Intelligent flight control method based on offline learning and online optimization

Abstract

The invention belongs to the technical field of intelligent control of fixed-wing aircrafts, and relates to an intelligent flight control method based on offline learning and online optimization. Firstly, an aircraft six-degree-of-freedom kinetic equation is established; secondly, performing systematic pulling deviation processing on the key pneumatic parameters according to a deviation working condition in which the pneumatic coefficient possibly exists, and performing control parameter optimization under a large number of working conditions by adopting a simulation method consistent with reference controller parameter optimization; on the basis, a mapping relation model between the aerodynamic coefficient deviation value and the optimal control parameter is constructed through offline training, and a neural network fitting model capable of being directly called is formed; finally, in the actual flight process, aerodynamic coefficient deviation information under the current working condition is obtained in real time through an online aerodynamic parameter identification module, an offline trained neural network model is called, dynamic correction and real-time optimization are conducted on control parameters, and accurate matching of a control strategy and the actual flight working condition is achieved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for fixed-wing aircraft, and relates to an intelligent flight control method based on offline learning and online optimization. Background Technology

[0002] Against the backdrop of the continuous expansion and deepening of aviation mission scenarios, the aerodynamic shape and handling characteristics of advanced aircraft are becoming increasingly complex. The interactions between its multiple subsystems, including aerodynamics, propulsion, and control, are significantly enhanced, exhibiting typical characteristics of strong coupling, strong nonlinearity, and strong time-varying features. This change directly leads to a significant deviation between the dynamic models upon which control system design relies and actual flight conditions. This discrepancy makes it difficult to establish mathematical models that accurately describe the dynamic characteristics of advanced aircraft across the entire mission profile. This not only significantly increases the technical difficulty of dynamic modeling but also poses new challenges to the design accuracy, robustness, and adaptability of control systems, becoming one of the key bottlenecks restricting the performance improvement of advanced aircraft. At the same time, the increasing complexity and uncertainty of modern aviation mission scenarios are placing higher standards on the intelligent adaptive capabilities of advanced aircraft systems. Traditional control methods based on precise mathematical models are unable to cope with the dynamic changes and uncertainties under complex operating conditions and can no longer meet the control quality requirements of advanced aircraft; while technical paths relying solely on experience-based optimization are also difficult to adapt to the high-safety and high-reliability technical specifications of the aviation field. Summary of the Invention

[0003] To address the significant uncertainties in the dynamic models of advanced aircraft under complex operating conditions and the insufficient adaptability of traditional control methods, this invention proposes an intelligent flight control method based on offline learning and online optimization, building upon the existing benchmark PID control law framework. By constructing a closed-loop control system of "offline modeling and learning - online identification and optimization - dynamic parameter updating," this method aims to more efficiently solve the model uncertainty problem and further improve the control performance of advanced aircraft under complex operating conditions.

[0004] First, the six-degree-of-freedom dynamic equations of the aircraft are established. Second, for operating conditions where aerodynamic coefficients may deviate, key aerodynamic parameters are systematically adjusted, and a simulation method consistent with the optimization of baseline controller parameters is used to optimize control parameters under numerous operating conditions. Based on this, a mapping relationship model between aerodynamic coefficient deviation and optimal control parameters is constructed through offline training, forming a directly callable neural network fitting model. Finally, during actual flight, the aerodynamic coefficient deviation information under the current operating condition is obtained in real time by the online aerodynamic parameter identification module, and the offline trained neural network model is called to dynamically correct and optimize the control parameters in real time, achieving precise matching between the control strategy and the actual flight conditions.

[0005] The technical solution of the present invention is as follows:

[0006] An intelligent flight control method based on offline learning and online optimization, the specific steps of which are as follows:

[0007] Step 1: Establish the six-degree-of-freedom dynamic equations of the aircraft

[0008] The aerodynamic forces acting on the aircraft and their corresponding moments are expressed in the following form:

[0009] (1)

[0010] In the formula, As resistance, For lift, It is a lateral force; For rolling torque, For pitching moment, For yaw moment, For dynamic pressure, This represents the characteristic surface area of ​​the organism. For the average aerodynamic chord length, For wingspan. The drag coefficient, The lift coefficient, This is the lateral force coefficient; This is the rolling moment coefficient. This is the pitch moment coefficient. Yaw moment coefficient. When the aircraft is in cruise and normal maneuvering conditions, the dimensionless coefficient is defined as follows:

[0011] (2)

[0012] In the formula, , , They represent , , The baseline value, , , They represent , , The baseline value; They represent right The derivative, They represent right The derivative, They represent right The derivative; They represent right The derivative, They represent right The derivative, They represent right The derivative; For the angle of attack, Sideslip angle, For flight speed, Indicates the aileron deflection angle. Indicates the rudder deflection angle. This refers to the elevator deflection angle. These are the roll rate, yaw rate, and pitch rate, respectively.

[0013] The velocity component expressions in the body coordinate system are as follows:

[0014] (3)

[0015] In the formula, the superscript " " represents the first derivative, For aircraft quality, It is the acceleration due to gravity. For aircraft engine thrust, In the body coordinate system Velocity components on the three axes In the body coordinate system Aerodynamic components on the three axes The pitch angle, This is the roll angle.

[0016] The expressions for the angular velocity components of the three axes of aircraft rotation are as follows:

[0017] (4)

[0018] In the formula, Representing the body relative to the body coordinate system Moment of inertia of the shaft Represents the entire cross-section of the aircraft relative to the aircraft coordinate system. shaft and The product of inertia of the axis, For the thrust line in the body coordinate system The coordinates of the projection point on the surface.

[0019] After obtaining the above equations for angular velocity and velocity components, the equations for angular displacement and linear displacement are described in the ground coordinate system and the body coordinate system, respectively.

[0020] Equations of linear displacement motion in the ground coordinate system:

[0021] (5)

[0022] Equations of angular displacement motion in the body coordinate system:

[0023] (6)

[0024] In the formula, This is the forward flight distance; Flight altitude; Lateral distance; This is the yaw angle.

[0025] Step 2: Design of Control Parameter Optimization Strategy

[0026] Step 2.1, Aerodynamic derivative pull simulation

[0027] By analyzing the static stability coefficient of the aircraft and elevator rudder effect Within the range of -30% to 30%, pull-off operations are performed at 1% intervals to generate various pull-off scenarios.

[0028] Step 2.2: Optimal Control Law Optimization Process

[0029] For each type of pull-off, the corresponding control law needs to be matched and optimized.

[0030] First, the basic control law is determined. The aircraft pitch angle attitude control system adopts the proportional-integral-derivative (PID) control law, the expression of which is:

[0031] (7)

[0032] in, The elevator deflection angle serves as the control input. Here is the pitch angle, and here is the actual value. This is the pitch angle command value. For pitch rate, This is the difference between the commanded pitch angle value and the actual pitch angle value. For proportional gain, For integral gain, This is the differential gain.

[0033] After determining the basic control law, multiple different combinations of control laws are generated in its vicinity using a random algorithm. These randomly generated control laws are then successively substituted into the constructed aircraft attitude control simulation model for simulation. During the simulation, the model calculates the overshoot and response time for each control law combination. Overshoot... Defined as the percentage difference between the maximum value and the steady-state value of the system output response, i.e.

[0034] (8)

[0035] in, To output the maximum value of the response, This is the steady-state value.

[0036] The response time is determined by the time required for the output response in the simulation model to reach 98% of its steady-state value.

[0037] After obtaining the overshoot and response time corresponding to each set of control laws, comparisons and selections are performed, and the set with the smallest overshoot is selected as the optimal control gain for that pull-off condition. By performing an optimization process for each pull-off condition, an optimal control law suitable for each pull-off condition is finally found, thus constructing a dataset containing pull-off conditions with multiple aerodynamic derivatives and their corresponding optimal control laws.

[0038] Step 3: Offline neural network model fitting

[0039] A neural network is trained using the aerodynamic parameter pull and optimal control law dataset constructed in step 2 to learn the direct mapping relationship between changes in aircraft dynamic characteristics and controller parameters. Specifically, the network input is the aircraft static stability coefficient. With elevator rudder effect The pull-off combination; the output corresponds to the optimal control gain obtained after optimization under this aerodynamic configuration. and Through this training process, the neural network will be able to capture the complex influence of aerodynamic uncertainties on control law design, thereby constructing an intelligent matching model that can quickly infer approximate optimal control parameters based on real-time aerodynamic parameters, laying the foundation for realizing offline learning and online optimization of intelligent flight control.

[0040] The Error Backpropagation Algorithm (BP) was chosen as the training method for the neural network because BP networks are feedforward neural networks composed of multiple layers of feedforward units. Furthermore, through the backpropagation process of errors, BP networks can effectively train a series of complex input-output pattern mappings without explicitly describing the mathematical formulas of these mappings beforehand. During the learning process, the BP network follows the steepest descent principle, selectively adjusting the weights and thresholds of each parameter in the network through repeated backpropagation processes to minimize the overall sum of squared network errors.

[0041] Step 4: Online aerodynamic identification and neural network application

[0042] After completing the offline neural network construction in step 3, the online optimization phase begins. First, based on real-time flight data, the elevator effect and static stability coefficient under the current flight condition are identified online. These two actual identified values ​​are compared with the baseline aerodynamic derivatives determined during the aircraft design phase to calculate the real-time aerodynamic parameter deviation. This deviation is then used as an input signal and passed to the pre-trained neural network model. The neural network infers in real time and outputs a set of optimal control law parameters that best match the current aerodynamic parameter deviation. Finally, these real-time generated control law parameters are updated to the flight control computer, enabling dynamic online optimization and adjustment of flight control commands.

[0043] The beneficial effects of this invention are:

[0044] This invention addresses the challenge of insufficient control performance in traditional aircraft attitude control systems under model uncertainty. It constructs an innovative control system that integrates offline learning and online optimization, and focuses on exploring the key applications of neural networks. Compared with traditional PID control, it demonstrates significant advantages in key performance indicators such as overshoot, response time, robustness, and stability. This provides a more reliable and efficient solution for aircraft attitude control, significantly improving aircraft control performance under complex flight conditions. Attached Figure Description

[0045] Figure 1 This is a flowchart of an intelligent flight control method based on offline learning and online optimization;

[0046] Figure 2 This is a basic structural block diagram of a BP neural network;

[0047] Figure 3 It is a BP neural network Prediction error percentage;

[0048] Figure 4 It is a BP neural network Prediction error percentage;

[0049] Figure 5 This is a diagram of the flight altitude tracking control loop structure using the PID control method.

[0050] Figure 6 This is a diagram of the flight speed tracking control loop structure using the PID control method;

[0051] Figure 7 This is a diagram of the pitch angle tracking control loop structure of the PID control method;

[0052] Figure 8 This is a flowchart of online aerodynamic parameter identification based on recursive least squares.

[0053] Figure 9The pitch angle simulation results are based on intelligent flight control with offline learning and online optimization.

[0054] Figure 10 The results are speed simulations of intelligent flight control based on offline learning and online optimization.

[0055] Figure 11 This is a simulation result of the elevator in intelligent flight control based on offline learning and online optimization. Detailed Implementation

[0056] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0057] The overall flowchart of the intelligent flight control method based on offline learning and online optimization is as follows: Figure 1 As shown, the specific implementation method is as follows:

[0058] Step (1) Input the initial state and give the target state.

[0059] A fixed-wing aircraft flies at a speed of 100 m / s at an altitude of 1000 m, with a trim pitch angle of 7.5°. The aircraft then jumps from level flight to a pitch angle of 10°, returns to the trim pitch angle, jumps to 6.5°, and then returns to the trim pitch angle. The pitch angle command tracking effect is observed.

[0060] Table 1 shows the data and status data for fixed-wing aircraft:

[0061] Table 1 Simulation Parameters

[0062]

[0063]

[0064] Step (2) Establish a dynamic model

[0065] To simplify the aircraft dynamics modeling process, the following assumptions are made regarding the aircraft's aerodynamic characteristics and motion laws:

[0066] We assume that gravitational acceleration is a constant and does not change with the aircraft's altitude; we disregard the curvature of the Earth and treat it as a plane; we assume that the ground coordinate system is an inertial reference system and that the Earth is stationary; we treat the aircraft as a rigid body, ignoring the effects of elastic deformation and the changes in the aircraft's mass during short-period attitude control.

[0067] Choosing an appropriate coordinate system is crucial for accurately analyzing the motion and mechanical characteristics of an aircraft. This invention employs a ground coordinate system, a body coordinate system, and a velocity coordinate system in its model construction. All these coordinate systems follow the North-Sky-East coordinate system to ensure the consistency and accuracy of the model.

[0068] a) Ground coordinate system

[0069] Origin of coordinate system Selected at the aircraft's center of gravity, The axis points in the direction of the aircraft's initial flight path. The axis points upward along the vertical line. The axis is perpendicular to the other two axes and forms a right-handed coordinate system. It is mainly used to determine the coordinate position of the aircraft's center of mass in space and the aircraft's attitude in space.

[0070] b) Body coordinate system

[0071] This coordinate system also sets its origin at the aircraft's center of mass. The axis is completely aligned with the longitudinal axis of the aircraft, and its positive direction is towards the nose of the fuselage; The axis lies within the longitudinal plane of symmetry of the machine body, and... The axes are perpendicular, with the positive direction upwards; The spatial orientation of the axis must satisfy the rules of the right-hand rectangular coordinate system, specifically perpendicular to... The plane (i.e. the longitudinal symmetry plane of the body) forms a complete relative coordinate system for the body.

[0072] c) Velocity coordinate system

[0073] The origin of this coordinate system still lies at the aircraft's center of mass. The direction of the axis is consistent with the real-time velocity vector of the aircraft's center of mass; The axis is located within the longitudinal plane of symmetry of the fuselage, and... The axis is vertical and its positive direction is upward; The axes are set according to the logic of a right-handed rectangular coordinate system, perpendicular to... The plane (the plane formed by the velocity vector and the longitudinal symmetry plane) ultimately forms a coordinate system that is dynamically related to the velocity state.

[0074] The aerodynamic forces acting on the aircraft and their corresponding moments are expressed in the following form:

[0075] (9)

[0076] In the formula, As resistance, For lift, It is a lateral force. For rolling torque, For pitching moment, For yaw moment, For dynamic pressure, This represents the characteristic surface area of ​​the organism. For the average aerodynamic chord length, For wingspan, in the pitch moment analysis scenario, the characteristic length needs to be selected as the mean aerodynamic chord length. For the analysis of rolling moment and yaw moment, the characteristic length should be the wingspan. . The drag coefficient, The lift coefficient, This is the lateral force coefficient; This is the rolling moment coefficient. This is the pitch moment coefficient. Yaw moment coefficient. When the aircraft is in cruise and normal maneuvering conditions, the dimensionless coefficient is defined as follows:

[0077] (10)

[0078] In the formula, , , They represent , , The baseline value, , , They represent , , The baseline value; They represent right The derivative, They represent right The derivative, They represent right The derivative; They represent right The derivative, They represent right The derivative, They represent right The derivative; For the angle of attack, Sideslip angle, For flight speed, Indicates the aileron deflection angle. Indicates the rudder deflection angle. This refers to the elevator deflection angle. These are the roll rate, yaw rate, and pitch rate, respectively.

[0079] The velocity component expressions in the body coordinate system are as follows:

[0080] (11)

[0081] In the formula, the superscript " " represents the first derivative, For aircraft quality, It is the acceleration due to gravity. For aircraft engine thrust, In the body coordinate system Velocity components on the three axes In the body coordinate system Aerodynamic components on the three axes The pitch angle, This is the roll angle.

[0082] The expressions for the angular velocity components of the three axes of aircraft rotation are as follows:

[0083] (12)

[0084] In the formula, These represent the moments of inertia of the machine about each axis of the machine's coordinate system. Represents the entire cross-section of the aircraft relative to the aircraft coordinate system. shaft and The product of inertia of the axis, For the thrust line in the body coordinate system The coordinates of the projection point on the surface.

[0085] After obtaining the above equations for angular velocity and velocity components, the equations for angular displacement and linear displacement are described in the ground coordinate system and the body coordinate system, respectively.

[0086] Equations of linear displacement motion in the ground coordinate system:

[0087] (13)

[0088] Equations of angular displacement motion in the body coordinate system:

[0089] (14)

[0090] In the formula, This is the forward flight distance; Flight altitude; Lateral distance; This is the yaw angle.

[0091] Step 3: Design of control parameter optimization strategy

[0092] Step 3.1, Aerodynamic derivative pull simulation

[0093] By analyzing the static stability coefficient of the aircraft and elevator rudder effect Within a range of -30% to 30%, yaw operations were performed at 1% intervals, and the complex coupling relationships between them were carefully considered to generate a variety of unique yaw scenarios. This yaw simulation is based on a large amount of actual flight data and aerodynamic theory. In real flight, disturbances such as atmospheric turbulence and structural deformation encountered by the aircraft cause the aerodynamic derivatives to fluctuate within a certain range. By selectively setting the yaw range and intervals, the range of aerodynamic characteristic changes that the aircraft may encounter in actual flight can be comprehensively and realistically covered as much as possible. These rich and diverse yaw scenarios provide sufficient and representative sample data for subsequent searches for optimal control laws adapted to different disturbance scenarios.

[0094] During aircraft flight, the aerodynamic derivative is a crucial bridge connecting the aircraft's motion state with aerodynamic forces and moments. Its mathematical expression is: Aerodynamic Force and aerodynamic torque This can be expressed as aircraft motion parameters (such as angle of attack). Elevator deflection Functions of (etc.), i.e. , The aerodynamic derivative is the partial derivative of these functions with respect to the corresponding motion parameters, such as the static stability coefficient. and elevator rudder effect These derivatives precisely describe the rate of change of aerodynamic forces and moments with minute variations in aircraft motion parameters, playing a decisive role in aircraft attitude control.

[0095] Step 3.2: Optimal Control Law Optimization Process

[0096] After completing step 3.1, which simulates various disturbances (pulling) during real flight, the corresponding control law needs to be matched and optimized for each disturbance scenario. This optimization process follows a rigorous scientific procedure to ensure the robustness and optimal performance of the control system.

[0097] First, the basic control law is determined. The aircraft pitch angle attitude control system adopts the proportional-integral-derivative (PID) control law, the expression of which is:

[0098] (15)

[0099] in, The elevator deflection angle serves as the control input. Here is the pitch angle, and here is the actual value. This is the pitch angle command value. For pitch rate, This is the difference between the commanded pitch angle value and the actual pitch angle value. For proportional gain, For integral gain, These three gain parameters, representing the differential gain, are fundamental elements of the control law. Their values ​​directly affect the control system's response to aircraft attitude deviations. This invention selects the proportional gain. and differential gain As an optimization parameter, the proportional gain. The differential gain determines the instantaneous response strength of the controller to attitude deviations; the larger the deviation, the stronger the control effect. This approach focuses on adjusting the control action based on the rate of change of attitude deviation, which can effectively suppress system overshoot and make aircraft attitude adjustments more stable.

[0100] After determining the basic control law, multiple different combinations of control laws are generated in its vicinity using a random algorithm. This random generation method fully explores the parameter space surrounding the basic control law, avoiding the omission of possible optimal solutions. Subsequently, the randomly generated control laws are substituted one by one into the constructed aircraft attitude control simulation model for simulation. During the simulation, the model accurately calculates the overshoot and response time corresponding to each control law. Overshoot It is usually defined as the percentage difference between the maximum value and the steady-state value of the system output response, i.e.

[0101] (16)

[0102] in, To output the maximum value of the response, This is the steady-state value.

[0103] Overshoot, as a key indicator for evaluating the accuracy of aircraft attitude adjustments, directly reflects the maximum degree to which the aircraft deviates from the desired attitude during the adjustment process. Response time The response time is an important parameter for measuring the speed of aircraft attitude adjustment. It represents the time required for the aircraft to reach the desired attitude from receiving the attitude adjustment command. Typically, the response time can be determined by the time required for the output response in the simulation model to reach 98% of its steady-state value.

[0104] After obtaining the overshoot and response time for each set of control laws, a systematic comparison and selection process is performed, choosing the set with the smallest overshoot as the optimal control gain for that pull-off condition. By performing this optimization process for each pull-off condition, an optimal control law suitable for each pull-off condition is finally found, laying a solid foundation for achieving precise attitude control of the aircraft under different aerodynamic characteristics. Thus, through the aforementioned steps, a dataset containing pull-off conditions with various aerodynamic derivatives and their corresponding optimal control laws has been constructed.

[0105] Step (4) Offline neural network model fitting

[0106] A neural network is trained using the aerodynamic parameter pull and optimal control law dataset constructed in step 3 to learn the direct mapping relationship between changes in aircraft dynamic characteristics and controller parameters. Specifically, the network input is the aircraft static stability coefficient. With elevator rudder effect The pull-off combination; the output corresponds to the optimal control gain obtained after optimization under this aerodynamic configuration. and Through this training process, the neural network will be able to capture the complex influence of aerodynamic uncertainties on control law design, thereby constructing an intelligent matching model that can quickly infer approximate optimal control parameters based on real-time aerodynamic parameters, laying the foundation for realizing offline learning and online optimization of intelligent flight control.

[0107] The Error Backpropagation Algorithm (BP) was chosen as the training method for the neural network because BP networks are feedforward neural networks composed of multiple layers of feedforward units. Furthermore, through the backpropagation process of errors, BP networks can effectively train a series of complex input-output pattern mappings without explicitly describing the mathematical formulas of these mappings beforehand. During the learning process, the BP network follows the steepest descent principle, selectively adjusting the weights and thresholds of each parameter in the network through repeated backpropagation processes to minimize the overall sum of squared network errors.

[0108] A backpropagation (BP) network is a type of neural network with feedforward characteristics, consisting of multiple layers of feedforward units. Through backpropagation of errors, the BP network can effectively train a series of complex input-output pattern mappings without explicitly describing the mathematical formulas for these mappings beforehand. During the learning process, the BP network follows the steepest descent method, repeatedly performing backpropagation to selectively adjust the weights and thresholds of each parameter in the network, thereby minimizing the sum of squared errors of the overall network. Furthermore, the basic architecture of a BP neural network mainly includes three parts: an input layer, hidden layers, and an output layer. Figure 2 As shown.

[0109] In practical applications, the training process of a BP neural network mainly consists of two stages: forward propagation and backward propagation. First, through forward propagation, the signal is gradually transmitted from the input layer to the hidden layers, and corresponding values ​​are calculated in the process. Next, the signal continues from the hidden layers to the output layer. Then, through backward propagation of the error, the signal is traced back from the output layer to the hidden layers, and corresponding values ​​are calculated in the process. Finally, the signal is transmitted from the hidden layers back to the input layer, thus achieving the gradual training of the weights and thresholds from the hidden layers to the output layer.

[0110] The BP neural network designed above was used to fit and test offline data. The network prediction results are as follows: Figure 3 and Figure 4 As shown: An offline neural network was trained using aerodynamic data obtained through pull-off simulation. Simulation results show that the offline neural network can effectively control the control parameters in the control loop. and The prediction errors were within 0.08% and 0.25% respectively, indicating that the offline neural network had high fitting accuracy, laying the foundation for subsequent online optimization of aerodynamic parameters.

[0111] Step (5) Intelligent flight control simulation for offline learning / online optimization

[0112] Step (5.1) Overall Design Idea of ​​Intelligent Flight Control Based on Offline Learning / Online Optimization

[0113] An aircraft flight model is constructed based on real-time flight data, CFD, and wind tunnel test data. This model is then used as the basis for ground-based control law design, determining a baseline parameter set for the control law with optimal performance. To address the issue of aerodynamic derivative variations during flight, the aerodynamic derivatives of the offline aircraft model are biased and extensive simulations are conducted. From a massive dataset, control parameter increments that satisfy stability margins and offer the best control performance are selected. A mapping between aerodynamic derivative deviations and optimal control parameter increment datasets is established. Then, aerodynamic derivatives are calculated based on the offline flight model, and the difference between this and the online-identified aerodynamic derivatives yields the actual aerodynamic derivative deviation during flight. This deviation is then substituted into the mapping relationship between aerodynamic derivative deviations and optimal control parameter increments to obtain the optimal control parameter increments. Combining these with a series of methods such as adjusting amplitude limits, scaling safety factors, and smoothing transitions, the control parameters are optimized. Several key steps are involved:

[0114] (a) Based on the baseline control law, a large number of offline simulations are performed. The optimal control parameters are determined by the constructed flight system evaluation index, and the relevant aerodynamic derivatives of the pitching moment coefficient are calculated on this basis.

[0115] (b) Based on the offline-designed reference control law, the aerodynamic derivatives such as the derivative of the pitch moment coefficient with respect to the angle of attack and the derivative of the pitch moment coefficient with respect to the elevator are subjected to ±30% bias and a large number of simulation tests are conducted. Appropriate evaluation indicators are selected from the constructed flight system evaluation index system as performance evaluation indicators. The optimal control parameters for direct pitch moment control are screened from a large amount of simulation experimental data, with the controller stability margin as the basic standard.

[0116] (c) The aerodynamic derivatives with and without deviation are subtracted to determine the deviation of the aircraft aerodynamic derivatives. Then, the deviation of each set of aerodynamic derivatives and the corresponding increment of the optimal control parameters are stored. The mapping relationship between the aerodynamic derivative deviation and the optimal control increment is constructed by interpolation table / neural network.

[0117] (d) During the actual flight of the aircraft, the real-time aerodynamic derivative of the aircraft is identified based on the online recursive least squares identification algorithm, and compared with the aerodynamic derivative of the ground reference model. The corresponding deviation is calculated, and the deviation is input into the previously constructed fitting mapping relationship to obtain the optimal control increment under the current condition and compensate into the control law, thereby improving the flight control performance and accuracy of the aircraft.

[0118] Step (5.2) Design of Aircraft Trajectory and Attitude Tracking Control Law

[0119] In practical system control, constructing an accurate mathematical model of the controlled object is difficult, and its parameters are prone to change, leading to high costs for control theory analysis and synthesis. PID control technology, due to its simple structure and convenient parameter adjustment, has become a preferred solution. Its core principle is to calculate and output a control signal based on the input deviation using a proportional-integral-derivative function relationship. Through long-term engineering practice and theoretical verification, this method has proven to be effective for controlling most nonlinear controlled objects.

[0120] The aircraft trajectory tracking command calculation focuses on altitude and speed command tracking, employing a guidance law with PID as its basic configuration, such as... Figure 5 ~ Figure 7 The specific form of the control law is as follows:

[0121] The altitude control law is shown in the following formula:

[0122] (17)

[0123] The speed control law is shown in the following formula:

[0124] (18)

[0125] The attitude control law is shown in the following equation:

[0126] (19)

[0127] in, These are the pitch angle, altitude, and speed command values, respectively. These are the actual values ​​of pitch angle, altitude, and velocity, respectively. This is the elevator command value. This is the pitch rate value. These are the proportional, derivative, and integral parameters of the height control loop. These are the proportional and integral parameters of the speed control loop. These are the proportional, integral, and derivative parameters of the pitch angle control loop, respectively.

[0128] Step (5.3) Online aerodynamic identification and neural network application

[0129] A time-domain recursive least squares identification algorithm is used to achieve online dynamic identification. First, the identification algorithm is verified. Then, based on the identification, sensor noise is added according to the actual situation, and excitation and filters are added for comparative simulation analysis.

[0130] Its process is as follows Figure 8 As shown.

[0131] For parameter estimation, the relationship between the measured output and the model parameters is more important. If the output is given by the following equation, the model is called a linear parametric model.

[0132] (20)

[0133] In the formula, It is the observation matrix, which describes the linear relationship between the input and output. It is the vector of model parameters to be estimated.

[0134] Here we assume a matrix It is known that the measurement equation can be expressed as:

[0135] (twenty one)

[0136] In the formula, This is the actual measured output. It measures noise.

[0137] When using a model of the least squares (LS) form, no information about and Any uncertainty, i.e., no information about and The probability expression. The least squares estimate can be obtained through the following corollary: for a given... , The “best” estimate is obtained by minimizing the weighted sum of the squared errors between the measured output and the model output;

[0138] (twenty two)

[0139] In the known Record all observations and outputs prior to time point [time]. Then the parameters estimated in the first k time steps are:

[0140] (twenty three)

[0141] In the formula, yes Information matrix at time, yes The parameter estimates at time 1.

[0142] If the variance of the measured noise is Then the covariance matrix of the estimated value is:

[0143] (twenty four)

[0144] Let the information matrix be... .

[0145] Note: Initial values ​​are provided here. recursively from The moment begins.

[0146] (25)

[0147] (26)

[0148] (27)

[0149] For unbiased estimates of parameters that are considered as deterministic quantities: ,get:

[0150] (28)

[0151] Obtain the estimated value The corresponding covariance matrix:

[0152] (29)

[0153] In practical applications, as the total amount of information continues to accumulate, the positive definiteness of the information matrix gradually weakens, and its optimization effect on new information gradually approaches zero. This phenomenon is defined as data saturation. To address this issue, this research proposes a recursive algorithm incorporating a forgetting factor, the specific formula of which is as follows:

[0154] (30)

[0155] (31)

[0156] In real flight environments, pitch moment coefficients cannot be directly obtained from sensors and must be derived and calculated based on other observable parameters. Therefore, it is necessary to construct an observation model between observable parameters and moment coefficients to provide the necessary observational data for the parameters to be identified in subsequent identification processes.

[0157] (32)

[0158] In the formula, This is the pitch moment coefficient. For the average aerodynamic chord length, The characteristic surface area of ​​the organism. For dynamic pressure, For roll acceleration, , , For rotational inertia, These are the roll rate, yaw rate, and pitch rate.

[0159] Similarly, the observation model for the lift coefficient is as follows:

[0160] (33)

[0161] In the formula, It's the quality of the aircraft. It's the aircraft's acceleration.

[0162] Based on the established observation model and recursive least squares identification algorithm, online optimization of key aerodynamic parameters and control laws of the aircraft can be further realized in a closed loop. Specifically, after completing the online identification of dynamics through the time-domain recursive least squares identification algorithm, the elevator effect and static stability coefficient, two core control-related parameters, are first extracted from the identification results. Then, the difference between these two parameters and the reference aerodynamic derivatives determined during the aircraft design phase is calculated to obtain the parameter deviation. Next, this deviation is used as an input signal to a preset neural network model. Utilizing the nonlinear mapping capability of the neural network, the optimal control law matching the current parameter deviation is output. Finally, the optimal control law is updated to the aircraft flight control system in real time, realizing the dynamic adjustment of flight control commands. This effectively compensates for the impact of parameter deviation on flight control, significantly improving the control accuracy and stability of the aircraft under complex flight conditions.

[0163] Based on the controller designed above, an intelligent flight control simulation based on offline learning / online optimization was performed. The simulation results are as follows: Figures 9-11As shown in the simulation results curves, the following conclusions can be drawn: After receiving a pitch angle step command, the average pitch angle overshoot before control optimization was 3.24%, while the average pitch angle overshoot after optimization using intelligent flight control based on offline learning / online optimization was 1.58%, a reduction of 51.23% compared to the original. The average pitch angle response time before control optimization was 1.41s, while the average pitch angle response time after optimization using intelligent flight control based on offline learning / online optimization was 1.04s, a reduction of 26.24% compared to the original. It can be seen that introducing a neural network to update the control gain online can significantly improve the pitch angle tracking control capability.

Claims

1. An intelligent flight control method based on offline learning and online optimization, characterized in that, The specific steps are as follows: Step 1: Establish the six-degree-of-freedom dynamic equations of the aircraft; Step 2: Design of control parameter optimization strategy; Step 2.1, Aerodynamic derivative pull simulation By analyzing the static stability coefficient of the aircraft and elevator rudder effect Within the range of -30% to 30%, pull-off operations are performed at 1% intervals to generate various pull-off scenarios; Step 2.2: Optimal Control Law Optimization Process For each type of pull-off, the corresponding control law needs to be matched and optimized; This allows us to construct a dataset containing various aerodynamic derivative deflection scenarios and their corresponding optimal control laws. Step 3: Offline neural network model fitting; A neural network is trained using the aerodynamic parameter pull and optimal control law dataset constructed in step 2 to learn the direct mapping relationship between changes in aircraft dynamic characteristics and controller parameters; specifically, the network input is the static stability coefficient of the aircraft. With elevator rudder effect The pull-off combination; the output corresponds to the optimal control gain obtained after optimization under this aerodynamic configuration. and Thus, an intelligent matching model can be constructed that can quickly infer approximate optimal control parameters based on real-time aerodynamic parameters. Step 4: Online aerodynamic identification and neural network application After completing the construction of the offline neural network in step 3, the online optimization phase begins. First, based on real-time flight data, the elevator effect and static stability coefficient under the current flight condition are identified online. These two actual identified values ​​are then compared with the baseline aerodynamic derivatives determined during the aircraft design phase to calculate the real-time aerodynamic parameter deviation. Subsequently, this deviation is used as an input signal and transmitted to the pre-trained neural network model. The neural network infers in real time and outputs a set of optimal control law parameters that best match the current aerodynamic parameter deviation. Finally, these real-time generated control law parameters are updated to the flight control computer, realizing dynamic online optimization and adjustment of flight control commands.

2. The intelligent flight control method based on offline learning and online optimization according to claim 1, characterized in that, Step 1 is as follows: The aerodynamic forces acting on the aircraft and their corresponding moments are expressed in the following form: (1) In the formula, As resistance, For lift, It is a lateral force; For rolling torque, For pitching moment, For yaw moment, For dynamic pressure, The characteristic surface area of ​​the organism; For the average aerodynamic chord length, For wingspan; The drag coefficient, The lift coefficient, This is the lateral force coefficient; This is the rolling moment coefficient. This is the pitch moment coefficient. Yaw moment coefficient; when the aircraft is in cruise and normal maneuvering conditions, the dimensionless coefficient is defined as follows: (2) In the formula, , , They represent , , The baseline value, , , They represent , , The baseline value; They represent right The derivative, They represent right The derivative, They represent right The derivative; They represent right The derivative, They represent right The derivative, They represent right The derivative; For the angle of attack, Sideslip angle, For flight speed, Indicates the aileron deflection angle. Indicates the rudder deflection angle. This refers to the elevator deflection angle. These are the roll rate, yaw rate, and pitch rate, respectively. The velocity component expressions in the body coordinate system are as follows: (3) In the formula, the superscript " " represents the first derivative, For aircraft quality, It is the acceleration due to gravity. For aircraft engine thrust, In the body coordinate system Velocity components on the three axes In the body coordinate system Aerodynamic components on the three axes The pitch angle, For roll angle; The expressions for the angular velocity components of the three axes of aircraft rotation are as follows: (4) In the formula, Representing the body relative to the body coordinate system Moment of inertia of the shaft, Represents the entire cross-section of the aircraft body relative to the aircraft body coordinate system. shaft and The product of inertia of the axis, For the thrust line in the body coordinate system The coordinates of the projection point on the surface; After obtaining the above equations for angular velocity and velocity components, the equations for angular displacement and linear displacement are described in the ground coordinate system and the body coordinate system, respectively. Equations of linear displacement motion in the ground coordinate system: (5) Equations of angular displacement motion in the body coordinate system: (6) In the formula, This is the forward flight distance; Flight altitude; Lateral distance; This is the yaw angle.

3. The intelligent flight control method based on offline learning and online optimization according to claim 1, characterized in that, Step 2.2 is as follows: First, the basic control law is determined. The aircraft pitch angle attitude control system adopts the proportional-integral-derivative (PID) control law, the expression of which is: (7) in, The elevator deflection angle serves as the control input. Here is the pitch angle, and here is the actual value. This is the pitch angle command value. For pitch rate, This is the difference between the commanded pitch angle value and the actual pitch angle value. For proportional gain, For integral gain, This is the differential gain; After determining the basic control law, multiple different combinations of control laws are generated in its vicinity using a random algorithm. Subsequently, the randomly generated control laws are successively substituted into the constructed aircraft attitude control simulation model for simulation. During the simulation, the model calculates the overshoot and response time corresponding to each control law combination; the overshoot... Defined as the percentage difference between the maximum value and the steady-state value of the system output response, i.e. (8) in, To output the maximum value of the response, This is the steady-state value; The response time is determined by the time required for the output response in the simulation model to reach 98% of its steady-state value. After obtaining the overshoot and response time corresponding to each control law, they are compared and filtered, and the group with the smallest overshoot is selected as the optimal control gain for that pull-off case. By performing the optimization process for each pull-off case, the optimal control law that fits each pull-off case is finally found, thus constructing a dataset containing pull-off cases with multiple aerodynamic derivatives and their corresponding optimal control laws.

4. The intelligent flight control method based on offline learning and online optimization according to claim 1, characterized in that, In step 3, the backpropagation algorithm (BP) is selected as the training method for the neural network.

Images