Fault-tolerant control method for hypersonic vehicle based on linear active disturbance rejection control of long short-term memory neural network

Abstract

The application discloses a kind of based on long short-time memory neural network linear active disturbance rejection control hypersonic vehicle fault-tolerant control method, including establishing the mathematical model of hypersonic vehicle and the model of actuator failure, including its dynamic model and kinematic model;Linear active disturbance rejection control technology is used to design controller, for the actuator failure and abnormal situation that hypersonic vehicle can appear corresponding fault-tolerant control strategy is designed, linear active disturbance rejection controller and fault-tolerant control strategy are integrated into the control system of hypersonic vehicle;During flight, linear extended state observer will constantly evaluate current flight state and total disturbance, according to real-time control demand and fault condition, using long short-time memory neural network automatically adjusts observer and controller parameter, ensure the safe operation of hypersonic vehicle.

Description

Technical Field

[0001] This invention relates to the field of hypersonic vehicle control systems and fault-tolerant control, specifically to a fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control. Background Technology

[0002] With the rapid development of science and technology, hypersonic vehicles have broad application prospects in military, aviation, and space exploration fields. Hypersonic vehicles face complex aerodynamic environments and high levels of uncertainty during flight, making the design of their control systems highly challenging. Traditional control methods often fail to meet the control requirements of hypersonic vehicles in complex environments; therefore, researching a novel fault-tolerant control method is of great significance.

[0003] Hypersonic vehicles are affected by various factors during flight, such as uncertainties, unmodeled dynamics, coupling effects, parameter perturbations, and disturbance faults. Linear active disturbance rejection control (ADRC) can treat these effects as a "total disturbance" and compensate for them through observation using an extended state observer. Unlike ordinary low-speed vehicles, each aerodynamic parameter in a hypersonic vehicle is related to the angle of attack and Mach number, and includes uncertainties caused by aerodynamic parameter perturbations, thus enhancing the coupling between channels. By selecting appropriate linear ADRC parameters, good quality and tracking performance can be obtained, and the effects of the "total disturbance" and a wide range of aerodynamic parameter perturbations can be overcome, exhibiting strong robustness.

[0004] While linear active disturbance rejection control (ADRC) can monitor the flight status and fault information of hypersonic vehicles, its effectiveness depends heavily on parameter tuning. Good parameter design yields the desired control performance, but poor parameter design significantly reduces environmental adaptability, potentially leading to the collapse of the entire control system. In fault-tolerant control of hypersonic vehicles, online adjustment of controller parameters is necessary to improve environmental adaptability. Therefore, fault-tolerant control methods for hypersonic vehicles based on long short-term memory neural network-based linear ADRC must balance accuracy, reliability, and timeliness. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and to propose a fault-tolerant control for hypersonic vehicles based on linear active disturbance rejection control using a long short-term memory neural network.

[0006] The technical solution for achieving this invention is: a fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control, comprising the following steps:

[0007] Step S1: Input the aerodynamic parameters of the aircraft into the offline-trained neural network to obtain the modeling parameters of the hypersonic aircraft;

[0008] Step S2: Establish a dynamic model of the hypersonic vehicle's center of mass motion and a dynamic model of its rotation around the center of mass based on aerodynamic parameters;

[0009] Step S3: Use a linear active disturbance rejection controller as the controller for the hypersonic vehicle and perform error updates;

[0010] Step S4: Perform online adaptive updates of observer and controller parameters using a long short-term memory neural network.

[0011] Preferably, the aerodynamic parameters of the aircraft include: hypersonic vehicle velocity V, thrust P, angle of attack α, air resistance X, mass m, gravitational acceleration g, trajectory inclination angle θ, hypersonic vehicle lift Y, Earth radius R, flight altitude y, and ω. x M is the pitch angular velocity, and l is the pitch torque. x It is the pitch moment of inertia. It is the pitch angle.

[0012] Preferably, the aerodynamic parameters include the derivatives of the force coefficient and the moment coefficient with respect to the flight state quantities, and the flight state quantities include the angle of attack, Mach number, and rudder deflection angle.

[0013] Preferably, the dynamic model of the hypersonic vehicle's center of mass motion and the dynamic model of its rotation around the center of mass are specifically as follows:

[0014] The dynamic equations of motion of the center of mass of a hypersonic vehicle:

[0015]

[0016]

[0017]

[0018] In the formula: q and S are the characteristic areas of the hypersonic vehicle and the hypersonic vehicle, respectively; m is the mass of the hypersonic vehicle; g is the acceleration due to gravity; P is the thrust of the hypersonic vehicle engine; V is the flight speed; and V is the flight acceleration. Angle of attack α, sideslip angle β, pitch angle Tilt angle γ v c x , These are the reciprocals of the drag coefficient, lift coefficient, and angle of attack, respectively, and the reciprocal of the lift coefficient and pitch angle, δ. z δ y These are pitch rudder deflection and yaw rudder deflection, respectively. These are the derivatives of the lateral force coefficient with respect to the sideslip angle and the lateral force coefficient with respect to the yaw rate, respectively. Vz Let ω be the rotational angular velocity of the velocity coordinate system relative to the z-axis of the ground coordinate system. Vy The rotational angular velocity of the velocity coordinate system relative to the y-axis of the ground coordinate system is the result of its decomposition.

[0019] The dynamic equations of a hypersonic vehicle rotating around its center of mass:

[0020]

[0021]

[0022]

[0023] In the formula: q and S are the characteristic areas of the hydrodynamic and hypersonic vehicles, respectively; L is the reference length of the hypersonic vehicle; V is the flight speed; α is the angle of attack; β is the sideslip angle; and ω is the sideslip angular velocity. ω x ω y ω z These are the angular velocities of the hypersonic vehicle rotating about its center of mass along the x, y, and z axes, respectively. These represent the angular accelerations of the hypersonic vehicle rotating about its center of mass along the x, y, and z axes, respectively. These are the control moment coefficients for the aileron, yaw, and pitch rudder, respectively. These are the dimensionless roll, yaw, and pitch damping moment coefficients, respectively. These are the derivatives of the yaw moment coefficient with respect to the sideslip angle and the damping moment coefficient under dimensionless yaw, respectively. These are the derivative of the pitch moment coefficient with respect to the angle of attack and the dimensionless pitch damping moment coefficient, respectively.

[0024] Preferably, the specific steps for using a linear active disturbance rejection controller as the controller for the hypersonic vehicle, and for performing error updates and measurement updates, are as follows:

[0025] Step S31: Set the initial value of the bandwidth of the linear extended state observer in the linear active disturbance rejection controller, and the initial value of the control parameters for state error feedback;

[0026] Step S32: Determine the observed estimates of the linear extended state observer for the angle of attack, the observed estimates of the angle of attack angular velocity, and the total estimates of fault information and internal and external disturbances based on the bandwidth of the linear extended state observer.

[0027] Step S33: Divide the output of the extended state observer with the angle of attack of the input signal and the differential angle of attack angular velocity of the input signal to obtain the corresponding error; determine the control surface angle control quantity acting on the hypersonic vehicle based on the control parameters of the state error feedback.

[0028] Preferably, the specific formulas for determining the observed estimates of the angle of attack, the observed estimates of the angle of attack angular velocity, and the total estimates of fault information and internal and external disturbances by the linearly extended state observer, based on the bandwidth of the linearly extended state observer, are as follows:

[0029]

[0030] in:

[0031]

[0032] x1 represents the desired angle of attack, and z1 represents the observer's estimated value for the angle of attack. z1 is the observer's estimated value of the angular velocity at the angle of attack, and z2 is the observer's estimated value of the angular velocity at the angle of attack. z0 is the observer's estimated value for the angle-of-attack angular acceleration, and z3 is the observer's total estimated value for fault information and internal and external disturbances. Let b0 be the derivative of the observer's total estimate of fault information and internal / external disturbances, and δ be the observer gain. α ω is the control input to the hypersonic vehicle model. o This represents the observer bandwidth.

[0033] Preferably, the control surface angle control quantity acting on the hypersonic vehicle is determined based on the control parameters fed back from the state error, and the specific formula is as follows:

[0034]

[0035] In the formula, x1 is the input signal, i.e., the desired angle of attack; x2 is the input signal, i.e., the angular velocity of attack; b is the controller gain; and δ... α Z is the output value of the state error feedback, i.e., the control surface quantity input to the hypersonic vehicle model; Z3 is the total estimate of the observer for fault information and internal and external disturbances; and K is the control surface quantity input to the hypersonic vehicle model. p ,k d These are the control parameters for state error feedback.

[0036] Preferably, the long short-term memory neural network is optimized using an adaptive moment estimation optimization algorithm, and the specific optimization method is as follows:

[0037] Calculate the first and second moment estimates of the gradient:

[0038] m t =β1m t-1 +(1-β1)g t

[0039]

[0040] Among them, g t It is the gradient at the current time t, m t It is the first moment estimate of the gradient, v t It is the second moment estimate of the gradient, where β1 and β2 are the decay rates;

[0041] Bias corrections are applied to the first-order and second-order moment estimates:

[0042]

[0043]

[0044] Update the parameters based on the bias-corrected values:

[0045]

[0046] Where, θ t It is a parameter vector, η is the learning rate, and ∈ represents a value of 10. -8 .

[0047] Preferably, the specific method for online adaptive updating of observer and controller parameters using a long short-term memory neural network is as follows:

[0048] The performance index is determined based on the error between the expected output and the actual output at the current moment. The specific formula is as follows:

[0049]

[0050]

[0051]

[0052]

[0053] In the formula, C eq Here, x1 represents the input signal, i.e., the desired angle of attack; x2 represents the input signal, i.e., the angular velocity of attack; z1 represents the observer's estimated value for the angle of attack; z2 represents the observer's estimated value for the angular velocity of attack; and y represents the actual angle of attack value of the hypersonic vehicle model. The mean square error of the difference between x1 and z1 The mean square error of the difference between x2 and z2 Let x1 be the mean square error of the difference between x1 and y, and k represent the current time.

[0054] The performance index is compared with a set threshold. If the performance index is less than the set threshold, no parameter adjustment is needed. The current bandwidth parameters of the linearly extended observer and the control parameters of the state error feedback are used as the parameters of the aircraft's linear active disturbance rejection fault-tolerant controller for fault-tolerant control. Otherwise, a long short-term memory neural network is used to update the observer and controller parameters. The specific method is as follows:

[0055] The error obtained in step 3, the neural network bias term 1, and the actual output angle of attack are used as inputs to the Long Short-Term Memory (LSTM) neural network. The LSM outputs the adaptively updated parameter values ​​through the output gate, and the output gate formula is as follows:

[0056] o (t) =σ(U o x (t) +W o h (t-1) )

[0057] h (t) =tanh(s (t) )o (t)

[0058] Where: o represents the output of the output gate, σ represents the sigmoid function, and U o Here, x represents the input weights, x represents the network's input information at the current moment, i.e., the current input angle of attack, and W represents the input weights. o is the recurrent weight, h is the neural network output, i.e., the adaptively adjusted parameter value, the superscript plus t indicates the current time step, and the superscript plus t-1 indicates the previous time step.

[0059] Preferably, the loss function of the long short-term memory neural network is a performance index determined based on the error value between the expected output and the actual output at the current moment.

[0060] Compared with the prior art, the significant advantages of this invention are:

[0061] 1. This invention addresses the problems of poor stability and reliability of hypersonic vehicles under fault conditions by combining machine learning methods with linear active disturbance rejection control, which effectively improves the adaptability and robustness of the system and enables it to adapt to different control requirements and environmental changes.

[0062] 2. This invention addresses the difficulty in designing parameters for linear active disturbance rejection controllers by employing a long short-time memory neural network as the basic parameter tuning algorithm and designing a method for online tuning of controller parameters based on machine learning, thereby improving the effectiveness and robustness of the control system.

[0063] 3. Compared with conventional linear active disturbance rejection control, the reliability and real-time performance of the control system are guaranteed; compared with conventional extended state observers, the accuracy of the control system is improved. Attached Figure Description

[0064] Figure 1 This is a control block diagram for a fault-tolerant control method.

[0065] Figure 2 Flowchart for adaptive parameter adjustment.

[0066] Figure 3 This is a training diagram for a long short-term memory neural network.

[0067] Figure 4 A graph showing the parameters being adjusted online.

[0068] Figure 5 This is a control error curve of a linear active disturbance rejection controller based on a long short-time memory neural network under fault-free conditions.

[0069] Figure 6 The diagram shows the control error curve of a linear active disturbance rejection controller based on a long short-term memory neural network under fault conditions.

[0070] Figure 7 This is a flowchart of the fault-tolerant control system. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0072] Existing traditional linear active disturbance rejection control methods are typically applied to aircraft with fixed parameters. However, for hypersonic vehicles, various factors influence flight, such as uncertainties, unmodeled dynamics, coupling effects, parameter perturbations, and disturbance faults. This makes control extremely difficult for controllers with fixed parameters, and also lacks robustness and adaptability. Therefore, this invention proposes a fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control, including the following:

[0073] Step S1: Input the modeling parameters of the aircraft into the offline-trained neural network to obtain the aerodynamic parameters of the hypersonic aircraft. This specifically includes the following steps:

[0074] Step S11: During aircraft flight, accurate aerodynamic parameters are needed to support online flight capability estimation. Additionally, during control system operation, to improve control accuracy and further compensate the control system, aerodynamic parameters related to the pitch moment coefficient must be obtained, including the derivative of the pitch moment coefficient with respect to the angle of attack. The derivatives of the rudder deflection angle are the most important. Since the research object of this invention is a hypersonic aircraft, the zero angle of attack pitch moment term and the zero angle of attack lift term are considered.

[0075] The modeling parameters for the hypersonic vehicle include: hypersonic vehicle velocity V, thrust P, angle of attack α, air resistance X, mass m, gravitational acceleration g, trajectory inclination angle θ, hypersonic vehicle lift Y, Earth radius R, flight altitude y, and ω. x M is the pitch angular velocity, and l is the pitch torque. x It is the pitch moment of inertia. It includes pitch angle, etc.

[0076] In order to achieve online estimation of flight capability during aircraft flight, relatively accurate [data / efficiencies] are required. (The derivative of the drag coefficient with respect to the angle of attack) Aerodynamic parameters such as the derivative of the lift coefficient with respect to the angle of attack provide support. Furthermore, during the operation of a hypersonic vehicle control system, aerodynamic uncertainties can lead to control instability, necessitating the acquisition of... And related aerodynamic derivatives of the pitching moment term, such as the derivative of the pitching moment coefficient with respect to the angle of attack. The derivative of the rudder deflection angle This is to compensate the control system and enhance attitude control stability. In this invention, aerodynamic parameters specifically refer to the derivatives of force coefficients and moment coefficients relative to flight state quantities (angle of attack, Mach number, rudder deflection, etc.).

[0077] Since the control input for a hypersonic vehicle model is the deflection angle, the deflection angle and aerodynamic torque can be converted using the following formula:

[0078] M xg =Q×S ref ×L ref ×C mxg

[0079] M yg =Q×S ref ×L ref ×C myg

[0080] M zg =Q×S ref ×L ref ×C mzg

[0081] Among them, S ref L represents the characteristic area of ​​a hypersonic vehicle. ref Let C be the characteristic length of the hypersonic vehicle, and Q be the dynamic pressure. mxg C myg C mzg The solution formula is as follows:

[0082]

[0083]

[0084]

[0085] In the formula, Δx, Δy, and Δz represent the deviation components of the actual centroid position relative to the reference centroid position (the origin of the coordinate system) along the X, Y, and Z axes, respectively, in the centroid coordinate system, with units of meters. Δx is positive when the actual centroid is closer to a fixed point than the reference centroid. Δy and Δz are defined according to the positive and negative directions of the centroid coordinates.

[0086] In this invention, the origins of the projectile coordinate system, velocity coordinate system, and aircraft coordinate system are all defined at the center of mass, so the deviation is 0, i.e., Δx = Δy = Δz = 0.

[0087] In the above formula Follow the formula below:

[0088] C mx =C mx (BASIC)+ΔC mx (Friction correction) + ΔC mx (D1)+ΔC mx (D2)+ΔC mx (D3)

[0089] C my =C my (BASIC)+ΔC my (Friction correction) + ΔC my (D1)+ΔC my (D2)+ΔC my (D3)

[0090] C mz =C mz (BASIC)+ΔC mz (Friction correction) + ΔC mz (D1)+ΔC mz (D2)+ΔC mz (D3)

[0091] Where ΔC mx (Friction correction) = 0, ΔC mx (Di), ΔC my (Di), ΔC mz (Di) can be obtained by table lookup and interpolation.

[0092] The aerodynamic parameters to be identified include: ΔC mx (Di), ΔC my (Di), ΔC mz (Di).

[0093] The identifiability and accuracy of aircraft aerodynamic parameters depend on both the mathematical model and identification method used, and the amount of information contained in the flight data regarding the parameters to be identified. Different inputs will elicit different motion modes of the aircraft system. When an aircraft is disturbed during flight, the disturbance precisely excites the aircraft's motion modes to a certain extent, laying the foundation for online identification of aerodynamic parameters.

[0094] By using a simple backpropagation (BP) neural network under different excitation functions, aerodynamic parameters are recorded and trained. This yields more accurate and faster modeling parameter values ​​than traditional interpolation methods.

[0095] Step S2: Establish the kinematic and dynamic models of the hypersonic vehicle, which specifically includes the following steps.

[0096] Step S21: Based on the aerodynamic parameters obtained in step S1, the dynamic equations of the hypersonic vehicle can then be established.

[0097] From the fundamental theorem of variable mass dynamics, the equation of motion of the center of mass of the aircraft in the inertial coordinate system is:

[0098]

[0099] For hypersonic vehicles, the velocity coordinate system is a moving coordinate system. The relationship between the dynamic equations established in the moving coordinate system and the inertial coordinate system, i.e., the ground coordinate system, can be expressed by the following formula:

[0100]

[0101] In the formula: Let V be the absolute derivative of the velocity V in the ground coordinate system. ω is the relative reciprocal of velocity V in the velocity coordinate system. V Let ω be the angular velocity of the velocity coordinate system relative to the ground coordinate system.

[0102] The formula for the aerodynamic vector R of a hypersonic vehicle is:

[0103]

[0104] In the formula: -X, Y, Z represent aerodynamic drag, aerodynamic lift, and lateral force, respectively; q and S represent dynamic pressure and the characteristic area of ​​the hypersonic vehicle, respectively; c x , These are the reciprocals of the drag coefficient, lift coefficient, and angle of attack, respectively, and the reciprocal of the lift coefficient and pitch angle, δ. z δ y These are pitch rudder deflection and yaw rudder deflection, respectively. These are the derivatives of the lateral force coefficient with respect to the sideslip angle and the lateral force coefficient with respect to the yaw rudder, respectively.

[0105] In the ground coordinate system, gravity G is along the negative Ay axis. Using the conversion formula between the ground coordinate system and the velocity coordinate system, we can obtain the expression for G in the velocity coordinate system as follows:

[0106]

[0107] The engine thrust vector P is along the positive ox1 axis in the projectile coordinate system, and its expression in the velocity coordinate system is:

[0108]

[0109] Because the velocity V of a hypersonic vehicle is in the same direction as the ox1 axis in the velocity coordinate system, the velocity vector can be expressed as...

[0110]

[0111] ω V The vector form is

[0112]

[0113] Combining all the above equations, we can obtain the dynamic equation for the motion of the center of mass of a hypersonic vehicle:

[0114]

[0115]

[0116]

[0117] Step S22: Establish the kinematic equations of the hypersonic vehicle.

[0118] The vector expression of the dynamic equations of a hypersonic vehicle rotating about its center of mass is:

[0119]

[0120] In the formula: H is the angular momentum of the hypersonic vehicle about its center of mass in the inertial coordinate system; M p The torque of the aerodynamic force about the center of mass; M R The torque of the engine thrust about the center of mass, because it is assumed that the engine thrust passes through the center of mass, M R =0

[0121] Similarly, based on the relationship between the absolute and relative reciprocals of the velocity coordinate system and the ground coordinate system, we can obtain...

[0122]

[0123] In the formula: The absolute derivative of the angular momentum H in the ground coordinate system

[0124] The relative derivative of the angular momentum H in the projectile coordinate system

[0125] The expression for angular momentum H is:

[0126] H = J·ω

[0127] In the formula: J is the inertia tensor

[0128] The angular momentum in the projectile coordinate system can be expressed as:

[0129]

[0130] In the formula: Let X be the moment of inertia of the hypersonic vehicle relative to the X, Y, and Z axes of the missile body coordinate system. The product of inertia of the hypersonic vehicle relative to the X, Y, and Z axes of the missile body coordinate system.

[0131] Since it is assumed that the principal inertial axis of the hypersonic vehicle coincides with the coordinate axis of the missile body coordinate system, the product of inertia is zero.

[0132] The above expression can be simplified to

[0133]

[0134] Aerodynamic torque M R In the projectile coordinate system, it can be represented as

[0135]

[0136] In the formula: Control moment coefficients for aileron, yaw, and pitch rudder, respectively. These are the dimensionless roll, yaw, and pitch damping moment coefficients, respectively. These are the derivatives of the yaw moment coefficient with respect to the sideslip angle and the damping moment coefficient under dimensionless yaw, respectively. These are the derivatives of the pitch moment coefficient with respect to the angle of attack, respectively, and the dimensionless pitch damping moment coefficient L is the reference length of the hypersonic vehicle.

[0137] Combining all the above equations, we obtain the dynamic equations for the hypersonic vehicle's rotation around its center of mass.

[0138]

[0139]

[0140]

[0141] Step S3: A linear active disturbance rejection controller (ADRC) is used as the controller for the hypersonic vehicle. Taking single-channel (pitch channel) attitude control as an example, the controller's input is the hypersonic vehicle's aerodynamic attitude angle (angle of attack), and the output is the hypersonic vehicle's actuator mechanism, i.e., the control surface angle. The control block diagram of the fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network ADRC is shown below. Figure 1 As shown, it specifically includes the following steps.

[0142] Step S31: Select the initial value of the bandwidth ω0 of the linear extended state observer in the linear active disturbance rejection controller, and the control parameter k for state error feedback. p k d Initial value.

[0143] Step S32: Determine the observed estimates of the linearly extended state observer for the angle of attack, the observed estimates of the angular velocity at the angle of attack, and the total estimates of fault information and internal and external disturbances based on the bandwidth of the linearly extended state observer.

[0144]

[0145] in:

[0146]

[0147] The main parameter in a linearly extended state observer is the observer bandwidth ω. o In the formula, x1 is the desired angle of attack, and z1 is the observer's estimated value of the angle of attack. z1 is the observer's estimated value of the angular velocity at the angle of attack, and z2 is the observer's estimated value of the angular velocity at the angle of attack. z0 is the observer's estimated value for the angle-of-attack angular acceleration, and z3 is the observer's total estimated value for fault information and internal and external disturbances. Let b0 be the derivative of the observer's total estimate of fault information and internal / external disturbances, and δ be the observer gain. α This refers to the control surface inputs to the hypersonic vehicle model.

[0148] Step S33: Error Update

[0149] The outputs of the extended state observer, z1, z2, and z3 in step S31, are respectively subtracted from the input signal (the desired angle of attack) and the derivative of the input signal (the angular velocity of attack) to obtain errors e1 and e2. Here, e1 represents the difference between the input signal angle of attack x1 and the output z1 of the extended state observer, and e2 represents the difference between the input signal angular velocity of attack x2 and the output z2 of the extended state observer.

[0150]

[0151] The sum of fault and disturbance information observed by the extended state observer, z3, is fed back to the state error feedback. Based on the control parameters of the state error feedback, the control surface angle control quantities acting on the hypersonic vehicle are determined.

[0152]

[0153] The main controller parameter is k p k d In the formula, x1 is the desired angle of attack, x2 is the desired angle of attack angular velocity, z1 is the observer's estimated value for the angle of attack, z3 is the observer's total estimate for fault information and internal and external disturbances, b is the controller gain, and δ... α This refers to the control parameters applied to the control surfaces of a hypersonic vehicle.

[0154] Step S4: Perform online adaptive updates of observer and controller parameters using a Long Short-Term Memory (LSTM) neural network. The flowchart is as follows: Figure 2 As shown, it specifically includes the following steps.

[0155] Step S41: Subtract the estimated value obtained by the observer in step S33 from the actual angle of attack output, and calculate the mean square error of the error value obtained in step S32;

[0156]

[0157]

[0158]

[0159] x1 is the input signal, i.e., the desired angle of attack; x2 is the input signal, i.e., the angular velocity of attack; z1 is the observer's estimated value of the angle of attack; z2 is the observer's estimated value of the angular velocity of attack; and y is the actual angle of attack value of the hypersonic vehicle model. The mean square error of the difference between x1 and z1 The mean square error of the difference between x2 and z2 The mean square error of the difference between x1 and y is given. The addition of k after the formula is to describe the current time.

[0160] The performance index is

[0161] Step S42: Predicted value correction

[0162] The performance index obtained in step S41 is used as the loss function of the long short-term memory neural network. The adaptive moment estimation optimization algorithm is used to optimize the long short-term memory neural network to calculate the observer and controller parameters at the next time step.

[0163] The adaptive moment estimation optimization algorithm is as follows:

[0164] First, calculate the first moment estimate (mean) and second moment estimate (variance) of the gradient:

[0165] m t =β1m t-1 +(1-β1)g t

[0166]

[0167] Among them, g t It is the gradient at the current time t, m t It is the first moment estimate of the gradient, v t This is the second moment estimate of the gradient, where β1 and β2 are the decay rates (typically taken as 0.9 and 0.999). Since m t and v t Initially, all values ​​are 0, requiring bias correction:

[0168]

[0169]

[0170] Then, update the parameters based on the above estimates:

[0171]

[0172] Where, θ t It is a parameter vector, η is the learning rate, and ∈ is a small constant added for numerical stability (usually taken as 10). -8 ).

[0173] Step S43: Analyze the performance index C obtained in step S41. eq Conduct an online evaluation; if performance metric C... eq Less than the set threshold σ U If the current linear extended observer bandwidth parameters and state error feedback control parameters are used as parameters for the aircraft's linear active disturbance rejection fault-tolerant controller, then no parameter adjustment is required. Otherwise, proceed to step S44. Iteration stopping criteria: The goal is to reduce the performance index to 0, i.e., the expected input angle of attack is equal to the angle of attack value estimated by the observer, the expected input angle of attack angular velocity is equal to the angle of attack angular velocity estimated by the observer, and the expected input angle of attack angular velocity is equal to the actual angle of attack. However, in reality, it is difficult to reduce it to 0, so if the performance index is less than the set value σ... U =0.05; of which As a condition for stopping the iteration, when the difference between the actual output angle of attack and the expected input angle of attack is within 2°, the subsequent online parameter adjustment work is stopped.

[0174] After steps S3 and S4 are completed, the error value between the expected output and the actual output at the current moment will be obtained. To minimize the error, the parameters of the observer and controller are updated online.

[0175] Simultaneously, based on the error obtained in step S32, the neural network deviation term 1 and the actual output angle of attack y of the hypersonic vehicle are used as inputs to the long short-term memory neural network, such as... Figure 3 As shown; the short-time memory neural network outputs adaptively updated parameter values ​​through an output gate. The output gate formula is as follows:

[0176] o (t) =σ(U o x (t) +W o h (t-1) )

[0177] h (t) =tanh(s (t) )o (t)

[0178] Where: o represents the output of the output gate, σ represents the sigmoid function, and U o Here, x represents the input weights, x represents the network's input information at the current moment, i.e., the current input angle of attack, and W represents the input weights. o is the recurrent weight, h is the neural network output, i.e., the adaptively adjusted parameter value, the superscript plus t indicates the current time step, and the superscript plus t-1 indicates the previous time step.

[0179] like Figure 4 The curve shown represents the online adjustment of parameters.

[0180] Step S45: Based on the principle of simplicity and ease of use, directly replace the initial values ​​set in step S31 with the online updated parameters, and use the output true angle of attack y and the errors e1 and e2 obtained in step S32 as the inputs of the linear expansion state observer, return to step S32, and recalculate.

[0181] Under this strategy, the linear active disturbance rejection controller proposed in this invention utilizes a long short-term memory neural network to dynamically optimize the observer bandwidth and controller parameters online in real time based on state error and system output information. While ensuring the BIBO stability of the closed-loop system, it improves the controller's observation and compensation accuracy for total disturbances, thereby enhancing the fault tolerance of the entire closed-loop system. This method endows the traditional control linear active disturbance rejection controller with data-driven adaptive capabilities. The algorithm is simple and effective, and has promising engineering application prospects.

[0182] The fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control proposed in this invention can effectively address attitude control during actuator failures. For example... Figure 5 As shown, this represents the control error of a linear active disturbance rejection controller based on a long short-time memory neural network under fault-free conditions, such as... Figure 6 As shown, the control error of the linear active disturbance rejection controller based on a long short-term memory neural network under fault conditions is represented. The red curve represents the control error obtained using the method of this invention, while the black curve represents the control error of a traditional PID controller. Comparing the two, the method proposed in this invention effectively reduces the control error regardless of whether a fault occurs. More importantly, this method employs online parameter updates, improving the robustness of the system.

[0183] This invention addresses the fault-tolerant attitude control problem of hypersonic vehicles in the event of actuator failure, proposing a fault-tolerant control strategy for hypersonic vehicles based on linear active disturbance rejection control (ADR) using a long short-term memory (LSTM) neural network. Corresponding fault-tolerant control strategies are designed to address potential actuator failures and anomalies in hypersonic vehicles, integrating the linear ADR controller and the fault-tolerant control strategy into the hypersonic vehicle's control system. During flight, a linear extended state observer continuously evaluates the current flight state and total disturbance. Based on real-time control requirements and fault conditions, the LTM neural network automatically adjusts the observer and controller parameters to ensure the safe operation of the hypersonic vehicle. This invention, by introducing an LTM neural network for adaptive parameter adjustment, organically combines linear ADR with fault-tolerant control, effectively addressing various failures and anomalies that may occur in hypersonic vehicles and ensuring their safe operation. Furthermore, this invention exhibits high adaptability and robustness, capable of adapting to different control requirements and environmental changes.

[0184] This invention introduces a long short-term memory neural network for adaptive parameter adjustment, organically combining linear active disturbance rejection control with fault-tolerant control. This effectively addresses various faults and anomalies that may occur in hypersonic vehicles, ensuring their safe operation. Furthermore, this invention exhibits high adaptability and robustness, capable of adapting to different control requirements and environmental changes.

[0185] The above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control, characterized in that, Includes the following steps: Step S1: Input the aerodynamic parameters of the aircraft into the offline-trained neural network to obtain the modeling parameters of the hypersonic aircraft; Step S2: Establish a dynamic model of the hypersonic vehicle's center of mass motion and a dynamic model of its rotation around the center of mass based on aerodynamic parameters; Step S3: Employ a linear active disturbance rejection controller as the controller for the hypersonic vehicle and perform error updates. The specific steps are as follows: Step S31: Set the initial value of the bandwidth of the linear extended state observer in the linear active disturbance rejection controller, and the initial value of the control parameters for state error feedback; Step S32: Determine the observed estimates of the linear extended state observer for the angle of attack, the observed estimates of the angle of attack angular velocity, and the total estimates of fault information and internal and external disturbances based on the bandwidth of the linear extended state observer. Step S33: Divide the output of the extended state observer with the angle of attack of the input signal and the differential angle of attack angular velocity of the input signal to obtain the corresponding error; determine the control surface angle control quantity acting on the hypersonic vehicle based on the control parameters of the state error feedback; Step S4: Perform online adaptive updates of observer and controller parameters using a long short-term memory neural network.

2. The fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control according to claim 1, characterized in that, The aerodynamic parameters of the aircraft include: hypersonic vehicle speed. ,thrust Angle of attack air resistance ,quality gravitational acceleration Ballistic tilt angle Lift of hypersonic vehicles Earth's radius Flight altitude pitch angular velocity Pitch moment Pitch moment of inertia Pitch angle .

3. The fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control according to claim 2, characterized in that, The aerodynamic parameters include the derivatives of the force coefficient and moment coefficient with respect to the flight state quantities, which include the angle of attack, Mach number, and rudder deflection angle.

4. The fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control according to claim 1, characterized in that, The dynamic models for the motion of the center of mass and the rotation around the center of mass of the hypersonic vehicle are as follows: The dynamic equations of motion of the center of mass of a hypersonic vehicle: In the formula: , These are the characteristic areas of dynamic pressure and hypersonic vehicles, respectively. For the mass of a hypersonic vehicle, gravitational acceleration Hypersonic vehicle engine thrust Flight speed Flight acceleration Angle of attack Sideslip angle Pitch angle Tilt angle , , , These are the reciprocals of the drag coefficient, lift coefficient, and angle of attack, respectively, and the reciprocal of the lift coefficient and pitch angle. , These are pitch rudder deflection and yaw rudder deflection, respectively. , These are the derivatives of the lateral force coefficient with respect to the sideslip angle and the lateral force coefficient with respect to the yaw rudder, respectively. The rotational angular velocity of the velocity coordinate system relative to the z-axis of the ground coordinate system is given. The rotational angular velocity of the velocity coordinate system relative to the y-axis of the ground coordinate system is the result of its decomposition. The dynamic equations of a hypersonic vehicle rotating around its center of mass: In the formula: , These are the characteristic areas of dynamic pressure and hypersonic vehicles, respectively. The reference length and flight speed of the hypersonic vehicle. Angle of attack Sideslip angle Sideslip angular velocity , These are hypersonic vehicles orbiting their center of mass. Angular velocity of the shaft rotation, These are hypersonic vehicles orbiting their center of mass. Angular acceleration of the axis rotation, These are the control moment coefficients for the aileron, yaw, and pitch rudder, respectively. These are the dimensionless roll, yaw, and pitch damping moment coefficients, respectively. These are the derivatives of the yaw moment coefficient with respect to the sideslip angle and the damping moment coefficient under dimensionless yaw, respectively. These are the derivative of the pitch moment coefficient with respect to the angle of attack and the dimensionless pitch damping moment coefficient, respectively.

5. The fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control according to claim 1, characterized in that, The specific formulas for determining the observed estimates of the angle of attack, the observed estimates of the angular velocity at the angle of attack, and the total estimates of fault information and internal and external disturbances by the linear extended state observer, based on the bandwidth of the linear extended state observer, are as follows: in: For the desired angle of attack, This represents the observer's estimated value for the angle of attack. This represents the observer's estimated value for the angular velocity at the angle of attack. This is the observer's estimated value for the angular velocity at the angle of attack. This is the observer's estimated value for the angle-of-attack angular acceleration. This is the observer's overall estimate of the fault information and internal and external disturbances. This is the derivative of the observer's total estimate of the fault information and internal and external disturbances. For observer gain, These are the control surface inputs to the hypersonic vehicle model. This represents the observer bandwidth.

6. The fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control according to claim 1, characterized in that, The control surface angle control parameters acting on the hypersonic vehicle are determined based on the control parameters fed back from the state error. The specific formula is as follows: In the formula, The input signal is the desired angle of attack. The input signal is the angle of attack angular velocity. For controller gain, The output value of the state error feedback is the control surface input to the hypersonic vehicle model. This is the observer's overall estimate of the fault information and internal and external disturbances. These are the control parameters for state error feedback.

7. The fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control according to claim 1, characterized in that, The Long Short-Term Memory Neural Network is optimized using an adaptive moment estimation optimization algorithm. The specific optimization method is as follows: Calculate the first and second moment estimates of the gradient: in, It is the current moment. gradient, It is a first-moment estimate of the gradient. It is a second-moment estimate of the gradient. and It is the attenuation rate; Bias corrections are applied to the first-order and second-order moment estimates: Update the parameters based on the bias-corrected values: in, It is a parameter vector. It's the learning rate. The value is .

8. The fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control according to claim 1, characterized in that, The specific method for online adaptive updating of observer and controller parameters using a long short-term memory neural network is as follows: The performance index is determined based on the error between the expected output and the actual output at the current moment. The specific formula is as follows: In the formula, For performance indicators, The input signal is the desired angle of attack. The input signal is the angle of attack angular velocity. This represents the observer's estimated value for the angle of attack. This is the observer's estimated value for the angular velocity at the angle of attack. These are the actual angles of attack values ​​for the hypersonic vehicle model. for and Mean square error of the difference for and Mean square error of the difference for and Mean square error of the difference Indicates the current time; The performance index is compared with a set threshold. If the performance index is less than the set threshold, no parameter adjustment is needed. The current bandwidth parameters of the linearly extended observer and the control parameters of the state error feedback are used as the parameters of the aircraft's linear active disturbance rejection fault-tolerant controller for fault-tolerant control. Otherwise, a long short-term memory neural network is used to update the observer and controller parameters. The specific method is as follows: The error obtained in step 3, the neural network bias term 1, and the actual output angle of attack are used as inputs to the Long Short-Term Memory (LSTM) neural network. The LSM outputs the adaptively updated parameter values ​​through the output gate, and the output gate formula is as follows: in: Indicates the output gate output. This refers to the sigmod function. These are the input weights. It is the network input information at the current moment, that is, the input angle of attack at the current moment. It is the recurrent weight. It is the output of the neural network, i.e., the adaptively adjusted parameter values, indicated by the superscript. This indicates the current time, indicated by the superscript. Indicates the previous moment.

9. The fault-tolerant control method for hypersonic vehicles based on long short-term memory neural network linear active disturbance rejection control according to claim 1, characterized in that, The loss function of the Long Short-Term Memory Neural Network is a performance index determined based on the error between the expected output and the actual output at the current moment.

Images