Intelligent control method of hypersonic vehicle driven by coupled information

Abstract

The present invention belongs to the technical field of hypersonic aircraft control, and relates to a coupling information driven hypersonic aircraft intelligent control method. First, a longitudinal dynamics model of an elastic hypersonic aircraft is established. Then, a coupling database based on sampling analysis is constructed, and the coupling degree matrix is ​​calculated by simulating different flight envelope data using a sampling statistical algorithm to form a coupling database to quantify the mutual influence between variables. Further, coupling data driven deep neural network training is carried out, and a long short-term memory network is adopted, and the coupling database is used as a data sample for offline training to enhance the autonomous fine control capability of the aircraft in a strong coupling environment. Finally, an online controller is designed to implement an intelligent compensation scheme for coupling information, and the aircraft maneuvering control is divided into a speed loop and an attitude loop, and the corresponding control instructions and neural network update laws are designed to achieve real-time and intelligent flight control.

Description

Technical Field

[0001] The present invention belongs to the technical field of hypersonic aircraft control, and relates to a coupled information driven hypersonic aircraft intelligent control method. Background Art

[0002] Hypersonic vehicles utilize a combination of propulsion systems to achieve high performance over a wide range. This leads to strong coupling and extremely complex characteristics between the airframe, dynamic modes, trajectory, guidance, and attitude control, posing significant challenges to control system design. On the one hand, fully accounting for coupling in vehicle modeling is difficult, and obtaining an accurate coupling model of the vehicle system is difficult. A comprehensive and in-depth understanding of the coupling mechanisms of hypersonic vehicles is therefore required. On the other hand, the vehicle system exhibits strong coupling and strong nonlinear characteristics, necessitating significant consideration of coupling effects during control. This requires further development of high-precision control under complex and uncertain conditions while meeting comprehensive performance requirements. Therefore, it is necessary to study the coupling characteristics of hypersonic vehicles under the stacking of various physical fields and design controllers that consider this coupling to improve flight quality.

[0003] Traditional control designs employ decoupling or ignore coupling, resulting in numerous cases of loss of control and failure, further demonstrating that these approaches cannot meet the performance requirements of the higher speed and wider envelope of hypersonic vehicles. Although numerous research results have been published on the design of hypersonic control systems, the impact of coupling analysis on control performance is often overlooked. Coupling analysis aims to reveal the dynamic relationships between stacked multi-physics fields and the laws governing their mutual influence. Therefore, it is necessary to design a hypersonic control system that fully utilizes coupling analysis, significantly reducing actuator control requirements and effectively improving control efficiency and the dynamic performance of the control system. This approach aims to advance hypersonic vehicle technology toward higher reliability and improved stability.

[0004] In "Hypersonic Vehicle Attitude Coordinated Control Method Based on Coupling Analysis" (CN 107085435 A), Wang Yuhui, Zhen Wubin, Ying Junqi, and others proposed a method for hypersonic vehicle attitude coordination based on coupling analysis, which effectively enhances the vehicle's controllability and maneuverability. However, compared to coupled data-driven analysis methods, this method assumes that the coupling relationship is a static constant throughout the flight mission. In reality, coupling characteristics are a dynamic process throughout the flight mission, thus failing to fully reveal the coupling relationship between variables, hindering the improvement of flight quality.

[0005] In "Hypersonic Vehicle Attitude Coupling Control Method" (CN110609564 A), Wang Zhi, Dong Wenqiang, Yang Ming, Gong Yulian, and Li Maomao proposed a coupled attitude control method for hypersonic aircraft, improving the rapidity of the vehicle's roll maneuver response. However, compared to coupled data-driven analysis methods, this method relies on a simplified model of kinematic coupling analysis, lacks a systematic analysis of attitude coupling, and fails to integrate coupling with controller design, potentially impacting control system performance.

[0006] Feng Xingkai, Wang Yuhui, Qin Xu, and others proposed a hypersonic vehicle coordinated control method based on dynamic coupling analysis in "Hypersonic Aircraft Attitude Coordinated Control Based on Dynamic Equations" (see Journal of Jilin University, March 2018, Vol. 36, No. 2, pp. 133-141), which comprehensively and objectively reflects the coupling situation under actual flight conditions. However, this method is relatively cumbersome in the design process. For complex systems, the dynamic equations may be very complex and difficult to establish and solve. Compared with the use of neural networks for The coupling database obtained by training with flight envelope data and the coupling matrix calculated by the dynamic equations in a single flight mission are insufficient to cope with the changes of hypersonic aircraft, and lack the real-time analysis and feedback of data-driven decision-making.

[0007] Shou Yingxin and Han Xudong proposed a channel-coupled coordinated robust adaptive control algorithm in "Channel-Coupled Coordinated Robust Adaptive Control Algorithm for Hypersonic Vehicles" (Science China, Vol. 54, No. 10, September 2024, pp. 2308-2325). Using composite learning and RISE techniques, they significantly improved the system's dynamic performance and steady-state error. However, compared to long short-term memory (LSTM) networks, this approach cannot better capture long-term dependencies in data or patterns and features in time series data. Furthermore, the control parameter tuning pressure is excessive, and the algorithm is highly dependent on the model.

[0008] Existing coupling analysis techniques mostly rely on a priori assumptions and pre-set theoretical models, making it difficult to comprehensively analyze the nonlinearities and dynamic coupling within a system. This results in inflexible and poorly adaptable control strategies in practical applications. Furthermore, coupling characteristic analysis has traditionally been an open-loop approach, making it difficult to integrate with closed-loop controllers. This can lead to unstable control systems and significant chattering in real-world operation, especially when dealing with uncertain systems. Current technological trends are driving data-driven and intelligent learning technologies (such as machine learning and artificial intelligence) to play a crucial role in improving system performance and adaptability. Long-term time-travel (LSTM) deep neural networks, driven by coupling data, can process large coupling databases, leverage gating mechanisms and inherent memory capabilities, and better capture long-term dependencies in the data, learn patterns and features, and organize potential coupling patterns. Furthermore, combining deep neural network coupling analysis with intelligent control enhances system robustness and ensures aircraft flight performance in various complex environments. Summary of the Invention

[0009] To address the control challenges of strongly coupled hypersonic vehicles, this paper proposes a coupling information-driven intelligent control method for hypersonic vehicles. First, a longitudinal dynamics model for an elastic hypersonic vehicle is established, integrating rigid body, aerodynamic, structural elasticity, and thrust coupling models to comprehensively describe the vehicle's dynamic characteristics. Next, a coupling database based on sampling analysis is constructed. By simulating data from different flight envelopes and calculating the coupling degree matrix using a sampling statistical algorithm, the coupling database is formed to quantify the mutual influence between variables. Furthermore, coupling data-driven deep neural network training is performed using a long short-term memory network (LSTM) and the coupling database as data samples for offline training to enhance the vehicle's autonomous and precise control capabilities in strongly coupled environments. Finally, an online controller is designed to implement an intelligent compensation scheme for coupling information. The vehicle's maneuvering control is divided into a velocity loop and an attitude loop, and corresponding control instructions and neural network update laws are designed to achieve real-time, intelligent flight control. This coupling information-driven intelligent control method for hypersonic vehicles not only improves the vehicle's control accuracy and stability, but also enhances its autonomous adaptability and survivability, providing strong technical support for the research, development, and application of hypersonic vehicles.

[0010] The technical solutions of the present invention are as follows:

[0011] A coupled information driven hypersonic vehicle intelligent control method is as follows:

[0012] Step (1) Construction of longitudinal dynamics model of elastic hypersonic vehicle

[0013] Based on the rigid body mathematical model, aerodynamic coupling model, structural elastic coupling model and thrust coupling model, a longitudinal dynamics model of elastic hypersonic vehicle is established:

[0014] (1)

[0015] in, is the hypersonic vehicle speed, is the flight altitude, is the track angle, is the pitch angle, is the pitch angular rate, is the angle of attack, and satisfies . is the mass of the aircraft, is the acceleration due to gravity, For the The moment of inertia of the shaft. represents the elastic mode, represents the damping factor, represents the natural oscillation frequency, represents the generalized force, denotes the first and second modes considered. are thrust, lift, and drag, respectively. is the pitching moment, and its specific expression is:

[0016] (2)

[0017] in, is the dynamic pressure, is the distance from the center of mass to the center of reference moment, is the wing reference area, The body coordinate axis system Axial force, is the mean aerodynamic chord length, is the elevator surface deflection angle; is the aileron rudder surface deflection angle; is the total lift coefficient, is the total thrust coefficient; is the total drag coefficient, and They represent the basic drag coefficient, the incremental drag coefficient caused by the elevator, and the incremental drag coefficient caused by the aileron; and They represent the basic lift coefficient, the incremental lift coefficient caused by the elevator, and the lift caused by the aileron; is the total pitching moment coefficient, and They represent the basic pitching moment coefficient, the incremental pitching moment coefficient caused by the elevator, and the incremental pitching moment coefficient caused by the aileron. is the incremental coefficient of the pitch moment caused by the pitch angular rate.

[0018] Total thrust coefficient The expression is as follows:

[0019] (3)

[0020] in, is the throttle opening, and Represent the basic resistance coefficient respectively. hour, , ;when hour, , ; The dynamic process of the engine is a second-order system:

[0021] (4)

[0022] in, It is the control signal of the throttle opening. is the natural oscillation frequency, It is the throttle opening that adjusts the damping.

[0023] Step (2) Construction of coupled database based on sampling analysis and dynamic equations

[0024] Based on the longitudinal dynamics model of the elastic hypersonic vehicle constructed in the previous step, this step first presents the coupled data analysis methods based on sampling analysis and dynamic equations, and then obtains a coupled database with sufficient samples based on these two methods.

[0025] 2.1) Provide the construction method of the sampling and analysis method database:

[0026] set up and are two mutually coupled variable groups in the system. Sampling and calculating the variables yields and ,use express right The impact quantification index is right The impact of can be expressed as

[0027] (5)

[0028] in, for right The degree of influence.

[0029] Similarly, right The influence degree can be expressed as

[0030] (6)

[0031] in, for right The degree of influence.

[0032] Furthermore, the variables in the system and The coupling degree can be expressed as

[0033] (7)

[0034] For the elastic hypersonic vehicle longitudinal dynamics model in step (1), define the state variable group and control input variable groups The coupling matrix is , trajectory variable group and posture variable group The coupling matrix is , posture variables and posture variable group The coupling matrix is , the specific expression is:

[0035] (8)

[0036] (9)

[0037] (10)

[0038] The flight position remains unchanged at the beginning and end, and the attack angle and flight speed instructions are changed to simulate the A set of flight envelope data is obtained, and then the coupling matrix is ​​calculated using sampling statistics algorithm in each flight envelope, so that Flight envelopes obtained Dimensional coupled database 、 、 .

[0039] 2.2) The construction method of the dynamic equation analysis method database is given:

[0040] Consider the following nonlinear system:

[0041] (11)

[0042] in, is the state variable, is the input variable.

[0043] Based on nonlinear theory, the analysis method of dynamic equations can be expressed as follows:

[0044] For the nonlinear system shown in Equation (11), the dynamic coupling matrix between the state variables is It can be defined as:

[0045] (12)

[0046] Similarly, variables With variables The dynamic coupling matrix between It can be defined as:

[0047] (13)

[0048] For the elastic hypersonic vehicle longitudinal dynamics model in step (1), define the state variable group and control input variable groups The coupling matrix is , trajectory variable group and posture variable group The coupling matrix is , posture variables and posture variable group The coupling matrix is , the specific expression is:

[0049] (14)

[0050] (15)

[0051] (16)

[0052] in, , is the state variable group Middle variables, , is the control input variable group Middle variables, . , is the trajectory variable group Middle variables, , is the posture variable group Middle variables, , , is the posture variable group Middle variables, .

[0053] The flight position remains unchanged at the beginning and end, and the attack angle and flight speed instructions are changed to simulate the Then, the coupling matrix is ​​calculated using the dynamic equation method in each flight envelope. Flight envelopes obtained Dimensional coupled database 、 、 .

[0054] 2.3) Hybrid database construction based on sampling analysis and dynamic equation coupling analysis

[0055] The coupled database is obtained by sampling analysis method and dynamic equation method, and then through screening and compensation, the final result is Total coupling database of dimensions 、 、 , providing sufficient training data for the next step (3) of deep neural network training.

[0056] Step (3) Deep neural network training based on coupled data drive

[0057] This step will use the offline data from the coupling database of different input variables used in the offline training to train the PID control framework as data samples. The offline neural network training process mainly includes the steps of dataset generation, training, verification, and testing.

[0058] 3.1) Dataset Generation: This step is to generate reliable and representative data that can cover the complex flight conditions and missions of hypersonic vehicles, that is, to generate the total coupled database in step (2). 、 、 , thereby training a neural network prior knowledge model with high accuracy and generalization. First, a coupling degree database must be collected and generated, covering a variety of flight conditions and missions, various cross-coupling interactions, model uncertainties, and the influence of multiple sources of disturbances. This data must simulate the performance of actual aircraft in various complex environments and missions as comprehensively as possible to ensure the adequacy of model training and its generalization capability. Next, simulation data generation is performed to simulate various possible flight conditions and disturbances, generating a large amount of training data. Specifically, first, by varying the aerodynamic force and torque model parameters, data containing model errors is generated to simulate model uncertainty. Second, by simulating different environmental conditions (such as wind speed and airflow variations), external disturbance data is generated to simulate environmental disturbances. Third, by simulating different flight missions, a generalized coupling database is generated to simulate complex flight missions and conditions. Finally, the dataset is divided into training, validation, and test sets with a certain distribution ratio for subsequent design verification.

[0059] 3.2) Training, Validation, and Testing: After the dataset is generated, it is used to train the model under the PID control framework. The neural network used for training is the Long Short-Term Memory (LSTM) network. First, the database in step (2) is divided into input data samples and target data samples. The input data samples are historical data used for model input, and the target data samples are future data to be predicted. During the training process, the model parameters, the parameters in the LSTM, and the control strategy are continuously adjusted to enable it to effectively cope with various flight missions, model uncertainties, and multi-source disturbances. Each LSTM has three gate structures to control the flow of gradients: the forget gate, the input gate, and the output gate.

[0060] The forget gate decides whether to retain or forget information by considering the input of the previous time step and the input of the current time step. Its calculation process is as follows:

[0061] (17)

[0062] Where, is the value of the forget gate, Indicates a connection operation. is the weight matrix of the forget gate, is the bias term of the forget gate, yes Input values ​​of the moment grid: The total coupling matrix , select the top 80% variable groups As input; Total coupling data of dimensional , select the top 80% variable groups As input; Total coupling data of dimensional , select the top 80% variable groups As input; is the output value of LSTM at the previous moment, The function is The activation function is as follows:

[0063] (18)

[0064] Input gate passes Activation function to calculate new memory unit candidate values , is the value of the input gate, quilt The processed value range is located in interval, and By multiplying each element, we can determine the candidate value of the memory unit For update , the calculation process is as follows:

[0065] (19)

[0066] (20)

[0067] Update the value of a memory cell:

[0068] (twenty one)

[0069] Where, and is the weight matrix, and is the bias term, is the unit state at the current moment, is the unit state at the previous moment, " means element-wise multiplication, The specific form of the function is as follows:

[0070] (twenty two)

[0071] Output gate usage The activation function determines the output. The activation function converts the value of the memory unit into Mapped to a range between -1 and 1. The calculation process is as follows:

[0072] (twenty three)

[0073] (twenty four)

[0074] Where, is the value of the output gate, is the output of the LSTM unit, is the weight matrix, is the bias matrix.

[0075] In the LSTM forward propagation, the original data is input into the neural network, and each neuron calculates an output value based on the input data and its own weight, that is, 、 、 、 、 The values ​​of the five vectors are processed by each layer of neurons and gradually passed to the next layer, and finally the prediction results of the neural network are obtained at the output layer.

[0076] The training algorithm of LSTM is the back propagation algorithm, which calculates the error term of each neuron The direction of back propagation of the LSTM error term is: the error term propagates to the upper layer as the error input of the LSTM unit in the upper layer. The gradient of each weight after the error is calculated based on the error to obtain the updated weight parameters to reduce the value of the loss function. Then, the weight and bias values ​​are updated according to the gradient of the loss function with respect to the weight and bias, so that the loss function gradually decreases. The gradient descent method is generally used to update the parameters:

[0077] (25)

[0078] (26)

[0079] Where, is the learning rate, is the loss function, 、 They are Moment and Moment The weights of the layers, including 、 、 and . 、 They are Moment and Moment Layer bias, including 、 、 and .

[0080] In this step, based on the effective data set, after training, verification and testing, the generated LSTM neural network covers the main features of the coupled data. The obtained LSTM neural network has the function of updating weights and biases and will be used as a coupled neural network in the design of subsequent intelligent controllers.

[0081] Step (4) Design of intelligent controller based on coupled deep neural network

[0082] This step makes full use of the coupled neural network generated by updating the weights and biases through LSTM in step (3) to carry out the design of the intelligent controller. The details are as follows:

[0083] First, the maneuver control model of the hypersonic vehicle is divided into a velocity loop and an attitude loop. The velocity loop is expressed as:

[0084] (27)

[0085] Where: and represents the transformed kinetic model coefficient, represents the coupling effect in the speed loop, represents the external disturbance of the speed loop, Represents the control input of the speed loop, i.e., the throttle opening. The speed reference signal is , and the corresponding tracking error is .

[0086] The pitch angle reference signal is defined as , the altitude of hypersonic vehicles By pitch angle Obtained by tracking the proportional-integral-derivative (PID) algorithm In the design, the left elevon rudder deflection angle and the right elevon rudder deflection angle are considered to be equal, that is, , the attitude loop is given by:

[0087] (28)

[0088] Where: and represents the transformed kinetic model coefficient, represents the coupling effect in the attitude loop, represents the external disturbance of the attitude loop, represents the control input of the attitude loop, i.e., the elevator angle, is the pitch angular velocity. The pitch angular velocity reference signal is , the tracking error of the pitch angular velocity is .

[0089] The design fuel equivalence ratio command and elevator control command are:

[0090] (29)

[0091] Where: , is the positive parameter to be designed, 、 、 and represents the nonlinear function approximated by the neural network, and express and The projection operator is used to avoid the occurrence of singular phenomena in the fuel equivalence ratio command and the elevator control command.

[0092] Among them, the neural network update law is designed as

[0093] (30)

[0094] Where: , , , , , , , , , , , , , , , is the positive parameter to be designed.

[0095] Beneficial effects of the present invention:

[0096] The present invention proposes a coupling information-driven intelligent control method for hypersonic aircraft, meeting the requirements for stable and robust performance faced by strongly coupled hypersonic aircraft flying across a wide speed range and airspace. Compared to traditional passive coupling suppression approaches, this method effectively resolves the conflict between complex uncertainty and high performance requirements. Firstly, it fully considers the difficulty in modeling coupled aircraft and obtaining an accurate coupling model for the aircraft system. By constructing a longitudinal dynamics model for an elastic hypersonic aircraft to represent the aircraft's equations of motion, a coupling database based on a sampling analysis algorithm and coupled data-driven deep neural network training are employed to gain a deeper understanding of the coupling mechanisms of hypersonic aircraft. Secondly, given the strong coupling and strong nonlinear characteristics of the aircraft system, an intelligent controller based on a coupled deep neural network is designed to further achieve high-precision control under complex and uncertain conditions while meeting comprehensive performance requirements. Therefore, the designed hypersonic aircraft control system fully utilizes coupling analysis to significantly reduce actuator control requirements and mitigate the reliance of traditional coupling analysis on a fixed number of sampling points. Using a neural network, it achieves rapid iterative coupling information feedback, learns patterns and features from the coupling database, and predicts the coupling matrix for any input state, effectively improving control efficiency and the dynamic performance of the control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 It is the overall flow chart of the coupled information driven hypersonic vehicle intelligent control method;

[0098] Figure 2 It is a framework for intelligent control methods of hypersonic vehicles driven by coupled information;

[0099] Figure 3 is the speed tracking curve of hypersonic vehicle;

[0100] Figure 4 is the position tracking curve of the hypersonic vehicle;

[0101] Figure 5 is the curve of hypersonic aircraft fuel command changing with time;

[0102] Figure 6 is the curve of the hypersonic aircraft rudder deflection command changing with time;

[0103] Figure 7 is the curve of the neural network weights ω1 and ω2 changing with time;

[0104] Figure 8 is the curve of the neural network weights ω3 and ω4 changing with time. DETAILED DESCRIPTION

[0105] The following further illustrates the embodiments of the present invention in conjunction with the accompanying drawings and technical solutions.

[0106] like Figure 1 and Figure 2 As shown in the figure, the present invention is a coupled information driven hypersonic aircraft intelligent control method. In this embodiment, the mass of the hypersonic aircraft is kg, and the power system uses a scramjet engine. The specific control process is as follows:

[0107] Step (1) Construction of longitudinal dynamics model of elastic hypersonic vehicle

[0108] Based on the rigid body mathematical model, aerodynamic coupling model, structural elastic coupling model and thrust coupling model, a longitudinal dynamics model of elastic hypersonic vehicle is established:

[0109] (1)

[0110] in, is the hypersonic vehicle speed, is the flight altitude, is the track angle, is the pitch angle, is the pitch angular rate, is the angle of attack, and satisfies . is the mass of the aircraft, is the acceleration due to gravity, For the The moment of inertia of the shaft. represents the elastic mode, represents the damping factor, represents the natural oscillation frequency, represents the generalized force, denotes the first and second modes considered. are thrust, lift, and drag, respectively. is the pitching moment, and its specific expression is:

[0111] (2)

[0112] in, is the dynamic pressure, is the distance from the center of mass to the center of reference moment, is the wing reference area, The body coordinate axis system Axial force, is the mean aerodynamic chord length, is the elevator surface deflection angle; is the aileron rudder surface deflection angle; is the total lift coefficient, is the total thrust coefficient; is the total drag coefficient, and They represent the basic drag coefficient, the incremental drag coefficient caused by the elevator, and the incremental drag coefficient caused by the aileron; and They represent the basic lift coefficient, the incremental lift coefficient caused by the elevator, and the lift caused by the aileron; is the total pitching moment coefficient, and They represent the basic pitching moment coefficient, the incremental pitching moment coefficient caused by the elevator, and the incremental pitching moment coefficient caused by the aileron. is the incremental coefficient of the pitch moment caused by the pitch angular rate.

[0113] Total thrust coefficient The expression is as follows:

[0114] (3)

[0115] in, is the throttle opening, and Represent the basic resistance coefficient respectively. hour, , ;when hour, , ; The dynamic process of the engine is a second-order system:

[0116] (4)

[0117] in, It is the control signal of the throttle opening. is the natural oscillation frequency, It is the throttle opening that adjusts the damping.

[0118] Step (2) Construction of coupling database based on sampling analysis algorithm

[0119] Based on the longitudinal dynamics model of the elastic hypersonic vehicle constructed in the previous step, this step first presents the coupled data analysis methods based on sampling analysis and dynamic equations, and then obtains a coupled database with sufficient samples based on these two methods.

[0120] 2.1) Provide a method for constructing a sampling and analysis method database

[0121] set up and are two mutually coupled variable groups in the system. Sampling and calculating the variables yields and ,use express right The impact quantification index is right The impact of can be expressed as

[0122] (5)

[0123] in, for right The degree of influence.

[0124] Similarly, right The influence degree can be expressed as

[0125] (6)

[0126] in, for right The degree of influence.

[0127] Furthermore, the variables in the system and The coupling degree can be expressed as

[0128] (7)

[0129] For the elastic hypersonic vehicle longitudinal dynamics model in step (1), define the state variable group and control input variable groups The coupling matrix is , trajectory variable group and posture variable group The coupling matrix is , posture variables and posture variable group The coupling matrix is , the specific expression is:

[0130] (8)

[0131] (9) (10)

[0132] The flight position remains unchanged at the beginning and end, and the attack angle and flight speed instructions are changed to simulate the Ten thousand sets of flight envelope data are then used to calculate the coupling matrix in each flight envelope using sampling statistics algorithm. Thousands of flight envelopes are obtained Dimensional coupled database 、 、 .

[0133] 2.2) The construction method of the dynamic equation analysis method database is given:

[0134] Consider the following nonlinear system:

[0135] (11)

[0136] in, is the state variable, is the input variable.

[0137] Based on nonlinear theory, the form of dynamic coupling is given as follows:

[0138] For the nonlinear system shown in Equation (11), the dynamic coupling matrix between the state variables is It can be defined as:

[0139] (12)

[0140] Similarly, variables With variables The dynamic coupling matrix between can be defined as:

[0141] (13)

[0142] For the elastic hypersonic vehicle longitudinal dynamics model in step (1), define the state variable group and control input variable groups The coupling matrix is , trajectory variable group and posture variable group The coupling matrix is , posture variables and posture variable group The coupling matrix is , the specific expression is:

[0143] (14)

[0144] (15)

[0145] (16)

[0146] in, , is the state variable group Middle variables, , is the control input variable group Middle variables, . , is the trajectory variable group Middle variables, , is the posture variable group Middle variables, , , is the posture variable group Middle variables, .

[0147] The flight position remains unchanged at the beginning and end, and the attack angle and flight speed instructions are changed to simulate the Ten thousand sets of flight envelope data are then used to calculate the coupling matrix in each flight envelope using the dynamic equation method. Thousands of flight envelopes are obtained Dimensional coupled database 、 、 .

[0148] 2.3) Hybrid database construction based on sampling analysis and dynamic equation coupling analysis

[0149] The coupled database is obtained by sampling analysis method and dynamic equation method, and then through screening and compensation, the final result is Total coupling database of dimensions 、 、 , providing sufficient training data for the next step (3) of deep neural network training.

[0150] Step (3) Deep neural network training based on coupled data drive

[0151] This step will use the offline data from the coupling database of different input variables used in the offline training to train the PID control framework as data samples. The offline neural network training process mainly includes the steps of dataset generation, training, verification, and testing.

[0152] 3.1) Dataset Generation: This step is to generate reliable and representative data that can cover the complex flight conditions and missions of hypersonic vehicles, that is, to generate the total coupled database in step (2). 、 、 , thereby training a neural network prior knowledge model with high accuracy and generalization. First, a coupling degree database must be collected and generated, covering a variety of flight conditions and missions, various cross-coupling interactions, model uncertainties, and the influence of multiple sources of disturbances. This data must simulate the performance of actual aircraft in various complex environments and missions as comprehensively as possible to ensure the adequacy of model training and its generalization capability. Next, simulation data generation is performed to simulate various possible flight conditions and disturbances, generating a large amount of training data. Specifically, first, by varying the aerodynamic force and torque model parameters, data containing model errors is generated to simulate model uncertainty. Second, by simulating different environmental conditions (such as wind speed and airflow variations), external disturbance data is generated to simulate environmental disturbances. Third, by simulating different flight missions, a generalized coupling database is generated to simulate complex flight missions and conditions. Finally, the dataset is divided into training, validation, and test sets with a certain distribution ratio for subsequent design verification.

[0153] 3.2) Training, verification and testing: After the data set is generated, the data is used for training under the PID control framework. The neural network used for training is the long short-term memory network (LSTM). First, the database in step (2) is divided into input data samples and target data samples. The input data samples are historical data used for model input, and the target data samples are future data to be predicted. During the training process, the model parameters, the parameters in the LSTM and the control strategy are continuously adjusted to enable it to effectively cope with various flight missions, model uncertainties and multi-source disturbances. Each LSTM has three gate structures to control the flow of gradients, namely: forget gate, input gate and output gate. The number of LSTM layers is selected to be 4, with 300 LSTM units per layer. The formula for LSTM forward propagation is as follows:

[0154] The forget gate decides whether to retain or forget information by considering the input of the previous time step and the input of the current time step. Its calculation process is as follows:

[0155] (17)

[0156] Where, is the value of the forget gate, Indicates a connection operation. is the weight matrix of the forget gate, is the bias term of the forget gate, yes Input values ​​of the moment grid: The total coupling matrix , before selecting variable groups As input; Total coupling data of dimensional , before selecting variable groups As input; Total coupling data of dimensional , before selecting variable groups As input; is the output value of LSTM at the previous moment, The function is The activation function is as follows:

[0157] (18)

[0158] Input gate passes Activation function to calculate new memory unit candidate values , is the value of the input gate, quilt The processed value range is located in interval, and By multiplying each element, we can determine the candidate value of the memory unit For update , the calculation process is as follows:

[0159] (19)

[0160] (20)

[0161] Update the value of a memory cell:

[0162] (twenty one)

[0163] Where, and is the weight matrix, and is the bias term, is the unit state at the current moment, is the unit state at the previous moment, " means element-wise multiplication, The specific form of the function is as follows:

[0164] (twenty two)

[0165] Output gate usage The activation function determines the output. The activation function converts the value of the memory unit into Mapped to a range between -1 and 1. The calculation process is as follows:

[0166] (twenty three)

[0167] (twenty four)

[0168] Where, is the value of the output gate, is the output of the LSTM unit, is the weight matrix, is the bias matrix.

[0169] In the LSTM forward propagation, the original data is input into the neural network, and each neuron calculates an output value based on the input data and its own weight, that is, 、 、 、 、 The values ​​of the five vectors are processed by each layer of neurons and gradually passed to the next layer, and finally the prediction results of the neural network are obtained at the output layer.

[0170] The training algorithm of LSTM is the back propagation algorithm, which calculates the error term of each neuron The direction of back propagation of the LSTM error term is: the error term propagates to the upper layer as the error input of the LSTM unit in the upper layer. The gradient of each weight after the error is calculated based on the error to obtain the updated weight parameters to reduce the value of the loss function. Then, the weight and bias values ​​are updated according to the gradient of the loss function with respect to the weight and bias, so that the loss function gradually decreases. The gradient descent method is generally used to update the parameters:

[0171] (25)

[0172] (26)

[0173] Where, is the learning rate, . is the loss function, 、 They are Moment and Moment The weights of the layers, including 、 、 and . 、 They are Moment and Moment Layer bias, including 、 、 and .

[0174] In this step, based on the effective data set, after training, verification and testing, the generated LSTM neural network covers the main features of the coupled data. The obtained LSTM neural network has the function of updating weights and biases and will be used as a coupled neural network in the design of subsequent intelligent controllers.

[0175] Step (4) Design of intelligent controller based on coupled deep neural network

[0176] This step makes full use of the coupled neural network generated by updating the weights and biases through LSTM in step (3) to carry out the design of the intelligent controller. The details are as follows:

[0177] First, the maneuver control model of the hypersonic vehicle is divided into a velocity loop and an attitude loop. The velocity loop is expressed as:

[0178] (27)

[0179] Where: and represents the transformed kinetic model coefficient, represents the coupling effect in the speed loop, represents the external disturbance of the speed loop, Represents the control input of the speed loop, i.e., the throttle opening. The speed reference signal is , and the corresponding tracking error is .

[0180] The pitch angle reference signal is defined as , the altitude of hypersonic vehicles By pitch angle Obtained by tracking the proportional-integral-derivative (PID) algorithm In the design, the left elevon rudder deflection angle and the right elevon rudder deflection angle are considered to be equal, that is, , the attitude loop is given by:

[0181] (28)

[0182] Where: and represents the transformed kinetic model coefficient, represents the coupling effect in the attitude loop, represents the external disturbance of the attitude loop, represents the control input to the attitude loop, i.e., the elevator angle. is the pitch angular velocity. The pitch angular velocity reference signal is , the tracking error of the pitch angular velocity is .

[0183] The design fuel equivalence ratio and command elevator control commands are:

[0184] (29)

[0185] Where: , , 、 、 and represents the nonlinear function approximated by the neural network, and express and The projection operator is used to avoid the occurrence of singular phenomena in the fuel equivalence ratio command and the elevator control command.

[0186] Among them, the neural network update law is designed as

[0187] (30)

[0188] In this embodiment: = 0.1, = 0.1, = 0.2, = 0.2, = 0.5, = 0.5, =0.4, = 0.4, = 0.1, = 0.1, = 0.1, = 0.1, = 1, = 2, = 1, = 10.

[0189] In order to verify this coupled information driven hypersonic aircraft intelligent control method, the simulation results are as follows Figure 3-8 shown. Figure 3 The speed tracking effect curve is given. The tracking signal rises synchronously with the speed command, indicating that the control system has an excellent tracking effect on the speed reference command. Figure 4The effect curve of altitude tracking is shown. In the first 60 seconds, there is a gap between the command and the tracking signal, and the position value of the tracking signal is higher than the command. After 60 seconds, the gap between the two gradually narrows until they almost completely overlap in the later stage, indicating that the altitude tracking signal gradually becomes consistent with the command. Figure 5 The fuel command changes are shown. It can be seen that the fuel command value is high at launch, indicating that the aircraft requires a large fuel supply to support high-speed flight. Starting around 20 seconds, the fuel command value gradually stabilizes, ultimately remaining at approximately 0.2. This indicates that after reaching a certain flight speed, the aircraft's fuel demand stabilizes to maintain sustained hypersonic flight. Figure 6 The graph shows the time-varying rudder angle command. Significant fluctuations in the first 20 seconds reflect the aircraft's adjustment process during takeoff. After 60 seconds, the rudder angle command remains relatively stable, indicating that the aircraft has entered a relatively stable flight state and no longer requires frequent rudder adjustments. Figures 7 to 8 The weight curves are shown in Figure 2. As the time approaches 0 seconds, each weight experiences a sharp decline. They then slowly recover and gradually enter a relatively stable phase. After 20 seconds, the neural network weights are essentially stable, and their fluctuations thereafter are very small. These simulations demonstrate the effectiveness of this coupled information-driven intelligent control method for hypersonic aircraft. It can improve the performance and stability of the control system, ensuring that hypersonic aircraft can achieve high-precision and high-reliability autonomous control in practical applications.

Claims

1. A method for intelligent control of a hypersonic vehicle driven by coupled information, characterized in that: The details are as follows: Step (1) Construction of longitudinal dynamics model of elastic hypersonic vehicle Based on the rigid body mathematical model, aerodynamic coupling model, structural elastic coupling model and thrust coupling model, the longitudinal dynamics model of elastic hypersonic vehicle is established: Where V is the speed of the hypersonic vehicle, h is the flight altitude, γ is the track angle, θ is the pitch angle, Q is the pitch rate, α is the angle of attack, and θ=γ+α is satisfied; m is the mass of the vehicle, g is the gravitational acceleration, I yy is the moment of inertia along the y-axis; η i represents the elastic mode, ζ i represents the damping factor, ω i represents the natural oscillation frequency, N i represents the generalized force, i=1,2 represents the first-order and second-order modes considered; T, L, D are thrust, lift, and drag respectively, M is the pitching moment, and the specific expression is: in, is the dynamic pressure, S is the wing reference area, C is the average aerodynamic chord length, X cg is the distance from the center of mass to the center of reference moment, Z is the force in the Z-axis direction of the body coordinate system, δ e is the left elevator rudder surface deflection angle; δ a is the right elevator rudder surface deflection angle; C L is the total lift coefficient, C T is the total thrust coefficient; C D is the total drag coefficient, and They represent the basic drag coefficient, the incremental drag coefficient caused by the left elevon and the incremental drag coefficient caused by the right elevon respectively; and They represent the basic lift coefficient, the lift increment coefficient caused by the elevator and the lift caused by the aileron; C m is the total pitching moment coefficient, and They represent the basic pitching moment coefficient, the incremental pitching moment coefficient caused by the left elevon rudder and the incremental pitching moment coefficient caused by the right elevon rudder, respectively. is the incremental coefficient of the pitch moment caused by the pitch angular rate; Total thrust coefficient C T The expression is as follows: Among them, λ is the throttle opening, and Represent the basic resistance coefficient respectively; when λ<1, When λ≥1, The dynamic process of the engine is a second-order system: Among them, λ c is the control signal of the throttle opening, ξ is the throttle opening adjustment damping, ω i is the natural oscillation frequency; Step (2) is based on the coupled database construction of sampling analysis and dynamic equations 2.1) The construction method of the sampling analysis method database is given: Let u j (j=1,2,…m) and x i (i=1,2…,n) are two mutually coupled variable groups in the system; sampling and calculating the variables obtains p(u j ) and p(x i ), use a ij Indicates u j x i The influence of quantitative indicators, then u j x i The influence is expressed as p(x i )=a i1 p(u1)+a i2 (u2)+…+a im p(u m ) (5) Among them, a ij for u j x i The degree of influence; Similarly, x i (i=1,2,…,n) for u j The influence degree of (j=1,2,…,m) can be expressed as p(u j )=b i1 p(x1)+b i2 p(x2)+…+b in p(x n ) (6) Among them, b ji For x i To u j The degree of influence; Furthermore, the variable u in the system j and x i The coupling degree can be expressed as For the elastic hypersonic vehicle longitudinal dynamics model in step (1), define the state variable group (V h γ θ αQ η i ) and the control input variable group (λ δ e δ a ) is η S , the coupling matrix of the trajectory variable group (h V) and the posture variable group (γ αQ) is χ S , the coupling matrix of the angle of attack α in the attitude variable and the attitude variable group (γ Q) is σ S , the specific expression is: The flight position remains unchanged at the beginning and end, and n sets of flight envelope data are simulated by changing the angle of attack command and the flight speed command. Then, the coupling degree matrix is ​​calculated in each flight envelope using the sampling statistical algorithm, so that the n flight envelopes can obtain a 7×3×n-dimensional coupling database η S , χ S , σ S ; 2.2) The construction method of the dynamic equation analysis method database is given: Consider the following nonlinear system: Among them, x is the state variable and u is the input variable; Based on nonlinear theory, the form of dynamic coupling is given as follows: For the nonlinear system shown in equation (11), the dynamic coupling matrix F between the state variables is x,x (x,u) can be defined as: Similarly, the dynamic coupling matrix F between variables x and u is x,u (x,u) is defined as: For the elastic hypersonic vehicle longitudinal dynamics model in step (1), define the state variable group (V h γ θ αQ η i ) and the control input variable group (λ δ e δ a ) is η D , the coupling matrix of the trajectory variable group (h V) and the posture variable group (γ αQ) is χ D , the coupling matrix of the angle of attack α in the attitude variable and the attitude variable group (γ Q) is σ D , the specific expression is: in, x i is the state variable set (V h γ θ α Q η i ) in which the i-th variable, i=1,2,…,7,u j is the control input variable group (λ δ e δ a ) in which the jth variable is j=1,2,3; x 1i is the i-th variable in the trajectory variable set (h V), i = 1, 2, x 2j is the jth variable in the attitude variable group (γαQ), j = 1, 2, 3, x 3j is the jth variable in the attitude variable group (γ Q), j = 1, 2; The flight position remains unchanged at the beginning and end. By changing the angle of attack command and the flight speed command, n sets of flight envelope data are simulated. Then, the coupling degree matrix is ​​calculated in each flight envelope using the dynamic equation method. Thus, a 7×3×n-dimensional coupling database η is obtained for n flight envelopes. D , χ D , σ D ; 2.3) Construction of hybrid database based on sampling analysis and dynamic equation coupling analysis The coupling database is obtained by sampling analysis method and dynamic equation method, and then after screening and compensation, the total coupling database η, χ, σ of 7×3×n dimensions is finally obtained, which provides sufficient training data for the deep neural network training in the next step (3); Step (3) Deep neural network training based on coupled data drive 3.1) Dataset generation: Generate the total coupling database η, χ, σ in step (2); 3.2) Training, verification and testing: After the data set is generated, the data is used for training under the PID control framework. First, the database in step (2) is divided into input data samples and target data samples. The input data samples are historical data used for model input, and the target data samples are future data to be predicted. During the training process, the model parameters, parameters in LSTM and control strategies are continuously adjusted to enable it to effectively cope with various flight missions, model uncertainties and multi-source disturbances. Each LSTM has three gate structures to control the flow of gradients, namely: forget gate, input gate and output gate. The forget gate decides whether to keep or forget information by considering the input of the previous time step and the input of the current time step. Its calculation process is as follows: f t =ι(W f ·[h t-1 ,x t ]+b f ) (17) In the formula, f t is the value of the forget gate, [,] represents the connection operation, W f is the weight matrix of the forget gate, b f is the bias term of the forget gate, x t is the input value of the grid at time t: the total coupling matrix η of 7×3×n dimensions, selecting the first 80% of n variable groups (V hγ θ α Q η i λ δ e δ a ) as input; the total coupled data χ of 7×3×n dimensions, the first 80% of n variable groups (h Vγ α Q) are selected as input; the total coupled data σ of 7×3×n dimensions, the first 80% of n variable groups (α γ Q) are selected as input; t-1 is the output value of LSTM at the previous moment, and the ι(x) function is the Sigmoid(x) activation function. The specific form is as follows: The input gate uses the tanh(x) activation function to calculate new memory cell candidate values. i t is the value of the input gate, i t After being processed by ι(x), the value range is in the interval [0,1], i t and By multiplying each element, we can determine the candidate value of the memory cell To update C t , the calculation process is as follows: i t =ι(W i ·[h t-1 ,x t ]+b i ) (20) Update the value of a memory cell: Where W C and W i is the weight matrix, b C and b i is the bias term, C t is the unit state at the current moment, C t-1 is the unit state at the previous moment, It means element-wise multiplication. The specific form of the tanh(x) function is as follows: The output gate uses the Sigmoid(x) activation function to determine the output, and the tanh(x) activation function converts the value of the memory cell C t is mapped to a range between -1 and 1; the calculation is as follows: o t =ι(W o ·[h t-1 ,x t ]+b o ) (23) In the formula, o t is the value of the output gate, h t is the output of the LSTM unit, W o is the weight matrix, b o is the bias matrix; In the LSTM forward propagation, the original data is input into the neural network, and each neuron calculates an output value based on the input data and its own weight, that is, f t 、i t , C t , o t 、h t The values ​​of the five vectors are processed by each layer of neurons and gradually passed to the next layer, and finally the prediction results of the neural network are obtained in the output layer; Step (4) Design of intelligent controller based on coupled deep neural network Using the coupled neural network generated by updating weights and biases through LSTM in step (3), the intelligent controller design is carried out as follows: Firstly, the maneuvering control model of the hypersonic vehicle is divided into a velocity loop and an attitude loop. The velocity loop is expressed as: Where: and represents the transformed kinetic model coefficient, represents the coupling effect in the speed loop, d V represents the external disturbance of the speed loop, λ represents the control input of the speed loop, i.e., the throttle opening; the speed reference signal is V d , the corresponding tracking error is e V =VV d ; Define the pitch angle reference signal as θ d , the height h of the hypersonic vehicle is obtained by the pitch angle θ through the PID algorithm d In the design, the left elevon rudder surface deflection angle is considered to be equal to the right elevon rudder surface deflection angle, that is, δ e =δ a , the attitude loop is given by: Where: and represents the transformed kinetic model coefficient, represents the coupling effect in the attitude loop, d Q represents the external disturbance of the attitude loop, δ e represents the control input of the attitude loop, i.e., the left elevator rudder deflection angle, Q is the pitch velocity; the pitch velocity reference signal is Q d , the tracking error of the pitch angular velocity is e Q =QQ d ; The design fuel equivalence ratio command and elevator control command are: Where: k V >0,k Q >0 is a positive parameter to be designed, and represents the nonlinear function approximated by the neural network, and express and The projection operator of Among them, the neural network update law is designed as In the formula: μ1>0, μ2>0, μ3>0, μ4>0, k1>0, k2>0, k3>0, k4>0, σ1>0, σ2>0, σ3>0, σ4>0, κ1>0, κ2>0, κ3>0, κ4>0 are the positive parameters to be designed.

2. The method for intelligent control of a hypersonic vehicle driven by coupled information according to claim 1, characterized in that: Step 3.1) includes: first, it is necessary to collect and generate a coupling degree database covering various flight conditions and flight missions, various cross-linked coupling effects, model uncertainties and multi-source disturbance effects; then, simulation data generation is performed to simulate various possible flight conditions and interference situations, and a large amount of training data is generated; specifically, first, by changing the aerodynamic force and torque model parameters, data containing model errors are generated to simulate model uncertainty; second, by simulating different environmental conditions, external interference data is generated to simulate environmental interference; third, by simulating different flight missions, a generalized coupling database is generated to simulate complex flight missions and conditions; finally, the data set is divided into training set, validation set and test set in a certain distribution ratio for subsequent design verification.

3. The method for intelligent control of a hypersonic vehicle driven by coupled information according to claim 1, characterized in that: In step 3.2), the LSTM training algorithm is the back propagation algorithm, which includes two directions: one is the back propagation along time, that is, starting from the current time t, calculating the error term at each moment; One is to propagate the error term to the upper layer; then, based on the corresponding error term, the gradient of each weight is calculated and the weight parameters of the model are updated to reduce the value of the loss function; Then, the values ​​of weights and biases are updated according to the gradient of the loss function to weights and biases, so that the loss function gradually decreases, and the gradient descent method is used to update the parameters: In the formula, is the learning rate, E is the loss function, W t l , are the weights of the lth layer at time t and time t+1, including W f , W C , W i and W o ; are the biases of the lth layer at time t and time t+1, including b f 、b C 、b i and b o .

Images