of the foldable wing configuration on the attitude control loop, and does not consider the aerodynamic uncertainty caused by the folding deflection, which has certain limitations.

[0006] Wang Peng, Chen Haolan, and Tang Guojian introduced an auxiliary system based on the control model for input saturation in "A control method for a deformable aircraft with input saturation and output limitation" (CN116795126A), which is used to guide the control quantity of the control model for input saturation out of the saturation state, and use the backstepping method to build a backstepping controller to track the instructions of the deformable aircraft. However, this patent only considers the control saturation and output limitation of the traditional control surfaces of the controlled object, and its research object does not consider the deformable influence of hypersonic aircraft. The method involved cannot handle the deformation control problem with strong time-varying characteristics, strong coupling characteristics and high uncertainty well.

[0007] In the "Preset Performance Control Method of Hypersonic Variant Aircraft Based on Interval Theory" (CN118466569A), Tan Junyue, Liao Yuxin, Xu Weiping and others used the dynamic surface control method to design the control law based on interval theory step by step, which effectively solved the problem that the traditional preset performance control scheme produces a large control action and causes singularity when the state error is close to the envelope. However, this method does not consider model uncertainties and external disturbance torques, and cannot be applied to real flight scenarios that consider variable configuration instructions in the guidance loop, which has certain limitations. Summary of the invention

[0008] In order to solve the above problems, the present invention proposes a hypersonic variable configuration aircraft intelligent control design method based on a hierarchical architecture. The method can take into account the uncertainty and nonlinearity of the hypersonic variable configuration aircraft model. By designing an intelligent decision layer, the instructions of the guidance control loop are decided in real time according to the overall flight mission, current flight status, constraints, and optimization indicators, thereby reducing the model uncertainty and nonlinearity caused by the variable configuration on the hypersonic flight control, and solving the coupling problem of attitude dynamics and center of mass dynamics caused by the variable configuration. At the same time, an intelligent active disturbance rejection control (ADRC) control structure based on the RBF neural network state expansion observer is designed to suppress the internal and external disturbances during the HMV flight process and track the guidance control instructions output by the intelligent decision layer.

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

[0010] The intelligent control design method of a hypersonic variable configuration aircraft based on a hierarchical architecture includes the following steps:

[0011] Step (1) Dynamic modeling of hypersonic variable configuration aircraft

[0012] First define the relevant coordinate system:

[0013] Body coordinate system : origin of the coordinate system coincides with the center of mass of the aircraft, The axis points from the center of mass to the front of the aircraft. The axis is located in the longitudinal symmetry plane of the aircraft and is Vertical and pointing upwards, Axis and , Vertical and form a right-hand tie.

[0014] Velocity coordinate system : origin of the coordinate system coincides with the instantaneous center of mass of the aircraft, The axis coincides with the velocity vector, The axis is located in the longitudinal symmetry plane of the aircraft and is Vertical and pointing upwards, Axis and , Vertical and form a right-hand tie.

[0015] After analyzing the stress on the longitudinal channel of the hypersonic variable configuration vehicle HMV, the longitudinal kinematic equation is obtained:

[0016] (1)

[0017] in, is the aircraft speed, is the aircraft altitude, is the pitch angle, is the angle of attack, is the pitch angular velocity, with the superscript “ ” denotes derivative; is the mass of the aircraft, For HMV around the body The moment of inertia of the shaft, , , They represent the lift, drag and thrust of the aircraft respectively. is the pitching moment, and the calculation formula is:

[0018] (2)

[0019] in, is the longitudinal characteristic length, is the characteristic area, is the atmospheric density. , , , are the lift coefficient, drag coefficient, thrust coefficient and pitch moment coefficient respectively, is the fuel equivalence ratio, , , It is expressed as:

[0020] (3)

[0021] in, , , is the aerodynamic coefficient corresponding to the 0° attack angle, and the extension deformation rate related, , , is the aerodynamic coefficient related to the change of angle of attack, , , is the aerodynamic coefficient associated with the elevator change, is the coefficient term related to the pitch angular velocity in the pitch moment coefficient. Defined as:

[0022] (4)

[0023] In the formula, For exhibition length; and are the minimum deformation extension and the maximum deformation extension respectively;

[0024] Step (2) Design of intelligent decision-making algorithm based on LSTM

[0025] The LSTM deep learning neural network is used to generate guidance and control instructions including altitude, speed, and deformation according to the overall flight mission, current state, constraints, and other information, to realize the intelligent decision-making function of the hypersonic deformable aircraft. The intelligent decision-making process of the present invention is as follows:

[0026] First, the training samples are extracted. The Gaussian pseudospectral method is used to generate the data sample pairs required for training. The Gaussian pseudospectral method is a numerical method for solving optimal control problems. By converting the continuous optimal control problem into a discrete nonlinear programming problem, the optimal trajectory that meets the constraints is found. The basic idea is to divide the time interval into several subintervals, and use the Gauss-Legendre integral point in each subinterval to approximate the changes in state and control quantities. By imposing constraints on the control and state quantities at the integral points, and fitting the dynamic equations, the continuous optimal control problem is converted into a discrete optimization problem. The Gaussian pseudospectral method can effectively plan the optimal trajectory that meets the given performance indicators and constraints under the conditions of handling deviations and process interference. Using the above advantages of the Gaussian pseudospectral method, the initial conditions, terminal conditions, process constraints, aerodynamic parameters, etc. of the trajectory are pulled to obtain a trajectory sample library under multiple flight missions, multiple indicators, and multiple uncertainties, that is, the data sample pairs required by the neural network. In each trajectory sample, the neural network input is set as information such as the flight mission and the current flight status, and the network output is the altitude guidance instruction. , speed guidance command , Transformation Instructions . Divide the sample data into training set and test set according to a certain ratio (usually 1:9 or 2:8).

[0027] Then, the LSTM deep learning neural network training is performed. The training set data is input into the LSTM deep learning neural network, and the parameters such as output nodes, number of hidden layers, number of hidden units, number of training times, etc. are set. The ADAM optimizer is used to train the neural network. The optimizer ADAM can iteratively update the weights of the neural network according to the training data, and has the characteristics of adaptive learning rate; and has good optimization capabilities for sparse gradient and noise problems.

[0028] Finally, the model is tested and verified. The test set data is input into the trained LSTM deep learning neural network, and the prediction accuracy of the model is verified by comparing the guidance instructions output by the test set with the true value, and the MAE indicator is used for accuracy evaluation.

[0029] Step (3) Design of HMV intelligent control algorithm

[0030] A HMV intelligent control algorithm is designed below to track the altitude and speed instructions output by the intelligent decision-making layer. The velocity loop uses traditional PD control, and the altitude loop uses a PID controller. In the attitude loop, an intelligent ADRC control structure based on the RBF neural network state expansion observer (RBF-NN-ESO) is designed to track the attitude control instructions output by the altitude loop. Through the time scale separation principle and the HMV dynamic model, the attitude control loop is divided into an outer loop angle control and an inner loop angular velocity control. The outer loop is responsible for tracking the desired attitude of the entire system and generating the desired attitude for the inner loop. The inner loop mainly tracks the angular velocity instructions generated by the outer loop. Then, ADRC control design based on the RBF neural network state expansion observer is performed on the inner and outer loops respectively, specifically including: inner and outer loop differential trackers , inner and outer loop nonlinear state error feedback control law , Inner and outer loop neural network state expansion observer The introduction of the neural network state expansion observer effectively solves the limitations of traditional ADRC in dealing with hypersonic vehicle variant control problems and improves the performance of the control system. It can better identify and compensate for system disturbances, reduce control jitter, and simplify controller design, providing an effective method for high-speed vehicle control.

[0031] Beneficial effects of the present invention:

[0032] The technical advantages of the present invention are:

[0033] 1. When solving the HMV flight configuration change decision problem, an intelligent decision-making algorithm based on the LSTM method is designed, combined with the Gaussian pseudo-spectral method to generate a large number of data samples, to achieve online optimal instruction decision-making considering flight missions, flight status, constraints, and optimization indicators, and obtain a continuous sequence of state quantities and control quantities, giving full play to the full potential within the configuration change range.

[0034] 2. Considering the challenges posed to attitude control by strong model uncertainty and severe internal and external disturbances caused by the configuration change process of hypersonic aircraft, an intelligent ADRC control algorithm based on RBF neural network state expansion observer is designed to achieve precise attitude control under the complex disturbances of moment of inertia, additional force, additional torque and aerodynamic parameters caused by the configuration change process.

[0035] The key points of the present invention are:

[0036] 1. A hypersonic variable configuration vehicle intelligent control design method based on a hierarchical architecture is adopted to realize online flight decision-making under multiple factors and explore the mission maneuverability of the HMV aircraft.

[0037] 2. The tandem intelligent ADRC control architecture based on the RBF neural network state expansion observer is adopted to effectively suppress the strong nonlinear disturbance of attitude control while having the characteristics of simple design and easy implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a basic flow chart of the intelligent control design method for hypersonic variable configuration aircraft based on a layered architecture.

[0039] Figure 2 This is a schematic diagram of the intelligent control structure of a hypersonic variable-configuration aircraft based on a layered architecture.

[0040] Figure 3 This is a schematic diagram of the LSTM cell structure.

[0041] Figure 4 It is a schematic diagram of the intelligent ADRC structure based on the RBF neural network state expansion observer.

[0042] Figure 5 It is a highly simulated result.

[0043] Figure 6 is the angle of attack simulation result.

[0044] Figure 7 is the pitch angle simulation result.

[0045] Figure 8 It is the fuel equivalence ratio command result.

[0046] Fig. 9 It is the speed control result.

[0047] Fig.10 This is the simulation result of the elevator rudder deflection. DETAILED DESCRIPTION

[0048] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0049] like Figure 1 As shown, the intelligent control design method of a hypersonic variable configuration aircraft based on a layered architecture of the present invention has the following implementation steps:

[0050] Step 1: Establish a dynamic model of a hypersonic variable-length aircraft for subsequent intelligent decision-making and intelligent control design.

[0051] Step 2: Design an LSTM-based intelligent decision-making algorithm to realize online optimal instruction decision-making considering flight missions, flight status, constraints, and optimization indicators.

[0052] Step 3: Design the HMV intelligent control algorithm, establish the basic intelligent control architecture, and realize attitude stability control under model uncertainty and disturbance conditions.

[0053] Step 4: Evaluate whether the designed method meets the performance requirements. If not, readjust the parameters of the second and third steps to complete the simulation calculation and performance test.

[0054] The specific steps are as follows:

[0055] Step 1: Establish the dynamic model of hypersonic variable-length aircraft:

[0056] After the force analysis of the longitudinal channel of HMV, the longitudinal kinematic equation can be listed:

[0057] (1)

[0058] in, is the aircraft speed, is the aircraft altitude, is the pitch angle, is the angle of attack, is the pitch angular velocity, with the superscript “ ” denotes derivative; is the mass of the aircraft, For HMV around the body The moment of inertia of the shaft, , , They represent the lift, drag and thrust of the aircraft respectively. is the pitching moment, and the calculation formula is:

[0059] (2)

[0060] in, is the longitudinal characteristic length, is the characteristic area, is the atmospheric density. , , , are the lift coefficient, drag coefficient, thrust coefficient and pitch moment coefficient respectively, is the fuel equivalence ratio, , , It is expressed as:

[0061] (3)

[0062] in, , , is the aerodynamic coefficient corresponding to the 0° attack angle, and the extension deformation rate related, , , is the aerodynamic coefficient related to the change of angle of attack, , , is the aerodynamic coefficient related to the elevator change, is the coefficient term related to the pitch angular velocity in the pitch moment coefficient. Defined as:

[0063] (4)

[0064] In the formula, For exhibition length; and are the minimum deformation extension and the maximum deformation extension respectively;

[0065] Step 2: Design an intelligent decision-making algorithm based on LSTM.

[0066] In deep learning, the Long Short Term Memory (LSTM) network has the characteristics of long-term memory and short-term memory, and can realize the prediction function of multi-feature input and multi-feature output. It plays an important role in solving the problems of highly nonlinear data features and multi-variable prediction. Therefore, the present invention proposes an intelligent decision-making algorithm for hypersonic variable configuration aircraft based on LSTM network to make effective decisions on guidance and control instructions in flight mission scenarios. The overall design framework is as follows: Figure 2 shown.

[0067] The computational memory unit of the LSTM model is composed of a forget gate, an input gate, and an output gate ( Figure 3 ). Assume that hidden units, and the batch size is , the input number is . Then the forward propagation process of a cell unit of LSTM can be described as:

[0068] { f t = σ ( f ˜ t ) = σ ( W f ⋅ [ h t − 1 , x t ] + b f ) i t = σ ( i ˜ t ) = σ ( W i ⋅ [ h t − 1 , x t ] + b i ) g t = t a n h ( g ˜ t ) = t a n h ( W g ⋅ [ h t − 1 , x t ] + b g ) o t = σ ( o ˜ t ) = σ ( W o ⋅ [ h t − 1 , x t ] + b o ) C t = f t ∗ C t − 1 + i t ∗ g t h t = o t ∗ t a n h ( C t ) y t = W y ⋅ h t + b y (5)

[0069] In formula (5), for The hidden state at a moment in time, representing short-term memory. is the state parameter at the current moment, [ h t − 1 , x t ] The time step is concatenated from the above two vectors. The gating unit is defined as follows: , is the forget gate, which can be based on the hidden layer and current status (hereinafter referred to as [ h t − 1 , x t ] Instead), long-term memory of the previous moment Selectively "forget". , is the input gate, according to [ h t − 1 , x t ] Determine the weights of the network that need to be updated with temporary memory, Based on [ h t − 1 , x t ] The temporary memory generated, Together they determine what information in the current network layer needs to be updated. It is the memory state, indicating long-term memory, through the input gate and the forget gate [ h t − 1 , x t ] Obtained after calculation. is the output gate, according to [ h t − 1 , x t ] To determine The weights output to the hidden layer. The current hidden state. It is the output to the next layer of network, representing the predicted value. is the weight matrix of the forget gate, is the weight matrix of the input gate, is the weight matrix of temporary memory, is the weight matrix of the output gate, is the weight matrix of the predicted values. is the bias matrix of the forget gate, is the bias matrix of the input gate, is the bias matrix of temporary memory, is the bias matrix of the output gate, is the bias matrix of the predicted value, is the Hadamard product. and tanh represent the Sigmoid activation function and than activation function.

[0070] It is not difficult to see The elements of are all in the range of 0 to 1. The elements of When the LSTM neural network works, a cell unit first passes through Input, and what was learned at the last moment (representing short-term memory), and then the long-term memory is realized through the sigmoid function of the forget gate based on the mixed information of the two Forgetting occurs, and then passing through the input gate, some of the current information is selected for new understanding and added to the long-term memory. Then the tanh function is performed through the current long-term memory and short-term memory to obtain the current short-term memory, which in turn affects the next cell to form a recursive network.

[0071] The LSTM deep learning neural network is used to generate guidance and control instructions including altitude, speed, and deformation according to the overall flight mission, current state, constraints, and other information, to realize the intelligent decision-making function of the hypersonic deformable aircraft. The intelligent decision-making process of the present invention is as follows:

[0072] First, training samples are extracted. The data sample pairs required for training are obtained by using the Gaussian pseudospectral method. The Gaussian pseudospectral method can effectively plan the optimal trajectory that meets the given performance indicators and constraints under the conditions of handling deviations and process interference. Using the advantage of the Gaussian pseudospectral method, a trajectory sample library under multiple flight missions, multiple indicators, and multiple uncertainties is generated. The flight mission and current flight status are input into the neural network to generate altitude guidance instructions. , speed guidance command , Transformation Instructions , realizing online decision-making, guidance and control instructions while taking into account state constraints, and giving full play to the flight potential of variable configuration aircraft.

[0073] Then the LSTM network training is performed. The data pairs generated by the training are input into the LSTM neural network. In this embodiment, the output nodes are set to 3, the number of hidden layers is 1, the number of hidden units is 124, the number of training times is set to 500, and the neural network is trained using the ADAM optimizer. The optimizer ADAM can iteratively update the weights of the neural network according to the training data, and has the characteristics of an adaptive learning rate; and has good optimization capabilities for sparse gradient and noise problems.

[0074] Finally, the model is tested and verified. The test set data is input into the trained LSTM network, and the prediction accuracy of the model is verified by comparing the guidance instructions output by the test set with the true value, and the MAE indicator is used for accuracy evaluation.

[0075] Step 3: Design HMV intelligent control algorithm.

[0076] After getting the guidance control command, the guidance controller is designed. The speed loop uses the traditional PD control, and the altitude loop uses the PID controller. In the attitude loop, an ADRC control structure based on the RBF neural network state expansion observer is considered.

[0077] The traditional active disturbance rejection control technology (ADRC) is considered to have good anti-disturbance ability in engineering applications based on its "NLSEF + ESO" structure, and has less dependence on the mathematical model of the system. By using ESO, not only each state variable can be estimated, but also model uncertainty and external disturbances can be estimated. However, the bandwidth and parameters of ESO will seriously affect the performance of ADRC, and jitter will occur when they do not match the disturbance characteristics. On the other hand, the moment of inertia, additional force, additional torque and aerodynamic parameters caused by the HMV configuration change process are usually difficult to accurately identify, and the flight environment is complex, which greatly increases the uncertainty of the model and the internal and external disturbances to the system.

[0078] Based on the above reasons, the present invention designs an intelligent ADRC control structure based on RBF neural network state expansion observer (RBF-NN-ESO) for the attitude control of HMF, tracks the attitude control instructions output by the altitude loop, and divides the attitude control loop into an outer loop angle control and an inner loop angular velocity control through the time scale separation principle and the HMV dynamic model. The outer loop is responsible for tracking the desired attitude of the entire system and generating the desired attitude for the inner loop, which mainly tracks the angular velocity instructions generated by the outer loop. Then, the ADRC control design based on the RBF neural network state expansion observer is performed on the inner and outer loops respectively, specifically including: inner and outer loop differential trackers , inner and outer loop nonlinear state error feedback control law , Inner and outer loop neural network state expansion observer The introduction of the neural network state expansion observer effectively solves the limitations of traditional ADRC in dealing with hypersonic aircraft variant control problems and improves the performance of the control system. It can better identify and compensate for system disturbances, reduce control jitter, and simplify controller design, providing an effective method for high-speed aircraft control. The overall framework is as follows Figure 4 shown.

[0079] The height controller and speed controller adopt the following configurations respectively:

[0080] (6)

[0081] in, are the height control and speed control errors, They are the speed and altitude instructions output by the intelligent decision-making link. The speed controller generates a fuel equivalence ratio instruction based on the size of the speed control error. , which is input into the HMV dynamic model, and the height controller controls the error according to the height Generate angle of attack control instructions , this instruction is the expected instruction of the attitude loop In this embodiment, is a controller parameter.

[0082] The following intelligent cascade ADRC strategy is used to design the inner and outer loops of attitude control to track the angle of attack command output by the altitude controller. First, the control object of ADRC is extracted from equation (1) as follows:

[0083] (7)

[0084] in , , , .in It does not directly participate in posture control, but affects the control effect as a posture state parameter. is a nonlinear function, represents the disturbance of the system, is the system control gain function. After the system form is established, The form is also determined simultaneously.

[0085] For the outer loop controller, the controlled object is the first equation in equation (7), where is the equivalent control quantity. For the outer loop differential tracker The designs are as follows:

[0086] (8)

[0087] in, They represent the outer loop differential tracker The tracking signal and differential output, represents the desired angle of attack, is the control amount limit, which is 5 in this embodiment. is the integration step length, which is 0.01s in this embodiment. It is the fastest control synthesis function. Tracking signal output by the module Finally, the input of the nonlinear state error feedback control law is obtained by subtracting the observed value of the neural network state expanded observer.

[0088] Then design the outer loop neural network state expansion observer The present invention introduces a neural network into ADRC to replace the traditional extended state observer, that is, an extended state observer based on a neural network, and identifies the disturbance caused by the HMV configuration change process through the nonlinear network structure of the neural network. Feedback from the HMV dynamics model , outer loop virtual control quantity Perform an expanded state estimation on the outer loop dynamic system and obtain the state The observation results of the total disturbance of the outer loop system , and the disturbance observation is controlled by the outer loop gain Compensation to virtual control amount , to achieve the suppression of disturbances, The specific design process is as follows:

[0089] { e 1 = z 1 1 − x 1 z ˙ 1 1 = W ^ 1 T σ → 1 ( x ^ 1 ) − β 1 ⋅ e 1 + Ψ 1 ( ⋅ ) ⋅ u 1 W ^ ˙ 1 = Γ 1 ( − σ → 1 ( x ^ 1 ) [ f a l ( e 1 , ε 1 , ρ 1 ) ] T − K W 1 W ^ 1 ) (9)

[0090] in, Yes The state observation, for The state observation error is The neural network is Valuation of x ^ 1 = [ x 1 T , x 3 T , V ] T , is the parameter to be adjusted. , is the neural network update rate parameter. In this embodiment .

[0091] In formula (9), The function is expressed as:

[0092] (10)

[0093] in, , , is the parameter to be adjusted.

[0094] Next, the nonlinear state error feedback control law of the outer loop is designed: :

[0095] (11)

[0096] in, express The tracking signal output is The error in state observation, represents the output of the outer loop NLSEF, To adjust the parameters, in this embodiment .

[0097] For the inner loop controller, the controlled object is the second equation in equation (7), where Control quantity. Differential tracker for the inner loop The designs are as follows:

[0098] (12)

[0099] in, represents the inner loop differential tracker The tracking signal and differential output, Represents the output of the outer loop NLSEF, which is the desired control quantity of the inner loop. is the control amount limit, which is 5 in this embodiment. is the integration step length, which is 0.01s in this embodiment.

[0100] For the inner loop system, the neural network state dilation observer of the inner loop Feedback from the HMV dynamics model , inner loop control quantity Perform an expanded state estimation on the outer loop dynamic system and obtain the state The observation results of the total disturbance of the inner loop system are , and the disturbance observation is controlled by the outer loop gain Compensation to control amount , to achieve the suppression of disturbances. Design As shown in the following formula.

[0101] { e 2 = z 2 1 − x 2 z ˙ 2 1 = W ^ 2 T σ → 2 ( x ^ 2 ) − β 2 ⋅ e 2 + Ψ 2 ( ⋅ ) ⋅ u W ^ ˙ 2 = Γ 2 ( − σ → 2 ( x ^ 2 ) [ f a l ( e 2 , ε 2 , ρ 2 ) ] T − K W 2 W ^ 2 ) (13)

[0102] in, represents the control quantity of the inner loop, which is the control instruction of the system in equation (7). Yes The state observation, for The state observation error is The neural network is Valuation of x ^ 2 = [ x 1 T x 2 T ] T , To adjust the parameters, is the neural network update rate parameter.

[0103] The nonlinear state error feedback control law (NLSEF) of the inner loop is as follows.

[0104] (14)

[0105] in express The tracking signal output is The error in state observation, represents the control quantity of the inner loop, To adjust the parameters, in this embodiment .

[0106] Step 4: Simulation verification

[0107] With the help of commonly used computer numerical calculation and simulation software, a simulation test is carried out on the designed "A hypersonic variable configuration aircraft intelligent control design method based on layered architecture". After selecting the design parameters within the allowable range, simulation calculations are carried out and performance tests are performed. If the performance meets the requirements, the design is completed; otherwise, the design parameters need to be adjusted, and the simulation calculations are carried out again and the performance tests are carried out.

[0108] The following simulation scenario design was conducted. In the first 50 seconds, the speed was maintained at 6 Ma, and the altitude reference command increased from the initial state of 26 km to the expected value of 26.2 km. The initial values ​​of the control variables were designed as follows: the initial value of the fuel equivalence ratio was 0.774, and the initial value of the elevator command was -4.3°.

[0109] The results are as follows Figure 5~Figure 10 As shown, Figure 5 The tracking effect of the aircraft on the altitude instruction obtained by online intelligent decision-making is introduced. After the LSTM-based intelligent decision-making method is introduced, the control instruction is the optimal instruction considering multiple tasks and multiple optimization indicators. Figure 6 , Figure 7 They are the state results of angle of attack and pitch angle respectively. It can be seen that under the intelligent ADRC control based on RBF neural network state expansion observer, although the attitude is affected by internal and external disturbances, the overall process remains bounded without divergence and excessive oscillation. Fig.10 is the simulation result of elevator rudder deflection, Fig. 9 For speed control results, Figure 8 is the fuel equivalence ratio control instruction, and it can be seen that the attitude control has no effect on the speed control loop. Under the action of the intelligent control design method for hypersonic variable configuration aircraft based on a hierarchical architecture proposed in the present invention, the speed and altitude of the hypersonic variable configuration aircraft are effectively tracked.

Claims

1. A hypersonic variable configuration aircraft intelligent control design method based on a hierarchical architecture, characterized in that: The following steps are involved: Step (1) Dynamic modeling of hypersonic variable configuration aircraft First define the relevant coordinate system: Body coordinate system O b X b Y b Z b :Origin of coordinate system O b coincides with the center of mass of the aircraft, O b X b The axis points from the center of mass to the front of the aircraft, O b Y b The axis is located in the longitudinal symmetry plane of the aircraft and is aligned with O b X b Vertical and pointing upward, O b Z b Axis and O b X b , O b Y b vertical and form a right-hand system; Velocity coordinate system O a X a Y a Z a :Origin of coordinate system O a coincides with the instantaneous center of mass of the aircraft, O a X a The axis coincides with the velocity vector, O a Y a The axis is located in the longitudinal symmetry plane of the aircraft and is aligned with O a X a Vertical and pointing upward, O a Z a Axis and O a X a , O a Y a vertical and form a right-hand system; After analyzing the stress on the longitudinal channel of the hypersonic variable configuration vehicle HMV, the longitudinal kinematic equation is obtained: Where V is the aircraft speed, H is the aircraft altitude, θ is the pitch angle, α is the angle of attack, ω z is the pitch angular velocity, the superscript "·" indicates the derivative; m is the mass of the aircraft, I Z For HMV around the body O b Z b The moment of inertia of the axis, L, D, T represent the lift, drag and thrust of the aircraft respectively, M is the pitch moment, and the calculation formula is: in, is the longitudinal characteristic length, S ref is the characteristic area, ρ is the atmospheric density; C L , C D , C T , C M are lift coefficient, drag coefficient, thrust coefficient and pitch moment coefficient respectively, φ is the fuel equivalence ratio, C L , C D , C M It is expressed as: in, is the aerodynamic coefficient corresponding to the 0° attack angle, and is related to the extension deformation rate ζ b related, is the aerodynamic coefficient related to the change of angle of attack, is the aerodynamic coefficient related to the elevator change, is the coefficient term related to the pitch angular velocity in the pitch moment coefficient; the elongation deformation rate ζ b Defined as: Where b is the span; b min and b max are the minimum deformation extension and the maximum deformation extension respectively; Step (2) Design of intelligent decision-making algorithm based on LSTM The LSTM deep learning neural network is used to generate guidance and control instructions including altitude, speed, and deformation according to the overall flight mission, current status, and constraint information, and realize the intelligent decision-making function of the hypersonic variant aircraft; the intelligent decision-making process is as follows: First, the training samples are extracted; the Gaussian pseudo-spectral method is used to generate the data sample pairs required for training, and the initial conditions, terminal conditions, process constraints, and aerodynamic parameters of the trajectory are biased to obtain a trajectory sample library under multiple flight missions, multiple indicators, and multiple uncertainties, that is, the data sample pairs required by the neural network. In each trajectory sample, the neural network input is set as the flight mission and current flight status information, and the network output is the altitude guidance instruction H. cmd , speed guidance command V cmd , deformation instruction b ; Divide the sample data into training set and test set according to the proportion; Then, the LSTM deep learning neural network is trained; the training set data is input into the LSTM deep learning neural network, the output nodes, number of hidden layers, number of hidden units, and number of training times are set, and the ADAM optimizer is used to train the neural network; Finally, the model is tested and verified; the test set data is input into the trained LSTM deep learning neural network, and the prediction accuracy of the model is verified by comparing the guidance instructions output by the test set with the true value, and the MAE indicator is used for accuracy evaluation; Step (3) HMV intelligent control algorithm design A HMV intelligent control algorithm is designed to track the altitude and speed instructions output by the intelligent decision-making layer; PD control is used for the speed loop, and PID controller is used for the altitude loop. An intelligent ADRC control structure based on the RBF neural network state expansion observer is designed in the attitude loop to track the attitude control instructions output by the altitude loop. Through the time scale separation principle and the HMV dynamic model, the attitude control loop is divided into an outer loop angle control and an inner loop angular velocity control. The outer loop is responsible for tracking the desired attitude of the entire system and generating the desired attitude for the inner loop. The inner loop tracks the angular velocity instructions generated by the outer loop; then the ADRC control design based on the RBF neural network state expansion observer is performed on the inner and outer loops respectively, including: inner and outer loop differential trackers TD o ,TD i , inner and outer loop nonlinear state error feedback control law NLSEF o ,NLSEF i , Inner-outer loop neural network state expansion observer NNESO O ,NNESO i ; The details are as follows: The height controller and speed controller adopt the following configurations respectively: Among them, e H 、e V are the height control and speed control errors, V cmd , H cmd They are the speed and altitude instructions output by the intelligent decision-making link. The speed controller generates the fuel equivalence ratio instruction φ based on the size of the speed control error. cmd , is input into the HMV dynamic model, and the height controller is based on the height control error e H Generate angle of attack control command α cmd , which is the expected instruction x of the attitude loop d ; is the controller parameter; The following intelligent cascade ADRC strategy is used to design the inner and outer loops of attitude control to track the angle of attack command output by the altitude controller. First, the control object of ADRC is extracted from equation (1) as follows: where x1 = α, x2 = ω z , x3=θ,u=δ e ; F i and i is a nonlinear function, i=1,2, F i represents the disturbance of the system, Ψ i is the system control gain function; For the outer loop controller, the controlled object is the first equation in equation (7), where x2 is the equivalent control quantity; for the outer loop differential tracker TD o The designs are as follows: Among them, v 11 and v 12 They represent the outer loop differential tracker TD o The tracking signal and differential output, x d =α cmd represents the desired angle of attack, r is the control limit, dh is the integration step, and fsun(·) is the optimal control comprehensive function; TD o The output tracking signal v 11 After that, the input of the nonlinear state error feedback control law is obtained by subtracting the observed quantity of the neural network state expansion observer; Then design the outer loop neural network state expansion observer NNESO o ; In ADRC, a neural network is introduced to replace the extended state observer, that is, the extended state observer based on the neural network is used to identify the disturbance caused by the HMV configuration change process through the nonlinear network structure of the neural network; NNESO o The outer loop dynamic system is expanded and estimated by using x1 fed back by the HMV dynamic model and the outer loop virtual control quantity u1 to obtain the observation of the state x1 and the observation results of the total disturbance of the outer loop system. The disturbance observation is compensated to the virtual control quantity u1 through the outer loop control gain Ψ1(·) to suppress the disturbance. o The specific design process is as follows: Among them, z 11 is the state observation of x1, e1 is the state observation error of x1, is the neural network’s estimate of F1(x1,x3,V), β1, ε1, ρ1 are the parameters to be adjusted, Γ1, is the neural network update rate parameter; The fal(·) function in formula (9) is expressed as: The nonlinear state error feedback control law NLSEF of the outer loop o as follows: Among them, es1 represents TD o The tracking signal output is consistent with NNESO O The error of state observation, u1 represents the output of the outer loop NLSEF, r1 is the adjustment parameter; For the inner loop controller, the controlled object is the second equation in equation (7), where u is the control quantity; for the inner loop differential tracker TD i The designs are as follows: Among them, v 21 ,v 22 Denotes the inner loop differential tracker TD i The tracking signal and differential output of the outer loop NLSEF, u1 represents the output of the outer loop NLSEF, is the desired control amount of the inner loop, r is the control amount limit, and dh is the integration step size; For the inner loop system, the inner loop neural network state expansion observer NNESO i The outer loop dynamic system is expanded and estimated by using x2 and the inner loop control quantity u fed back by the HMV dynamic model to obtain the observation of the state x2 and the observation results of the total disturbance of the inner loop system. The disturbance observation is compensated to the control quantity u through the inner loop control gain Ψ2(·) to suppress the disturbance; Design NNESO i As shown in the following formula; Where u represents the control quantity of the inner loop, which is the control instruction of the system in equation (7), and z 21 is the state observation of x2, e2 is the state observation error of x2, is the neural network’s estimate of F2(x1,x2), β2, ε2, ρ2 are adjustment parameters, Γ2, is the neural network update rate parameter; The nonlinear state error feedback control law NLSEF of the inner loop i as follows; Where es2 means TD i The tracking signal output is consistent with NNESO i The error of state observation, u represents the control amount of the inner loop, r2 is the adjustment parameter.

Images