A cross-domain variable-configuration high-speed aircraft adaptive attitude control method, system and storage medium based on neural network identification

Abstract

The present invention relates to the field of aircraft control technology. The present invention provides a method, system, and storage medium for adaptive attitude control of a cross-domain, variable-configuration, high-speed aircraft based on neural network identification. The method addresses the problem that existing methods require unified parameterized modeling of the aircraft's aerodynamics and identification of the parameters in advance. This method fails to address the attitude control challenges posed by aerodynamic uncertainty when the aerodynamic parameter model of the cross-domain, variable-configuration, high-speed aircraft is completely unknown. The present invention constructs a three-channel neural network model, including a pitch channel neural network model, a yaw channel neural network model, and a roll channel neural network model. A loss function is constructed based on the weighted sensitivity of the influence of rudder error on attitude changes, and a three-channel rudder error solver neural network model is trained. An online transfer learning sample library is constructed for online transfer learning, and a backstepping controller is used to solve the aerodynamic torque, ultimately solving for the yaw command.

Description

Technical Field

[0001] The present invention relates to the field of aircraft control technology, and in particular to a method, system and storage medium for adaptive attitude control of a cross-domain variable-configuration high-speed aircraft based on neural network identification. Background Art

[0002] In recent years, the demand for repeatable earth-to-space transportation across large airspaces and speed ranges has grown rapidly. Cross-domain variable-configuration high-speed aircraft, with their ability to change their shape, can adjust their configuration according to different mission requirements and flight environments, achieving the optimal aerodynamic shape throughout the entire flight process. A key issue in the flight missions of wide-domain variable-configuration aircraft is how to achieve aircraft attitude control in a wide-domain, high-speed environment. However, while variable-configuration hypersonic aircraft have a larger flight envelope and stronger maneuverability, configuration changes also introduce significant uncertainty into their aerodynamic models, posing challenges to controller design. Therefore, it is very meaningful to develop adaptive control methods based on aerodynamic identification.

[0003] Currently, commonly used methods for aircraft aerodynamic parameter identification are mainly divided into two categories: model-driven and data-driven. Traditional aircraft parameter identification methods, such as the least squares method and Kalman filtering, require establishing an aerodynamic model of the aircraft. Based on appropriate initial values ​​for iteration, the predicted aerodynamic data based on the aerodynamic model is compared with the actual values ​​to iteratively optimize the aerodynamic parameters. Data-driven parameter identification methods simulate the input-output relationship of the aircraft's aerodynamics using a neural network and extract aerodynamic parameters from the neural network through backpropagation. These two methods differ only in their technical approach; in essence, they both require parameterized modeling of the aircraft's aerodynamics in advance, allowing for parameter identification using flight data. However, due to the multiple influences of cross-domain, high speed, and variable configuration, it is difficult to uniformly parameterize the aerodynamic model of an aircraft. Therefore, there is an urgent need for an adaptive attitude control method for cross-domain, variable configuration, high-speed aircraft when the aerodynamic parameter model is completely unknown. Summary of the Invention

[0004] The technical problems to be solved by the present invention are:

[0005] Existing methods require unified parametric modeling of the aircraft aerodynamics in advance and identification of the parameters; they are unable to solve the attitude control problem caused by aerodynamic uncertainty when the aerodynamic parameter model of a cross-domain variable-configuration high-speed aircraft is completely unknown.

[0006] The present invention is to solve the above technical problems using the following technical solutions:

[0007] The present invention provides a method for adaptive attitude control of a cross-domain variable-configuration high-speed aircraft based on neural network identification, comprising the following steps:

[0008] Step 1: Collect wind tunnel test and historical flight data of the cross-domain variable configuration high-speed aircraft, segment the data, and construct an offline training sample library for the three-channel aerodynamic model of the cross-domain variable configuration high-speed aircraft, including: a pitch channel sample library, a yaw channel sample library, and a roll channel sample library; preprocess the data and divide the preprocessed data into a data set, a validation set, and a test set;

[0009] Step 2: Construct a three-channel neural network model, including a pitch channel neural network model, a yaw channel neural network model, and a roll channel neural network model. Use the training set to train the three channel neural network models separately, construct a loss function based on the sensitivity of the influence of the rudder error on the attitude change, and train the neural network until the loss function converges. During the training process, use the validation set to verify the model performance, and obtain the three-channel rudder solver neural network model.

[0010] Step 3: During the online flight, flight data is collected and the online transfer learning sample library of the three-channel aerodynamic model of the cross-domain variable-configuration high-speed aircraft is constructed using the aircraft attitude rotation dynamics equation;

[0011] Step 4: Using the three-channel rudder deflection solver neural network model, based on the online transfer learning sample library of the cross-domain variable-configuration high-speed aircraft three-channel aerodynamic model constructed in step 3, online transfer learning is performed at intervals of T to obtain a three-channel adaptive rudder deflection solver neural network model;

[0012] Step 5: Establish the aircraft attitude rotation dynamics equation, design a backstepping controller to obtain the aerodynamic torque, further reversely solve the torque coefficient, input the torque coefficient into the three-channel adaptive rudder yaw solver neural network model, and solve to obtain the yaw command.

[0013] Furthermore, the wind tunnel test and historical flight data of the cross-domain variable configuration high-speed aircraft in step 1 include: flight state data and aerodynamic moment coefficients, and the flight state data include angle of attack α, sideslip angle β, roll angle σ, pitch rudder deflection angle δ z , yaw rudder deflection angle δ y 、Roll rudder deflection angle δ x and the wing fold angle δ f ; Aerodynamic moment coefficients include: pitch moment coefficient C mz , yaw moment coefficient C my and the rolling moment coefficient C mx ;

[0014] The data is segmented to construct an offline training sample library of a three-channel aerodynamic model for a cross-domain variable-configuration high-speed aircraft, including: a pitch channel sample library [α δ f C mz δ z ], yaw channel sample library [β α δf C my δ y ] and the rolling channel sample library [βα δ f C mx δ x ].

[0015] Furthermore, the data preprocessing in step 1 includes normalizing the data and cleaning abnormal or erroneous data.

[0016] Furthermore, the three-channel neural network model in step 2 includes:

[0017] The pitch channel neural network model:

[0018] Input layer: The number of input layer nodes is 3;

[0019] Hidden layer: The hidden layer consists of 3 layers, and the neurons use Tanh function;

[0020] Output layer: The number of nodes in the output layer is 1, and the linear activation function is used;

[0021] The yaw channel neural network model and the roll channel neural network model both include:

[0022] Input layer: The number of input layer nodes is 4;

[0023] Hidden layer: The hidden layer of the yaw channel and the roll channel consists of 3 layers, and the neurons use the Tanh function;

[0024] Output layer: The number of nodes in the output layer is 1, and a linear activation function is used.

[0025] Furthermore, the construction process of the loss function in step 2 is:

[0026] The intervals are divided according to the different effects of the rudder error on the attitude change, and the sensitive intervals are assigned higher weights in the loss function.

[0027] The loss function uses the mean square error (MSE):

[0028]

[0029] Among them, y i is the true value, is the predicted value, n is the number of samples;

[0030] The constructed loss function is:

[0031] L=L1+k×L2

[0032] Wherein, L1 is the mean square error between the true value and the predicted value of the steering deflection corresponding to the non-important interval, L2 is the mean square error between the true value and the predicted value of the steering deflection corresponding to the important interval, and k is the importance weighting factor.

[0033] Furthermore, step 3 includes the following steps:

[0034] First, collect online data including the attitude angular velocity ω of the three channels of the aircraft x 、ω y 、ω z , the aircraft's angle of attack α, sideslip angle β, bank angle σ, dynamic pressure Q, and pitch rudder angle δ z , yaw rudder deviation δ y and the roll rudder bias δ x , folding angle δ f and the characteristic area S, characteristic lengths b and c of the aircraft;

[0035] By analyzing the attitude angular velocity ω of the three channels x 、ω y 、ω z Perform differential calculation to obtain attitude angular acceleration The true torque of the three channels is obtained by inversely solving the aircraft attitude rotation dynamics equation:

[0036]

[0037] Among them, M tx 、M ty 、M tz are the true moments of the roll channel, yaw channel, and pitch channel, respectively; I1~I9 are the moment of inertia parameters of the aircraft;

[0038] Furthermore, the true moment coefficients of the three channels are solved by inversely solving the definition of the aircraft aerodynamic torque:

[0039]

[0040] After data segmentation, we can obtain the online transfer learning sample library for the three-channel aerodynamic model of cross-domain variable-configuration high-speed aircraft, including the pitch channel sample library [α δ f C mz δ z ], yaw channel sample library [β α δ f C my δ y ] and the rolling channel sample library [β α δ f C mx δ x ].

[0041] Furthermore, step 5 includes the following steps:

[0042] First, the dynamic equation of the aircraft attitude rotation is established as:

[0043]

[0044] Simplifying it to:

[0045]

[0046] Among them, x1=[α β σ] T ,x2=[ω x ω y ω z ] T , the control input is u=[M x M y M z ];

[0047] Furthermore, the outer loop controller is set up, and the specific steps are as follows:

[0048] For the attitude angle equation:

[0049]

[0050] The angular tracking error is defined as:

[0051] e1=x1-x c

[0052] The angle tracking error change rate is:

[0053]

[0054] x 2c Considered as a virtual control quantity, select

[0055]

[0056] Where k1 is a diagonal positive definite matrix, and its form is: k1=diag(k 11 ,k 12 ,k 13 ), x 2c is the desired angular velocity;

[0057] Furthermore, the inner loop controller is set up, and the specific steps are as follows:

[0058] For the attitude angular velocity equation:

[0059]

[0060] The angular velocity tracking error is defined as:

[0061] e2=x2-x2c

[0062] The angular velocity tracking error change rate is:

[0063]

[0064] The final output u of the backstepping controller is:

[0065]

[0066] Where k2 is a diagonal positive definite matrix, and its form is: k2=diag(k 21 ,k 22 ,k 23 );

[0067] The final output u obtained by the backstepping controller is the required aerodynamic torque [M tx M ty M tz ], the required torque coefficients of the three channels are solved by inversely solving the definition of the aircraft aerodynamic torque:

[0068]

[0069] After normalizing the torque coefficient and related state quantities required at the current moment, they are input into the three-channel adaptive rudder deviation solver neural network model completed by online migration in step 4 to obtain the rudder deviation command.

[0070] The present invention provides an adaptive attitude control system for a cross-domain variable-configuration high-speed aircraft based on neural network identification. The system has a program module corresponding to the steps of the method described in any one of the above technical solutions, and executes the steps in the above-mentioned adaptive attitude control method for a cross-domain variable-configuration high-speed aircraft based on neural network identification during operation.

[0071] The present invention provides a computer-readable storage medium storing a computer program, wherein the computer program is configured to implement the steps of the cross-domain variable-configuration high-speed aircraft adaptive attitude control method based on neural network identification described in any one of the above technical solutions when called by a processor.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] The present invention proposes an adaptive control method based on neural network aerodynamic identification. Through the neural network "offline training + online transfer learning", an adaptive rudder deflection solver is obtained, which is introduced into the backstepping control loop to achieve adaptive attitude control. Compared with existing methods, it has the following advantages: 1) When the aerodynamic parameter model is completely unknown, the mapping relationship from aerodynamic torque to rudder deflection can be directly learned through the neural network to achieve aerodynamic model identification; 2) By performing a rudder deflection sensitivity analysis in advance, an importance weighting strategy is introduced to construct a loss function to achieve high-precision fitting; 3) Through the learning strategy of "offline training + online transfer learning", offline and online data are fully utilized, significantly reducing online training time. 4) By changing the output of the traditional controller from rudder deflection to torque, a neural network rudder deflection solver is introduced to achieve attitude tracking control in non-nominal environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 This is a flow chart of the cross-domain variable configuration high-speed aircraft adaptive attitude control method based on neural network identification in an embodiment of the present invention;

[0075] Figure 2 Schematic diagram of a pitch channel neural network in an embodiment of the present invention;

[0076] Figure 3 This is a flow chart of the online transfer learning sample library collection process in an embodiment of the present invention;

[0077] Figure 4 1 is a control loop diagram of a neural network-based adaptive posture controller in an embodiment of the present invention;

[0078] Figure 5 is a neural network training loss curve in an embodiment of the present invention;

[0079] Figure 6 This is a neural network steering command fitting curve diagram in an embodiment of the present invention;

[0080] Figure 7 Graphs showing attitude angle command tracking curves, angular velocity, and rudder deflection curves in an embodiment of the present invention. DETAILED DESCRIPTION

[0081] In order to enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below with reference to the accompanying drawings. Obviously, the described embodiments or examples are only some of the embodiments or examples of the present invention, and not all of them. Based on the embodiments or examples of the present invention, all other embodiments or examples obtained by those skilled in the art without creative work should fall within the scope of protection of the present invention.

[0082] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0083] Specific implementation plan 1: Figure 1 As shown, the present invention provides a cross-domain variable configuration high-speed aircraft adaptive attitude control method based on neural network identification, comprising the following steps:

[0084] Step 1: Collect wind tunnel test and historical flight data of the cross-domain variable configuration high-speed aircraft, segment the data, and construct an offline training sample library for the three-channel aerodynamic model of the cross-domain variable configuration high-speed aircraft, including: a pitch channel sample library, a yaw channel sample library, and a roll channel sample library; preprocess the data and divide the preprocessed data into a data set, a validation set, and a test set;

[0085] Step 2: Construct a three-channel neural network model, including a pitch channel neural network model, a yaw channel neural network model, and a roll channel neural network model. Use the training set to train the three channel neural network models separately, construct a loss function based on the sensitivity of the influence of the rudder error on the attitude change, and train the neural network until the loss function converges. During the training process, use the validation set to verify the model performance, and obtain the three-channel rudder solver neural network model.

[0086] Step 3: During the online flight, flight data is collected and the online transfer learning sample library of the three-channel aerodynamic model of the cross-domain variable-configuration high-speed aircraft is constructed using the aircraft attitude rotation dynamics equation;

[0087] Step 4: Using the three-channel rudder deflection solver neural network model, based on the online transfer learning sample library of the cross-domain variable-configuration high-speed aircraft three-channel aerodynamic model constructed in step 3, online transfer learning is performed at intervals of T to obtain a three-channel adaptive rudder deflection solver neural network model;

[0088] Step 5: Establish the aircraft attitude rotation dynamics equation, design a backstepping controller to obtain the aerodynamic torque, further reversely solve the torque coefficient, input the torque coefficient into the three-channel adaptive rudder yaw solver neural network model, and solve to obtain the yaw command.

[0089] Specific implementation plan 2: The wind tunnel test and historical flight data of the cross-domain variable configuration high-speed aircraft described in step 1, including: flight status data and aerodynamic torque coefficients, flight status data including angle of attack α, sideslip angle β, roll angle σ, pitch rudder deflection angle δ z , yaw rudder deflection angle δ y 、Roll rudder deflection angle δ x and the wing fold angle δ f ; Aerodynamic moment coefficients include: pitch moment coefficient C mz , yaw moment coefficient Cmy and the rolling moment coefficient C mx ;

[0090] The data is segmented to construct an offline training sample library of a three-channel aerodynamic model for a cross-domain variable-configuration high-speed aircraft, including: a pitch channel sample library [α δ f C mz δ z ], yaw channel sample library [β α δ f C my δ y ] and the rolling channel sample library [βα δ f C mx δ x ]. The rest of this embodiment is the same as the specific embodiment 1.

[0091] Specific implementation plan three: The data preprocessing described in step 1 includes normalizing the data, dividing all data by the absolute value of the maximum value, thereby scaling the data range to a fixed interval [-1, 1]. This helps improve the training efficiency of the neural network and prevents slow model convergence or gradient explosion / vanishing caused by inconsistent data dimensions. The normalization formula is as follows:

[0092]

[0093] Since there may be sensor failures in the wind tunnel data and aircraft historical data collection process, abnormal or erroneous data is cleaned to ensure data quality. The rest of this implementation plan is the same as the specific implementation plan 2.

[0094] Specific implementation plan four: Figure 2 As shown, the three-channel neural network model described in step 2 includes:

[0095] The pitch channel neural network model:

[0096] Input layer: The number of input layer nodes is 3, including the normalized attack angle α and folding angle δ f and the pitching moment coefficient C mz ;

[0097] Hidden layer: The hidden layer consists of three layers, with 264, 128, and 64 neurons in each layer, respectively. Since the input of the neural network is restricted to the interval [-1, 1] during the normalization process, the neurons use the Tanh function, and their output range is exactly between -1 and 1. This is consistent with the input and output data range after normalization, which can reduce numerical offsets during model training and promote gradient propagation and optimization.

[0098] Output layer: The number of nodes in the output layer is 1, which represents the aircraft at the angle of attack α and the folding angle δ. f In order to provide the pitch moment coefficient C mz The corresponding pitch rudder deflection δ z , the output layer uses a linear activation function;

[0099] The yaw channel neural network model and the roll channel neural network model both include:

[0100] Input layer: The number of input layer nodes is 4. The input of the yaw channel includes the normalized sideslip angle β, attack angle α, and folding angle δ. f and the yaw moment coefficient C my The input of the roll channel includes the normalized sideslip angle β, attack angle α, and fold angle δ f and the rolling moment coefficient C mx ;

[0101] Hidden layer: The hidden layer of the yaw channel and the roll channel consists of 3 layers, and the neurons use the Tanh function;

[0102] Output layer: The number of nodes in the output layer is 1. The output layer of the yaw channel represents the sideslip angle β, attack angle α, and folding angle δ of the aircraft. f In order to provide the yaw moment coefficient C my The corresponding yaw rudder deviation δ y The output layer of the roll channel represents the aircraft's sideslip angle β, attack angle α, and fold angle δ. f In order to provide the rolling moment coefficient C mx The corresponding roll rudder deflection δ x , the output layer uses a linear activation function. The rest of this implementation is the same as the specific implementation plan three.

[0103] Specific implementation plan 5: The construction process of step 2 loss function is:

[0104] During the attitude tracking process of an aircraft, the impact of rudder error on attitude changes varies significantly across different intervals. The impact of rudder error on attitude changes varies across different stages of attitude change, and its influence on attitude control is particularly significant within certain rudder error ranges. To further improve the training accuracy of neural networks, it is necessary to analyze these intervals. By assigning higher weights to intervals that are more sensitive to rudder error, the network can better fit these intervals, thereby improving overall control performance.

[0105] First, we analyze the sensitivity of rudder error. During the attitude tracking process, the impact of rudder error varies depending on the vehicle's state and control strategy. During the attitude change phase, the rudder deflection range is large (e.g., -20 to 20 degrees). In this case, a 0.1-degree rudder error has little impact on attitude change. However, during the attitude hold phase, the rudder deflection range is typically smaller. For example, when the pitch channel's angle of attack is tracking a constant angle of attack between 0 and 10 degrees, the rudder error range is only approximately -1 to -3 degrees. In this case, a 0.1-degree rudder error can result in an angle-of-attack tracking error of approximately 0.5 degrees, a significant impact. Therefore, we divide the ranges based on the different impacts of the rudder error on attitude change, and assign higher weights to sensitive intervals in the loss function.

[0106] Furthermore, based on the sensitivity analysis above, a loss function is constructed based on the importance-weighted concept to ensure improved fitting accuracy in the sensitive range of rudder error. The loss function uses the mean squared error (MSE). The loss function is defined as the mean of the squared differences between the predicted and true values, and the optimization goal is to minimize this loss function.

[0107]

[0108] Among them, y i is the true value, is the predicted value, n is the number of samples;

[0109] The constructed loss function is:

[0110] L=L1+k×L2

[0111] Where L1 is the mean square error between the true value and the predicted value of the steering deflection corresponding to the non-important interval ([-20°, 20°]), L2 is the mean square error between the true value and the predicted value of the steering deflection corresponding to the important interval, and k is the importance weighting factor.

[0112] The important interval of the pitch channel neural network loss function is [-4°, 0°], and the important interval of the yaw channel and roll channel neural network loss function is [-2°, 2°]. The importance weighting factor k is 10.

[0113] Furthermore, a gradient descent method (such as the Adam optimizer) is used to minimize the loss function. The Adam optimizer can adaptively adjust the learning rate, which generally accelerates convergence and works well with large-scale data. Training is terminated when the loss function no longer decreases significantly.

[0114] The entire training process is:

[0115] Use the training set data for forward propagation to calculate the steering deviation prediction values ​​of the three channels respectively.

[0116] The error between the predicted rudder deviation value and the true rudder deviation value is calculated using the importance-weighted loss function.

[0117] The gradient is calculated using the back-propagation algorithm, and the weights of the three-channel neural network are updated separately through the optimization algorithm.

[0118] After each round of iteration, the validation set is used to evaluate the generalization ability of the model to avoid overfitting.

[0119] Early stopping is used to avoid overfitting during training. If the validation set error does not significantly improve within a certain number of iterations, training is terminated early, resulting in a three-channel rudder deviation solver neural network model that has been trained offline. This implementation plan is otherwise identical to Specific Implementation Plan 4.

[0120] Specific implementation plan six: Figure 3 As shown, the specific steps of step 3 are as follows:

[0121] During the online flight, data such as aircraft speed, attitude change angular velocity, attitude angle, rudder deflection, and deformation are collected, and the aircraft attitude rotation dynamics are used to construct an online transfer learning sample library for the three-channel aerodynamic model of a cross-domain variable-configuration high-speed aircraft.

[0122] First, collect online data including the attitude angular velocity ω of the three channels of the aircraft x 、ω y 、ω z , the aircraft's angle of attack α, sideslip angle β, bank angle σ, dynamic pressure Q, and pitch rudder angle δ z , yaw rudder deviation δ y and the roll rudder bias δ x , folding angle δ f and the characteristic area S, characteristic lengths b and c of the aircraft;

[0123] By analyzing the attitude angular velocity ω of the three channels x 、ω y 、ω z Perform differential calculation to obtain attitude angular acceleration The true torque of the three channels is obtained by inversely solving the aircraft attitude rotation dynamics equation:

[0124]

[0125] Among them, M tx 、M ty 、M tz are the true moments of the roll channel, yaw channel, and pitch channel, respectively; I1~I9 are the moment of inertia parameters of the aircraft;

[0126] Furthermore, the true moment coefficients of the three channels are solved by inversely solving the definition of the aircraft aerodynamic torque:

[0127]

[0128] After data segmentation, we can obtain the online transfer learning sample library for the three-channel aerodynamic model of cross-domain variable-configuration high-speed aircraft, including the pitch channel sample library [α δ f C mz δ z ], yaw channel sample library [β α δ f C my δ y ] and the rolling channel sample library [β α δ f C mx δ x ]. The rest of this implementation plan is the same as the specific implementation plan five.

[0129] Specific implementation plan seven: Figure 4 As shown, step 5 includes the following steps:

[0130] First, the dynamic equation of the aircraft attitude rotation is established as:

[0131]

[0132] Simplifying it to:

[0133]

[0134] Among them, x1=[α β σ] T ,x2=[ω x ω y ω z ] T , the control input is u=[M x M y M z ];

[0135] Furthermore, the outer loop controller is set up, and the specific steps are as follows:

[0136] For the attitude angle equation:

[0137]

[0138] The angular tracking error is defined as:

[0139] e1=x1-x c

[0140] The angle tracking error change rate is:

[0141]

[0142] x 2c Considered as a virtual control quantity, select

[0143]

[0144] Where k1 is a diagonal positive definite matrix, and its form is: k1=diag(k 11 ,k 12 ,k 13 ), x 2c is the desired angular velocity;

[0145] Furthermore, the inner loop controller is set up, and the specific steps are as follows:

[0146] For the attitude angular velocity equation:

[0147]

[0148] The angular velocity tracking error is defined as:

[0149] e2=x2-x2c

[0150] The angular velocity tracking error change rate is:

[0151]

[0152] The final output u of the backstepping controller is:

[0153]

[0154] Where k2 is a diagonal positive definite matrix, and its form is: k2=diag(k 21 ,k 22 ,k 23 );

[0155] The final output u obtained by the backstepping controller is the required aerodynamic torque [M tx M ty M tz ], the required torque coefficients of the three channels are solved by inversely solving the definition of the aircraft aerodynamic torque:

[0156]

[0157] After normalizing the torque coefficient and related state variables required at the current moment, they are input into the three-channel adaptive rudder deflection solver neural network model completed in the online migration in step 4 according to a fixed format to obtain the rudder deflection command.

[0158] The present invention proposes a cross-domain variable-configuration high-speed aircraft adaptive attitude control method (algorithm) based on neural network identification, which is the underlying technical core of the present invention. Various products can be derived based on the algorithm.

[0159] Based on the method proposed in the present invention, a cross-domain variable-configuration high-speed aircraft adaptive attitude control system based on neural network identification is developed using a programming language. The system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned cross-domain variable-configuration high-speed aircraft adaptive attitude control method based on neural network identification during operation.

[0160] The developed system (software) computer program is stored on a computer-readable storage medium. The computer program is configured to implement the steps of the above-mentioned method for adaptive attitude control of a cross-domain, variable-configuration, high-speed aircraft based on neural network identification when called by a processor. This materializes the present invention on a carrier, becoming a computer program product.

[0161] Various implementations of the systems and techniques described herein can be realized in digital electronic circuitry, integrated circuitry, dedicated ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system comprising at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0162] The computer programs (also referred to as programs, software, software applications, or code) of the present invention include machine instructions for a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., a magnetic disk, an optical disk, a memory, a programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0163] The following examples are used to verify the beneficial effects of the present invention.

[0164] Example 1

[0165] This embodiment is based on the method of the above-mentioned implementation scheme, and uses a digital model of a foldable cross-domain variable-configuration high-speed aircraft for simulation verification.

[0166] In this embodiment, a flight experiment is first conducted using an aircraft under nominal conditions. The collected data is treated as wind tunnel data and historical flight data, thereby constructing an offline training sample library for a three-channel aerodynamic model of a cross-domain variable-configuration high-speed aircraft. The three-channel neural network model is constructed using the method described in step 2 above. The loss function is constructed using the importance weighting concept. The neural network is trained until the loss function converges. The convergence graph of the loss function is shown in the figure below. Figure 5 , we get the three-channel rudder deviation solver after offline training, and the rudder deviation command fitting curve is as follows: Figure 6 . Then, before the start of the online simulation, the aerodynamic parameters of the aircraft are randomly deflected by 25% to simulate the uncertainty brought to the aerodynamic model due to wide range, high speed and configuration changes. Then, the working conditions of flight altitude of 40km, flight speed of 3000km / s, initial angle of attack of 0° and tracking angle of attack instruction of 10° are selected, and the simulation is carried out under backstepping control without aerodynamic identification. At the same time, online flight data is collected to perform online transfer learning on the basis of the three-channel rudder deviation solver completed by offline training, and a more accurate rudder deviation solver is obtained. It is introduced into the backstepping control loop at the 3rd second for attitude control. From the simulation results Figure 7 It can be seen that the adaptive attitude control method of the present invention can enable a cross-domain variable-configuration high-speed aircraft to achieve attitude tracking control with aerodynamic uncertainty when the aerodynamic model is completely unknown.

[0167] Although the present invention is disclosed as above, the scope of protection disclosed by the present invention is not limited thereto. Those skilled in the art of the present invention may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A cross-domain variable-configuration high-speed aircraft adaptive attitude control method based on neural network identification, characterized in that: The following steps are involved: Step 1: Collect wind tunnel test and historical flight data of the cross-domain variable configuration high-speed aircraft, segment the data, and construct an offline training sample library for the three-channel aerodynamic model of the cross-domain variable configuration high-speed aircraft, including: a pitch channel sample library, a yaw channel sample library, and a roll channel sample library; preprocess the data and divide the preprocessed data into a data set, a validation set, and a test set; Step 2: Construct a three-channel neural network model, including a pitch channel neural network model, a yaw channel neural network model, and a roll channel neural network model. Use the training set to train the three channel neural network models separately, construct a loss function based on the sensitivity of the influence of the rudder error on the attitude change, and train the neural network until the loss function converges. During the training process, use the validation set to verify the model performance, and obtain the three-channel rudder solver neural network model. Step 3: During the online flight, flight data is collected and the online transfer learning sample library of the three-channel aerodynamic model of the cross-domain variable-configuration high-speed aircraft is constructed using the aircraft attitude rotation dynamics equation; Step 4: Using the three-channel rudder deflection solver neural network model, based on the online transfer learning sample library of the cross-domain variable-configuration high-speed aircraft three-channel aerodynamic model constructed in step 3, online transfer learning is performed at intervals of T to obtain a three-channel adaptive rudder deflection solver neural network model; Step 5: Establish the aircraft attitude rotation dynamics equation, design a backstepping controller to obtain the aerodynamic torque, further reversely solve the torque coefficient, input the torque coefficient into the three-channel adaptive rudder yaw solver neural network model, and solve to obtain the yaw command; The construction process of step 2 loss function is: Divide the intervals according to the different effects of the rudder error on the attitude change, and assign higher weights to the sensitive intervals in the loss function; The loss function uses the mean square error (MSE): Among them, y i is the true value, is the predicted value, n is the number of samples; The constructed loss function is: L=L1+k×L2 Wherein, L1 is the mean square error between the true value and the predicted value of the steering deflection corresponding to the non-important interval, L2 is the mean square error between the true value and the predicted value of the steering deflection corresponding to the important interval, and k is the importance weighting factor.

2. The cross-domain variable configuration high-speed aircraft adaptive attitude control method based on neural network identification according to claim 1 is characterized in that: The wind tunnel test and historical flight data of the cross-domain variable configuration high-speed aircraft described in step 1, including: flight state data and aerodynamic torque coefficients. The flight state data include angle of attack α, sideslip angle β, roll angle σ, pitch rudder deflection angle δ z , yaw rudder deflection angle δ y 、Roll rudder deflection angle δ x and the wing fold angle δ f ; Aerodynamic moment coefficients include: pitch moment coefficient C mz , yaw moment coefficient C my and the rolling moment coefficient C mx ; The data is segmented to construct an offline training sample library of a three-channel aerodynamic model for a cross-domain variable-configuration high-speed aircraft, including: a pitch channel sample library [α δ f C mz δ z ], yaw channel sample library [β α δ f C my δ y ] and the rolling channel sample library [β α δ f C mx δ x ].

3. The cross-domain variable configuration high-speed aircraft adaptive attitude control method based on neural network identification according to claim 2 is characterized in that: The data preprocessing in step 1 includes normalizing the data and cleaning abnormal or erroneous data.

4. The cross-domain variable configuration high-speed aircraft adaptive attitude control method based on neural network identification according to claim 3 is characterized in that: The three-channel neural network model described in step 2 includes: The pitch channel neural network model: Input layer: The number of input layer nodes is 3; Hidden layer: The hidden layer consists of 3 layers, and the neurons use Tanh function; Output layer: The number of nodes in the output layer is 1, and the linear activation function is used; The yaw channel neural network model and the roll channel neural network model both include: Input layer: The number of input layer nodes is 4; Hidden layer: The hidden layer of the yaw channel and the roll channel consists of 3 layers, and the neurons use the Tanh function; Output layer: The number of nodes in the output layer is 1, and a linear activation function is used.

5. The cross-domain variable configuration high-speed aircraft adaptive attitude control method based on neural network identification according to claim 4 is characterized in that: Step 3 includes the following steps: First, collect online data including the attitude angular velocity ω of the three channels of the aircraft x 、ω y 、ω z , the aircraft's angle of attack α, sideslip angle β, bank angle σ, dynamic pressure Q, and pitch rudder angle δ z , yaw rudder deviation δ y and the roll rudder bias δ x , folding angle δ f and the characteristic area S, characteristic lengths b and c of the aircraft; By analyzing the attitude angular velocity ω of the three channels x 、ω y 、ω z Perform differential calculation to obtain attitude angular acceleration The true torque of the three channels is obtained by inversely solving the aircraft attitude rotation dynamics equation: Among them, M tx 、M ty 、M tz are the true moments of the roll channel, yaw channel, and pitch channel, respectively; I1~I9 are the moment of inertia parameters of the aircraft; Furthermore, the true moment coefficients of the three channels are solved by inversely solving the definition of the aircraft aerodynamic torque: After data segmentation, we can obtain the online transfer learning sample library for the three-channel aerodynamic model of cross-domain variable-configuration high-speed aircraft, including the pitch channel sample library [α δ f C mz δ z ], yaw channel sample library [β α δ f C my δ y ] and the rolling channel sample library [β α δ f C mx δ x ].

6. The cross-domain variable configuration high-speed aircraft adaptive attitude control method based on neural network identification according to claim 5 is characterized in that: Step 5 includes the following steps: First, the dynamic equation of the aircraft attitude rotation is established as: Simplifying it to: Among them, x1=[α β σ] T , x2[ω x ω y ω z ] T , the control input is u=[M x M y M z ]; Furthermore, the outer loop controller is set up, and the specific steps are as follows: For the attitude angle equation: The angular tracking error is defined as: e1=x1-x c The angle tracking error change rate is: x 2c Considered as a virtual control quantity, select Where k1 is a diagonal positive definite matrix, and its form is: k1=diag(k 11 ,k 12 ,k 13 ), x 2c is the desired angular velocity; Furthermore, the inner loop controller is set up, and the specific steps are as follows: For the attitude angular velocity equation: The angular velocity tracking error is defined as: e2=x2-x 2c The angular velocity tracking error change rate is: The final output u of the backstepping controller is: Where k2 is a diagonal positive definite matrix, and its form is: k2=diag(k 21 ,k 22 ,k 23 ); The final output u obtained by the backstepping controller is the required aerodynamic torque [M tx M ty M tz ], the required torque coefficients of the three channels are solved by inversely solving the definition of the aircraft aerodynamic torque: After normalizing the torque coefficient and related state quantities required at the current moment, they are input into the three-channel adaptive rudder deviation solver neural network model completed by online migration in step 4 to obtain the rudder deviation command.

7. A cross-domain variable configuration high-speed aircraft adaptive attitude control system based on neural network identification, characterized in that: The system has a program module corresponding to the steps of the method described in any one of claims 1 to 6 above, and executes the steps in the above-mentioned cross-domain variable-configuration high-speed aircraft adaptive attitude control method based on neural network identification when running.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps in the cross-domain variable-configuration high-speed aircraft adaptive attitude control method based on neural network identification according to any one of claims 1 to 6 when called by a processor.

Images