Hypersonic vehicle attitude fault-tolerant control method based on data-driven technology

Abstract

The application discloses a kind of high supersonic aircraft attitude fault-tolerant control methods based on data driving technology, for the actuator failure that occurs in the flight process of high supersonic aircraft, considering the case of additive fault and external disturbance, based on hybrid data-driven model, an adaptive sliding mode robust control strategy is designed.First, for high supersonic aircraft, the principle of flight mechanics is used to establish a mathematical model, and BP neural network and RNN neural network are used to replace the parameters in the model that need to be tested by wind tunnel experiment.Second, based on the sliding mode control theory, an adaptive sliding surface is designed, so that the tracking error can quickly converge to zero under different fault conditions.The application identifies the nonlinear parameters in the controller by introducing a generalized learning system, improving the anti-interference performance of the system.This method is more effective and superior than traditional fault-tolerant control methods in the presence of additive faults and disturbances in high supersonic aircraft.

Description

Technical Field

[0001] This invention belongs to the field of fault-tolerant control and data-driven technology for aircraft, specifically relating to a fault-tolerant attitude control method for hypersonic aircraft based on data-driven technology. Background Technology

[0002] Hypersonic vehicles have become a highly anticipated research area in current military and civilian applications. Research on hypersonic vehicles involves multiple disciplines, such as aerodynamics, thermodynamics, materials science, and control engineering. The advantages of hypersonic vehicles lie in their ability to traverse the atmosphere at high speeds, possessing high maneuverability and ultra-long-range strike capabilities. However, hypersonic vehicles also face a series of challenges, such as complex aerodynamic and thermodynamic environments and high-temperature, high-pressure conditions that lead to material fatigue and ablation. Therefore, research on hypersonic vehicles has both significant application value and considerable challenges.

[0003] However, due to the harsh operating environment and conditions of hypersonic vehicles, they may encounter various unforeseen situations such as actuator failure. Actuator failure is one of the main causes of hypersonic vehicle failure, and its severity not only affects the performance of the hypersonic vehicle but may also threaten the safety and stability of the system. Therefore, research on actuator failure in hypersonic vehicles is of great significance for improving the performance and safety of hypersonic vehicles.

[0004] In hypersonic vehicles, actuator failures can involve multiple actuators, such as thrusters, servos, and valves. The types of actuator failures also vary, including jamming, slackness, performance loss, and constant gain. Based on the control signal loss caused by actuator failure, failures can be broadly categorized into additive and multiplicative failures. Currently, the mainstream solutions for actuator failures include hardware redundancy and analytical redundancy. Hardware redundancy employs multiple backup actuators, allowing the backup actuators to complete the task when the primary actuator fails. Analytical redundancy primarily utilizes fault-tolerant control techniques.

[0005] Due to the unique flight environment of hypersonic vehicles, the commonly used methods for calculating aerodynamic forces and moments are based on ground-based wind tunnel experimental data. However, wind tunnel data can introduce errors of more than 20% during actual flight. Furthermore, the use of simplified linear models to assist in controller design further increases these errors. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and to propose a design method for a fault-tolerant control system for hypersonic aircraft based on data-driven technology.

[0007] The technical solutions of the application are as follows: a high-speed aircraft attitude fault-tolerant control method based on a data-driven technology, comprising the following steps:

[0008] Step S1: establishing a high-speed aircraft attitude kinematics and dynamics model;

[0009] Step S2: inputting a rudder deflection angle into the attitude kinematics and dynamics model to obtain an aircraft attitude angle; inputting the obtained aircraft attitude angle into a trained cascade neural network to obtain a compensation amount; adding the compensation amount to the aircraft attitude angle obtained by the attitude kinematics and dynamics model to obtain a compensated aircraft attitude angle; using an integral link to obtain three-axis attitude angles; and regarding the parallel architecture of the attitude kinematics and dynamics model and the cascade neural network as a hybrid data-driven model;

[0010] Step S3: establishing a fault model for controller design; using a generalized learning system to estimate uncertain terms in the fault model in real time according to the output of the hybrid data-driven model in step S2;

[0011] Step S4: designing an adaptive sliding mode controller adaptive to actuator additive faults; converting a control torque output by the controller to act on a physical entity to ensure that the physical entity can still track a given attitude angle in the case of additive faults.

[0012] Preferably, the high-speed aircraft attitude kinematics and dynamics model is specifically as follows:

[0013]

[0014] wherein θ is a pitch angle, ψ is a yaw angle, and γ is a roll angle; M x , M y , and M z are aerodynamic moments of rotation around the x-axis, the y-axis, and the z-axis of a body coordinate system respectively; J x , J y , and J z are moments of inertia of rotation around the x-axis, the y-axis, and the z-axis of the body coordinate system respectively; ω x , ω y , and ω z are angular velocities of rotation around the x-axis, the y-axis, and the z-axis of the body coordinate system respectively; is a pitch angular velocity, is a yaw angular velocity, is a roll angular velocity; are angular accelerations of rotation around the x-axis, the y-axis, and the z-axis of the body coordinate system respectively.

[0015] Preferably, the specific method for inputting a rudder deflection angle into the attitude kinematics and dynamics model to obtain an aircraft attitude angle is as follows:

[0016] The rudder deflection angle is converted into aerodynamic moment, which is input into the attitude kinematics and dynamics model to obtain the attitude angle of the aircraft, wherein the conversion relationship between the aerodynamic moment and the rudder deflection angle is shown in the following formula:

[0017] M(t) = B(t)delta

[0018] Wherein M(t) e R 3×1 is the three aerodynamic moments rotating around the body coordinate system; B(t) e R 3×4 is the conversion matrix, and delta = [delta1, delta2, delta3, delta4] T is the rudder deflection angle corresponding to the four tail wings of the aircraft.

[0019] Preferably, the hybrid data-driven model adopts a BP neural network to calculate the aerodynamic parameters c ij and the conversion matrix B(t), the BP neural network is trained offline using wind tunnel data, and when used online, the rudder deflection angle delta i , the angle of attack alpha and the sideslip angle beta are input to output the aerodynamic parameters c ij and the conversion matrix B(t) for calculating the aerodynamic force.

[0020] Preferably, the cascade neural network adopts an RNN neural network.

[0021] Preferably, a fault model for controller design is established, and the specific method for estimating the uncertainty term in the fault model in real time according to the output of the hybrid data-driven model in step S2 is as follows:

[0022] According to the normal attitude kinematics and dynamics model, a hypersonic aircraft fault model is constructed for controller design, and the hypersonic aircraft fault model is specifically as follows:

[0023]

[0024] Wherein, F1(x) = 0; G1(x) = I 3×3 is the unit matrix; F2(x) = [d x ,d y ,d z ] T , d x is the coupling term and modeling uncertainty in the roll channel, d y is the coupling term and modeling uncertainty in the yaw channel, and d z is the coupling term and modeling uncertainty in the pitch channel; wherein G2(x) e R 3×3 is a square matrix containing the moment of inertia; f(t) e R 3×1 is a three-channel additive fault term;

[0025] F2(x) is the coupling term and modeling uncertainty term of the model, and the expression is as follows:

[0026]

[0027] Wherein, d x is the coupling term and modeling uncertainty term of the roll channel, d y is the coupling term and modeling uncertainty term of the yaw channel, d z is the coupling term and modeling uncertainty term of the pitch channel; J x , J y , J z are the moments of inertia around the x-axis, y-axis and z-axis of the body coordinate system respectively; θ is the pitch angle, ψ is the yaw angle, and γ is the roll angle; is the pitch angular velocity, is the yaw angular velocity, is the roll angular velocity; ω x , ω y , ω z are the angular velocities around the x-axis, y-axis and z-axis of the body coordinate system respectively; are the angular accelerations around the x-axis, y-axis and z-axis of the body coordinate system respectively;

[0028] G2(x) is a square matrix containing the moments of inertia, and the expression is as follows:

[0029]

[0030] Wherein, θ is the pitch angle, ψ is the yaw angle, and γ is the roll angle; J x , J y , J z are the moments of inertia around the x-axis, y-axis and z-axis of the body coordinate system respectively;

[0031] f(t) is the sum of the additive fault term and the environmental disturbance term, since the modeling uncertainty term F2(x) and the fault term f(t) are both unknown, the sum of the two is regarded as the total uncertainty of the system, denoted as Δ(x, t), and the fault model of the system is rewritten as follows:

[0032]

[0033] Wherein, F1(x) = 0; G1(x) = I 3×3 is the unit matrix; F2(x) = [d x , d y , d z ] T , d x is the coupling term and modeling uncertainty term of the roll channel, d y is the coupling term and modeling uncertainty term of the yaw channel, d zFor the pitch channel coupling terms and modeling uncertainties; where G2(x)∈R 3×3 Let Δ(x,t) be a square matrix including the moment of inertia; x ,Δ y ,Δ z ] T Let Δ be the total uncertainty of the system. x For the uncertainty term of the roll-off channel, Δ y For the uncertainty term of the yaw channel, Δ z This represents the uncertainty of the rollover path;

[0034] Step S32: Use a generalized learning system to estimate the system uncertainty Δ(x,t). The generalized learning system uses mixed data to drive the model's data for offline pre-training and adaptively updates the parameters online. The output of the generalized learning system is shown below:

[0035] A = W T Φ(z)+Hz+ε

[0036] Where z is the input of the neural network, W and H are the network weights, ε is the network estimation error, and Φ(·) is the radial basis function, as shown in the following expression:

[0037]

[0038] Where c represents the cluster center of the input data, and σ represents the bandwidth of that cluster center. In online applications, the W and H parameters in the network are updated using an adaptive law. The estimation expression for uncertainties in the generalized learning system is written as:

[0039]

[0040] in For the output of the neural network, and The adaptive parameter update method is as follows:

[0041]

[0042] in and The adaptive parameter is updated in a differential form; η1 and η2 are proportional coefficients; S i Φ is the sliding surface of the corresponding channel; Φ(Φ) is the radial basis function; z is the input of the neural network.

[0043] Preferably, the specific method for designing an adaptive sliding mode controller that adapts to additive faults in the actuator is as follows:

[0044] According to the fault model, an adaptive sliding mode fault-tolerant controller is designed, the controller input adaptive parameters, the estimated value of uncertain term, and the output three-axis aerodynamic moment is converted into rudder deflection angle through a conversion matrix and applied to the physical entity. The fault-tolerant control law contained in the controller is as follows:

[0045]

[0046] Wherein, u x , u y , u z are the control inputs of the roll, yaw and pitch channels, and θ d , ψ d , γ d are the expected values of the pitch angle, yaw angle and roll angle respectively; s x , s y , s z are the sliding surfaces of the roll, yaw and pitch channels respectively, is the adaptive parameter of the sliding surface of the roll, yaw and pitch channels; is the uncertain term of the roll, yaw and pitch channels, which is estimated and output by the generalized learning system.

[0047] Preferably, according to the updating method of the adaptive parameter, the three-channel sliding surface is defined as follows:

[0048]

[0049] Wherein s x , s y , s z are the sliding surfaces of the roll, yaw and pitch channels respectively; e x , e y , e z are the tracking errors of the roll, yaw and pitch channels; is the adaptive parameter of the sliding surface of the roll, yaw and pitch channels; is the derivative of the tracking error of the roll, yaw and pitch channels.

[0050] Preferably, the updating expression of the adaptive parameter is as follows:

[0051]

[0052] Wherein, is the differential updating form of the adaptive parameter.

[0053] Compared with the prior art, the present application has the following advantages:

[0054] 1. The mixed data-driven model is adopted for system modeling, the traditional dynamic model is retained, and the error caused by wind tunnel data is compensated by a neural network.

[0055] 2. The PID type adaptive sliding mode surface is designed, the gain parameter in the sliding mode surface is designed as an adaptive updating form, the sliding mode surface can adjust the gain online according to different fault conditions, and the purpose of rapid convergence is achieved.

[0056] 3. The nonlinear function in the estimation controller of the generalized learning system is introduced, compared with the traditional radial basis neural network, the estimation error of the generalized learning system is smaller, and the convergence speed is faster. DETAILED DESCRIPTION

[0057] Figure 1 It is a control system overall block diagram

[0058] Figure 2 It is a mixed data-driven model structure diagram

[0059] Figure 3 It is a generalized learning system topology diagram

[0060] Figure 4 It is a generalized learning system pitch channel uncertainty estimation curve

[0061] Figure 5 It is a generalized learning system yaw channel uncertainty estimation curve

[0062] Figure 6 It is a generalized learning system roll channel uncertainty estimation curve

[0063] Figure 7 It is a pitch channel angle tracking curve

[0064] Figure 8 It is a yaw channel angle tracking curve

[0065] Figure 9 It is a roll channel angle tracking curve DETAILED DESCRIPTION

[0066] In order to make the purpose, technical scheme and advantages of the present application more clear, specific implementation, the present application will be further described in detail below.

[0067] The traditional kinematics and dynamics equation is adjusted in a mixed data-driven manner, the system uncertainty is estimated by introducing a generalized learning system, and the present application provides a hyper-sonic vehicle attitude fault-tolerant control method based on data-driven technology.

[0068] To verify the effectiveness of the proposed data-driven adaptive fault-tolerant control scheme: in the simulation experiment, the high-speed hypersonic vehicle studied in the additive fault condition, with no adaptive traditional sliding mode control is compared. In the estimation of the uncertainty, respectively, using RBF neural network and generalized learning system as a comparison, to verify the effectiveness of the technology.

[0069] The overall structure diagram of the technology is shown in Figure 1 The present application provides a technical solution: a data-driven technology-based hypersonic vehicle attitude fault-tolerant control method, comprising the following steps: step S1, establishing a hypersonic vehicle kinematics, dynamics model, specifically including the following steps.

[0070] Step S11: In order to model the hypersonic vehicle, the body coordinate system is adopted, and the origin of the coordinate system coincides with the center of gravity of the vehicle. By using Newton-Lagrange formula, Newton-Euler formula, the kinematics and dynamics equations of the hypersonic vehicle can be obtained:

[0071]

[0072] In the above formula, the first three formulas are the kinematics equations describing the three Euler angles, and the last three formulas are the dynamics equations describing the three Euler angular velocities. Where θ is the pitch angle, ψ is the yaw angle, and γ is the roll angle; M x , M y , M z are the moments of the aerodynamic force around the body coordinate system; J x , J y , J z are the three-axis rotational inertia; ω x , ω x , ω x are the three-axis angular velocities. Considering the design of the hypersonic vehicle body as a tail cross, without side wing, the rudder deflection angle is taken as the control input of the model, and the aerodynamic moment coefficient is introduced, and the relationship between the moment coefficient and the aerodynamic moment is written as shown in the following formula:

[0073]

[0074] Where S ref is the characteristic area of the vehicle, L ref is the characteristic length of the vehicle, Q is the dynamic pressure, and the aerodynamic moment coefficients C mxg , C myg , C mzg and the relationship between the rudder deflection angle δ i (i=1, 2, 3, 4) of the four tail wings is given by the data obtained by the wind tunnel experiment of the vehicle on the ground. The relationship between the rudder deflection angle and the aerodynamic moment can be approximately linear mapping, written as shown in the following formula:

[0075] M(t) = B(t) δ

[0076] where M(t) ∈ R 3×1 are three aerodynamic moments around the body coordinate system; B(t) ∈ R 3×4 is a coefficient matrix calculated from wind tunnel test data; δ = [δ1, δ2, δ3, δ4] T are the deflection angles of the four control surfaces of the aircraft.

[0077] Step S12: Simplify the kinematic model and the dynamic model, and rewrite them into the form of affine nonlinear model. First, derive the kinematic model of each channel to obtain the following equation.

[0078]

[0079] Step S13: Substitute the dynamic model to obtain:

[0080]

[0081] Step S14: Take x = [x1 T ,x2 T ] T as the state variable of the system, where x1 = [γ, ψ, θ] T , Take u = [M x ,M y ,M z ] T as the control input, and write the following equation into the form of affine nonlinear model:

[0082]

[0083] where F1(x) = 0, G1(x) = I 3×3 , F2(x) = [d x ,d y ,d z ] T , G2(x) ∈ R 3×3 . F2(x) is the coupling term and the modeling uncertainty of the model, and its expression is as follows:

[0084]

[0085] G2(x) is a square matrix containing the moment of inertia, and its expression is as follows:

[0086]

[0087] Step S15: Convert the control input from aerodynamic torque to control surface deflection angle through linear transformation. The complete affine nonlinear system model is shown below:

[0088]

[0089] Where u t =δ=[δ1,δ2,δ3,δ4] T To simplify the controller design described later, based on virtual controller technology, u t The system is transformed into three equivalent rudder deflection angles through linear transformation, specifically represented as u. t =[v x ,δ y ,δ z ].

[0090] Step S2: Utilize the data measured by the sensors to establish a hybrid data-driven model. The model structure diagram is shown below. Figure 2 As shown, it specifically includes the following steps.

[0091] Step S21: Figure 2 The BP neural network in the example is trained offline using wind tunnel data, and the input rudder deflection angle δ is used online. i The angle of attack α and sideslip angle β can be used to output the aerodynamic parameters c for calculating aerodynamic forces. ij And the transformation matrix B(t). Since wind tunnel data is collected as discrete data, interpolation methods are usually used in practice to interpolate and solve for the uncollected data. However, high-dimensional interpolation methods have high computational complexity and slow calculation speed. Experiments have verified that using a trained BP neural network to solve for aerodynamic parameters can greatly reduce the calculation time and shorten the response time of the entire system.

[0092] Step S22: Figure 2 Recurrent Neural Networks (RNNs) in hypersonic vehicles have strong processing capabilities for sequential data. Since hypersonic vehicles typically do not experience abrupt changes in aerodynamic forces or torques during flight, RNNs are used to predict angular acceleration compensation and perform real-time angular acceleration compensation. By adding the acceleration output from the dynamic model to the compensation output from the RNN, the final actual angular acceleration is obtained, and then an integral step is used to calculate the three-axis attitude angles.

[0093] Step S3: Design a generalized learning system, based on... Figure 2 The output of the hybrid data-driven model is based on Figure 3 The topology design of the generalized learning system estimates the deviation between the dynamic model and the physical entity output in real time. The generalized learning system is a variant of the traditional RBF neural network, adding a direct mapping from the input layer to the output layer. Specifically, it includes the following steps:

[0094] Step S31: Introducing an additive fault term in the affine nonlinear model. The fault is represented as an unknown nonlinear term added at the end of the analytical form of the model. The corresponding system fault model is shown as follows:

[0095]

[0096] where f(t)∈R 3×1 is the sum of the additive fault term and the environmental disturbance term. Since both the modeling uncertainty term and the fault term are unknown, the sum of the two is regarded as the total uncertainty term of the system, denoted as Δ(x, t). Then the fault model of the system is rewritten as follows:

[0097]

[0098] Step S32: According to the topology of the generalized learning system Figure 2 , the expression of the estimated uncertainty term is determined. According to the topology structure diagram, the output of the generalized learning system can be derived as shown below:

[0099] Δ = W T Φ(z) + Hz + ε

[0100] where z is the input of the neural network, W and H are the network weights, ε is the network estimation error, and Φ(·) is the radial basis function, which is expressed as follows:

[0101]

[0102] In the above formula, c is the clustering center of the input data, and σ is the bandwidth of the clustering center, both of which are obtained through offline training of the existing data set, and W and H parameters in the network are updated through the adaptive law when used online.

[0103] The entire generalized learning system is regarded as a control system uncertainty estimation module, and the three-channel position information [x, y, z] and the three-channel attitude angular velocity of the hybrid data-driven model output are selected as the inputs of the generalized learning system; the three-channel control input [γ c , ψ c , θ c ] is selected as the input of the generalized learning system; and the uncertainty term of the model is selected as the output of the generalized learning system Then the estimation expression of the generalized learning system for the uncertainty term can be written as:

[0104]

[0105] where and the adaptive update expressions are shown as follows:

[0106]

[0107] Step S4, design an adaptive sliding mode control strategy to adapt to the additive fault of the actuator and the model uncertainty, estimate the uncertain term by using a generalized learning system to ensure the tracking performance of the system under fault-free conditions, and specifically include the following steps.

[0108] Step S41: According to the control target, let the expected value of the pitch angle be θ d , the expected value of the yaw angle be ψ d , and the expected value of the roll angle be γ d , and define the tracking error as:

[0109]

[0110] Step S42: Define the sliding mode surface function as a comprehensive PID type. The three-channel sliding mode surface function is defined as follows:

[0111]

[0112] Step S43: Wherein is the parameter of the sliding mode surface, x, y, and z correspond to the roll channel, the yaw channel, and the pitch channel, respectively. Since the high-speed supersonic aircraft will have a large degree of change in the mathematical model once the actuator fails during flight, if the sliding mode surface is set to a fixed coefficient at this time, the control performance will be poor. Therefore, in order to further improve the control performance in the fault state, the sliding mode surface parameter is set to an adaptive parameter instead. The expression of the adaptive sliding mode surface is as follows:

[0113]

[0114] Wherein, the differential expression of the adaptive parameter update is as follows:

[0115]

[0116] Step S44: According to the value of the sliding mode surface at the last time and the tracking error at the current time, update the sliding mode surface parameter at the current time by using the above formula, and then use the parameter to construct a new sliding mode surface to achieve the adaptive effect.

[0117] Verify the stability of the closed-loop system formed by the high-speed supersonic aircraft and the controller, and design the following Lyapunov function:

[0118]

[0119] Take the derivative of the Lyapunov function:

[0120]

[0121] where the derivative of the sliding surface is:

[0122]

[0123] Substitute into :

[0124]

[0125] Substitute the failure model of the yaw channel into the control law u y , which is the component of the output of the sliding mode controller on the yaw channel after transformation by the virtual controller. After simplification, we obtain:

[0126]

[0127] Substitute the control law u y into , and we obtain:

[0128]

[0129] Substitute the adaptive update law of the sliding surface of the yaw channel into the above equation, and we obtain:

[0130]

[0131] Substitute the adaptive update law of the generalized learning system into the above equation, and we obtain

[0132]

[0133] Combine like terms in the above equation, and we obtain:

[0134]

[0135] Define , and the above equation can be written as:

[0136]

[0137] Take as the upper bound of the estimated value Δ y , and we have , then

[0138]

[0139] Therefore, the Lyapunov stability of the system is proved.

[0140] According to the definition of the sliding surface, when t→∞, s y →0, and e y →0. That is, at this time, the yaw angle ψ on the yaw channel →ψ dangular velocity

[0141] The fault-tolerant control method for a hypersonic vehicle provided by the application can make the vehicle effectively track a given reference input in the case of additive faults.

[0142] Figures 4 to 6 For the three-channel uncertainty estimation results, the simulation experiment compares the estimation effects of the generalized learning system and the RBF neural network. From the comparison chart of the uncertainty estimation of the three channels, it can be seen that whether the generalized learning system or the RBF neural network is used, the uncertainty can be accurately estimated within a limited error range. At the same time, according to the local magnified chart, it can be seen that when the generalized learning system is used, the estimation accuracy of the uncertainty is higher than that of the traditional RBF neural network. Through calculation, the average estimation errors of the three channels of the generalized learning system are respectively: 2.57*10 -4 , 6.57*10 -4 , 8.52*10 -3 ; the average estimation errors of the three channels of the RBF neural network are respectively: 4.73*10 -3 , 1.11*10 -3 , 1.81*10 -2 . It can be seen that in each channel, the estimation error of the generalized learning system is smaller and the approximation accuracy is higher.

[0143] Figures 7 to 9 The system response curve under the additive fault, the physical entity is replaced by a traditional nonlinear model, and the fault is injected at the time of 30s. From the above curve chart, it can be seen that under normal working conditions, the use of the adaptive sliding surface can accelerate the convergence speed, and due to the accurate estimation of the system uncertainty by the generalized learning system, compared with the traditional RBF neural network, the overshoot of the angle tracking will be reduced, and the convergence speed will be faster. After the system is injected with an additive fault, it can be seen from the curve chart that when the adaptive sliding surface is not used, once the additive fault is too large or the sliding surface parameters are not reasonably selected, the angle tracking curve is easy to diverge. When the adaptive technology is used to dynamically adjust the sliding surface parameters, the robustness of the system is improved, and the convergence speed is accelerated. On this basis, due to the accurate estimation of the additive fault by the generalized learning system, the overshoot of the angle tracking of the system under the fault state is smaller, and the convergence speed is faster.

[0144] The application is directed to the attitude tracking problem of hypersonic vehicles under additive fault, and a fault-tolerant control strategy based on data driving is proposed. The control method mainly includes three modules: hybrid data-driven model module, generalized learning system estimation module and adaptive sliding mode controller module. The hybrid data-driven model is realized by the cascade of neural networks, the BP neural network is used to speed up the solution, and the RNN neural network is used to compensate the error between the nonlinear model and the physical entity; the generalized learning system is used to estimate the system uncertainty required by the designed controller; the adaptive sliding mode controller has stronger robustness to faults compared with the traditional controller. The comparison simulation results show that the method is more effective and superior than the traditional control method under additive fault.

[0145] The above-described embodiments are merely specific implementations of the present application, which are used to illustrate the technical solutions of the present application, but not to limit the same. The protection scope of the present application is not limited thereto. Although the present application 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 make modifications or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features, within the technical scope disclosed by the present application. The modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A data-driven attitude fault-tolerant control method for hypersonic vehicles, characterized in that, Includes the following steps: Step S1: Establish the attitude kinematics and dynamics model of the hypersonic vehicle; Step S2: Input the control surface deflection angle into the attitude kinematics and dynamics model to obtain the aircraft attitude angle; input the obtained aircraft attitude angle into the trained cascaded neural network to obtain the compensation amount; add the compensation amount to the aircraft attitude angle obtained from the attitude kinematics and dynamics model to obtain the compensated aircraft attitude angle; use the integral element to calculate the three-axis attitude angle. The architecture of parallel attitude kinematics and dynamics model and cascaded neural network is called hybrid data-driven model. Step S3: Establish a fault model for controller design. Based on the output of the hybrid data-driven model in Step S2, use a generalized learning system to estimate the uncertainties in the fault model in real time. The output of the generalized learning system is shown below: Δ=W T Φ(z)+Hz+ε Where z is the input to the neural network, W and H are the network weights, ε is the network estimation error, and Φ(·) is the radial basis function, expressed as follows: Where c represents the cluster center of the input data, and σ represents the bandwidth of that cluster center. In online applications, the W and H parameters in the network are updated using an adaptive law. The estimation expression for uncertainties in the generalized learning system is written as: in For the output of the neural network, and The adaptive parameter update method is as follows: in and This is the differential update form of the adaptive parameters; η1 and η2 are proportionality coefficients; S i This refers to the sliding surface of the corresponding channel; Step S4: Design an adaptive sliding mode controller to adapt to additive faults in the actuator. The controller output torque is converted and applied to the physical entity to ensure that the physical entity can still track the given attitude angle even under additive fault conditions. The specific method is as follows: Based on the fault model, an adaptive sliding mode fault-tolerant controller is designed. The controller takes adaptive parameters and estimated values ​​of uncertainties as inputs and outputs three-axis aerodynamic torques. The aerodynamic torques are transformed into control surface deflection angles using a transformation matrix and then applied to the physical entity. The fault-tolerant control law contained in the controller is given as follows: Among them, u x u y u z θ is the control input for the roll, yaw, and pitch channels. d ψ d γ d These are the expected values ​​of pitch angle, yaw angle, and roll angle, respectively; s x s y s z These are three sliding surfaces: roll, yaw, and pitch. The adaptive parameters for the three-channel sliding surface of roll, yaw, and pitch; The uncertainties in the roll, yaw, and pitch channels are estimated and output by the generalized learning system.

2. The hypersonic vehicle attitude fault-tolerant control method based on data-driven technology according to claim 1, characterized in that, The attitude kinematics and dynamics model of the hypersonic vehicle is specifically as follows: Where θ is the pitch angle, ψ is the yaw angle, and γ is the roll angle; M x M y M z These are the aerodynamic moments of rotation about the x-axis, y-axis, and z-axis of the body coordinate system, respectively; J x J y J z ω represents the moment of inertia of rotation about the x-axis, y-axis, and z-axis of the body coordinate system, respectively; x ω y ω z These are the angular velocities of rotation about the x-axis, y-axis, and z-axis of the body coordinate system, respectively. The pitch angular velocity, Yaw angular velocity, It is the roll angular velocity; These are the angular accelerations of rotation about the x-axis, y-axis, and z-axis of the body coordinate system, respectively.

3. The hypersonic vehicle attitude fault-tolerant control method based on data-driven technology according to claim 2, characterized in that, The specific method for obtaining the aircraft attitude angles by inputting the control surface deflection angle into the attitude kinematics and dynamics model is as follows: The control surface deflection angle is converted into aerodynamic torque, which is then input into the attitude kinematics and dynamics model to obtain the aircraft's attitude angles. The specific conversion relationship between aerodynamic torque and control surface deflection angle is shown in the following formula: M(t)=B(t)δ Where M(t)∈R 3×1 These are the three aerodynamic torques rotating about the body coordinate system; B(t)∈R 3×4 The transformation matrix is ​​δ = [δ1, δ2, δ3, δ4]. T This refers to the deflection angle of the control surfaces corresponding to the four tail fins of the aircraft.

4. The hypersonic vehicle attitude fault-tolerant control method based on data-driven technology according to claim 1, characterized in that, The hybrid data-driven model uses a BP neural network to calculate the aerodynamic parameters c of the aerodynamic forces. ij And the transformation matrix B(t), the BP neural network is trained offline using wind tunnel data, and the input is the control surface deflection angle δ when used online. i Given i = 1, 2, 3, 4, the angle of attack α and sideslip angle β can be used to output the aerodynamic parameters c for calculating aerodynamic forces. ij And the transformation matrix B(t).

5. The hypersonic vehicle attitude fault-tolerant control method based on data-driven technology according to claim 1, characterized in that, The cascaded neural network uses an RNN neural network.

6. The hypersonic vehicle attitude fault-tolerant control method based on data-driven technology according to claim 3, characterized in that, The specific method for establishing a fault model for controller design, and estimating the uncertainties in the fault model in real time using a generalized learning system based on the output of the hybrid data-driven model in step S2, is as follows: Step 31: Based on the normal attitude kinematics and dynamics model, construct a hypersonic vehicle fault model for controller design. The hypersonic vehicle fault model specifically includes: Where F1(x) = 0; G1(x) = I 3×3 It is the identity matrix; F2(x) = [d x ,d y ,d z ] T d x For the coupling terms and modeling uncertainties of the roll channel, d y For the coupling terms and modeling uncertainties of the yaw channel, d z For the pitch channel coupling terms and modeling uncertainties; where G2(x)∈R 3×3 Let f(t) be a square matrix including the moment of inertia; 3×1 For a three-channel additive fault term, u t =δ=[δ1,δ2,δ3,δ4] T ; F2(x) represents the coupling terms and modeling uncertainties of the model, expressed as follows: The expression for G2(x) is as follows: Since both the modeling uncertainty F2(x) and the fault term f(t) are unknowns, their sum is considered as the total uncertainty of the system, denoted as Δ(x,t). The fault model of the system is rewritten as follows: Where, Δ(x,t)=[Δ x ,Δ y ,Δ z ] T Δ x For the uncertainty term of the roll-off channel, Δ y For the uncertainty term of the yaw channel, Δ z This represents the uncertainty of the rollover path; Step S32: Use a generalized learning system to estimate the total uncertainty term Δ(x,t) of the system. The generalized learning system uses mixed data to drive the model's data for offline pre-training and online adaptive parameter updates.

7. The hypersonic vehicle attitude fault-tolerant control method based on data-driven technology according to claim 1, characterized in that, Based on the adaptive parameter update method, the three-channel sliding surface is defined as follows: Among them, e x e y e z For tracking errors in the roll, yaw, and pitch channels; This is the derivative of the tracking error for the roll, yaw, and pitch channels.

8. The hypersonic vehicle attitude fault-tolerant control method based on data-driven technology according to claim 7, characterized in that, The adaptive parameter update expression is as follows: in, This is the differential update form of the adaptive parameters.

Images