A method, apparatus, equipment, and medium for adaptive stability margin configuration of aircraft based on residual learning.
Patent Information
- Application Number
- CN202611154756.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies have high computational complexity when adjusting the stability margin of aircraft, making it difficult to meet real-time requirements and lacking control precision. In particular, they are difficult to achieve the desired stability margin when there are unknown input delays.
A residual learning-based approach is adopted to identify system parameters and time delay estimates through observation data. The residual learning network is used to determine controller parameters to compensate for residuals. Combined with online parameter estimation and adaptive control law, the adaptive stability margin configuration of the aircraft is realized.
It reduces the computational complexity of controller parameters, meets the real-time requirements of online aircraft adjustment, improves control accuracy, adapts to dynamic changes in model parameters and time delay, and ensures the safety of flight control.
Smart Images

Figure CN123067490A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of aircraft control technology, and in particular to an adaptive stability margin configuration method, apparatus, equipment and medium for aircraft based on residual learning. Background Technology
[0002] Stability margin is a crucial indicator of the safety of an aircraft control system. To ensure flight safety, the flight control system needs to maintain a certain stability margin. Currently, to adjust the stability margin to the desired level, the controller parameters are generally determined by directly solving the transcendental equations concerning the desired stability margin, estimated system parameters, and controller parameters. However, directly solving the transcendental equations is computationally complex, especially for aircraft systems with unknown input delays, where the difficulty increases further, making it difficult to meet the real-time requirements of online adaptive adjustment during flight. Furthermore, due to the influence of model parameter changes and time delay uncertainties during flight, the controller parameters obtained by direct solution are unlikely to meet the desired stability margin requirements, resulting in insufficient control accuracy.
[0003] Therefore, the existing technology still needs to be improved and enhanced. Summary of the Invention
[0004] The technical problem to be solved by this application is to provide a method, apparatus, device and medium for adaptive stability margin configuration of aircraft based on residual learning, which addresses the shortcomings of the existing technology.
[0005] To address the aforementioned technical problems, the first aspect of this application provides a method for adaptive stability margin configuration of aircraft based on residual learning, wherein the method specifically includes: Based on observation data, the estimated values of system parameters and time delay of the aircraft are identified, including elevator deflection angle and pitch rate. The approximate controller parameters of the aircraft are determined based on the system parameter estimates, the time delay estimates, and the preset expected stability margin. Using the approximate controller parameters as prior knowledge, the controller parameters are determined through a residual learning network based on the estimated system parameters, the estimated time delay, and the expected stability margin to compensate for the residuals. The target controller parameters are determined based on the controller parameter compensation residuals and the approximate controller parameters, and the aircraft is controlled based on the target controller parameters. The steps of identifying the estimated system parameters and time delay of the aircraft based on the observation data are re-executed according to preset conditions to achieve adaptive stability margin configuration during the flight process of the aircraft.
[0006] The aforementioned adaptive stability margin configuration method for aircraft based on residual learning, wherein the aircraft controller is: , in, This represents the transfer function of the controller. Represents the proportional gain coefficient. Represents the integration time constant. This represents the Laplace operator.
[0007] The adaptive stability margin configuration method for aircraft based on residual learning includes a compensation residual for the proportional gain coefficient and a compensation residual for the reciprocal of the integral time constant.
[0008] The aforementioned adaptive stability margin configuration method for aircraft based on residual learning, wherein the residual learning network adopts a multilayer perceptron as the basic regression network structure and a two-layer fully connected hidden layer architecture, with a Dropout layer introduced between each hidden layer, wherein the Huber loss function is used as the optimization objective for training the residual learning network during the training process.
[0009] The adaptive stability margin configuration method for aircraft based on residual learning, wherein the identification of the estimated system parameters and time delay values of the aircraft based on observation data specifically includes: Obtain system parameters and the parameter update law corresponding to time delay; Based on the observation data, the system parameter estimates and time delay estimates of the aircraft are determined using the parameter update law; The parameter update law is expressed as follows: , , in, The derivative of the generalized vector to be estimated is... Represents the covariance matrix. This represents the estimation error of the pitch angular velocity. express Relative to the generalized vector to be estimated gradient, Represents the normalization factor. The derivative of the covariance matrix is... Indicates the forgetting factor, Indicates time.
[0010] The aforementioned adaptive stability margin configuration method for aircraft based on residual learning, wherein the construction process of the parameter update law corresponding to the system parameters and time delay specifically includes: A stable filter is used to calculate the system model of the aircraft, and a regression model containing time delay nonlinear terms is constructed using the calculated system model. The estimation error of the system output is constructed based on the regression model. Introduce a generalized vector to be estimated with unknown parameters and time delay, and obtain the gradient of the estimation error relative to the generalized vector to be estimated; Based on the estimation error, the generalized vector to be estimated, and the gradient, a cost function with a forgetting factor is constructed. The parameter update law is determined based on the cost function using an approximate nonlinear least squares algorithm. The regression model is expressed as follows: , , , , in, Indicates system output, Indicates system control variables. Represents a vector of unknown parameters. Represents the regression vector. Indicates time, Indicates a stable filter. Indicates adoption Operator role The following signal, Indicates adoption effect The following signal, Indicates adoption effect The following signal, Indicates adoption effect The signal after.
[0011] The aforementioned residual learning-based adaptive stability margin configuration method for aircraft includes a step where, when determining the estimated system parameters and time delay values of the aircraft using the parameter update law, parameter boundary constraint compensation terms are configured for the system parameters and time delay to impose boundary restrictions on the estimated system parameters and time delay values. The parameter boundary constraint compensation terms are: , , in, The first vector to be estimated represents the generalized vector to be estimated. Boundary constraint functions for the parameters to be estimated. Indicates the first One parameter to be estimated, Indicates the first The lower limit value of each parameter to be estimated. Indicates the first The upper limit value of each parameter to be estimated. Indicates the first The update speed of each parameter to be estimated.
[0012] A second aspect of this application provides an adaptive stability margin configuration device for aircraft based on residual learning, wherein the adaptive stability margin configuration device for aircraft based on residual learning specifically includes: The identification module is used to identify the estimated system parameters and time delay of the aircraft based on observation data; The determination module is used to determine the approximate controller parameters of the aircraft based on the estimated system parameters, the estimated time delay, and the preset expected stability margin. The residual compensation module is used to determine the controller parameters to compensate for residuals based on the approximate controller parameters as prior knowledge, the estimated system parameters, the estimated time delay, and the expected stability margin through a residual learning network. The control module is used to determine the target controller parameters based on the controller parameter compensation residuals and the approximate controller parameters, control the aircraft based on the target controller parameters, and re-execute the steps of identifying the system parameter estimates and time delay estimates of the aircraft based on observation data according to preset conditions, so as to realize the adaptive stability margin configuration during the flight process of the aircraft.
[0013] A third aspect of this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the residual learning-based adaptive stability margin configuration method for aircraft as described above.
[0014] A fourth aspect of this application provides a terminal device, which includes: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the residual learning-based adaptive stability margin configuration method for aircraft as described above.
[0015] Beneficial Effects: This application provides a method, apparatus, device, and medium for adaptive stability margin configuration of an aircraft based on residual learning. The method includes: identifying estimated system parameters and time delay values of the aircraft based on observation data; determining approximate controller parameters of the aircraft based on the estimated system parameters, the estimated time delay values, and a preset desired stability margin; using the approximate controller parameters as prior knowledge, determining controller parameters to compensate for residuals based on the estimated system parameters, the estimated time delay values, and the desired stability margin through a residual learning network; determining target controller parameters based on the controller parameter compensation residuals and the approximate controller parameters, and controlling the aircraft based on the target controller parameters; and re-executing the steps of identifying estimated system parameters and time delay values of the aircraft based on observation data according to preset conditions to achieve adaptive stability margin configuration during aircraft flight. This application integrates online parameter estimation, residual learning networks, and adaptive control laws to achieve autonomous configuration of the system's phase margin and gain margin, ensuring stable tracking of the pitch angular velocity to the reference command. This reduces the computational complexity of solving the controller parameters, meets the real-time requirements of online aircraft adjustments, and compensates for parameter errors in approximate solutions through residual learning. This ensures that the configured controller parameters meet the desired stability margin requirements, improves control accuracy, and can adapt to dynamic changes in model parameters and time delays during flight, thus guaranteeing the safety of flight control. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 A flowchart of an adaptive stability margin configuration method for aircraft based on residual learning provided in an embodiment of this application.
[0018] Figure 2 This is a diagram showing the convergence effect of the residual learning network.
[0019] Figure 3 for The fitting effect diagram.
[0020] Figure 4 for The fitting effect diagram.
[0021] Figure 5 This is a diagram showing the error distribution of the phase margin.
[0022] Figure 6This is a graph showing the error distribution of the gain margin.
[0023] Figure 7 For stability margin A schematic diagram of the output response of the controlled system configured with a residual learning-based adaptive stability margin configuration method for aircraft.
[0024] Figure 8 For stability margin BODE diagram of the controlled system under the condition of configuring an adaptive stability margin configuration method for an aircraft based on residual learning.
[0025] Figure 9 For stability margin The controller parameter curves of the controlled system are configured using a residual learning-based adaptive stability margin configuration method for aircraft.
[0026] Figure 10 A schematic diagram of the adaptive stability margin configuration device for aircraft based on residual learning provided in the embodiments of this application.
[0027] Figure 11 A schematic block diagram of the terminal device provided in the embodiments of this application. Detailed Implementation
[0028] This application provides a method, apparatus, device, and medium for adaptive stability margin configuration of aircraft based on residual learning. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description, with reference to the accompanying drawings and embodiments, further illustrates this application. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0029] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.
[0030] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.
[0031] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.
[0032] The application content will be further explained below with reference to the accompanying drawings and the description of the embodiments.
[0033] For a small-disturbance model of the aircraft's longitudinal channel (i.e., the controlled system), its system model can be represented by a second-order transfer function with time delay. This second-order transfer function with time delay is: , in, The transfer function representing the pitch rate channel of the aircraft. Indicates the pitch rate of the aircraft. Indicates the elevator deflection angle. Indicates the damping ratio. Indicates the natural frequency. Represents the time constant. Indicates aerodynamic gain. Indicates time delay. This represents the Laplace operator.
[0034] In this system model, , , and For coefficient parameters, Due to the time delay in the control loop, the system output is the pitch rate of the aircraft. The control input is the elevator deflection angle. The deflection command is issued by the controller corresponding to the system model.
[0035] It is understood that the control objective of this application embodiment is the pitch rate of the aircraft. The control input is the elevator deflection angle. Assume the desired pitch rate tracks a reference signal. The actual pitch angular velocity of the aircraft measured by the sensors is Therefore, the difference between the reference signal and the actual measured pitch angular velocity is the tracking error, which can be expressed as: , in, This indicates the tracking error.
[0036] To achieve the control objective, this embodiment first constructs a controller for the system model. This controller can use a proportional-integral (PI) controller as the base control structure. The elevator deflection angle serves as the control input to the system model. It can be represented as: , in, Represents the proportional gain coefficient. Represents the integration time constant. This represents the Laplace operator.
[0037] The transfer function of the controller is expressed as: , in, This represents the transfer function of the controller.
[0038] Furthermore, when using the aforementioned controller, the open-loop amplitude-frequency characteristics of the controlled system... This represents the gain and phase frequency response of the control loop of the controlled system to the reference signal as the frequency of the input signal changes. This represents the magnitude of the phase lag in the control loop of the controlled system as the frequency of the input reference signal changes. This open-loop amplitude-frequency characteristic... Phase frequency characteristics It can be represented as: , , in, Angular frequency is used to characterize how fast a signal changes. It represents the imaginary unit.
[0039] Based on this, the phase margin included in the stability margin and gain margin They can be represented as follows: , , in, This indicates the amplitude crossover frequency, corresponding to the frequency point where the open-loop amplitude-frequency response curve crosses 0dB, i.e., the frequency point where the gain of the reference signal input to the controlled system drops below 1. This represents the phase crossover frequency, corresponding to the phase frequency response curve and... The frequency of intersection.
[0040] It should be noted that the gain margin One form is ratio-based gain margin. Used to derive integral laws; one form is based on decibel amplitude margin. It is used for the formulation of industry standards. Specifically, and Satisfy the following equation: , .
[0041] The above phase margin and gain margin The expression and and The system of equations formed by the satisfied equations reflects the relationship between system parameters and system stability margin.
[0042] In order to improve the phase margin and gain margin By configuring the desired phase margin and desired gain margin, the above equations can be directly solved to obtain the quantitative relationship between the desired stability margin, system parameters, time delay, and controller parameters. Then, with the desired stability margin, system parameters, and time delay known, the controller parameters can be directly determined, completing the controller parameter tuning. However, directly solving this equation system presents a computational challenge, making it impractical for engineering applications.
[0043] Based on this, embodiments of this application provide an adaptive stability margin configuration method for aircraft based on residual learning, such as... Figure 1 As shown, the aircraft adaptive stability margin configuration method based on residual learning specifically includes: S10. Based on observation data, identify the estimated system parameters and time delay of the aircraft.
[0044] Specifically, system parameters and time delays are crucial for configuring the stability margin near the desired value; however, system parameters and time delays may not be directly available. Therefore, this application embodiment utilizes an online parameter estimator to address uncertainties in the system model. Specifically, this application embodiment first acquires observation data, and then, based on this observation data, uses an online parameter estimator to estimate the system parameters and time delay to obtain estimated system parameter values and estimated time delay values. The observation data may include the aircraft's pitch rate and elevator deflection, where the pitch rate is an actual measured value, the elevator deflection is the control input of the controlled system, and the pitch rate is the output of the controlled system. The online parameter estimator may employ algorithms such as least squares.
[0045] However, in practical applications, unknown time delays introduce nonlinear characteristics into the controlled system, making it impossible to represent the generalized estimated vector, including system parameters and time delays, in a linear regression form. This results in the least squares algorithm being unable to simultaneously and accurately estimate the unknown system parameters and time delays. Therefore, this application employs an approximate nonlinear least squares algorithm to construct parameter update laws for the system parameters and time delays, and uses these parameter update laws to determine the estimated values of the system parameters and time delays.
[0046] Accordingly, the system parameter estimates and time delay estimates for identifying the aircraft based on observation data specifically include: Obtain system parameters and the parameter update law corresponding to time delay; Based on the observation data, the system parameter estimates and time delay estimates of the aircraft are determined using the parameter update law.
[0047] Specifically, the parameter update law is used to iteratively update the system parameter estimates and time delay estimates. In each iteration, based on the currently observed pitch rate and elevator deflection, the estimated residuals of the parameter estimates and time delay are calculated. These residuals are then used to correct the current system parameters and time delay to obtain the final system parameter estimates and time delay estimates. The parameter boundary constraint compensation term is used to limit the updates of the parameter estimates and time delay estimates within a preset physical range, preventing them from jumping out of the value range corresponding to their actual physical meaning, ensuring the stability of the iteration process, and ultimately obtaining converged system parameter estimates and time delay estimates.
[0048] The online parameter estimator depends on the relative order of the controlled system. For the relative order... For the controlled system, the parameter update law is expressed as: , , in, The derivative of the generalized vector to be estimated is... Represents the covariance matrix. This represents the estimation error of the pitch angular velocity. express Relative to the generalized vector to be estimated gradient, Represents the normalization factor. The derivative of the covariance matrix is... Indicates the forgetting factor, Indicates time.
[0049] In one embodiment, the process of constructing the system parameters and the parameter update law corresponding to the time delay specifically includes: A stable filter is used to calculate the system model of the aircraft, and a regression model containing time delay nonlinear terms is constructed using the calculated system model. The estimation error of the system output is constructed based on the regression model. Introduce a generalized vector to be estimated with unknown parameters and time delay, and obtain the gradient of the estimation error relative to the generalized vector to be estimated; Based on the estimation error, the generalized vector to be estimated, and the gradient, a cost function with a forgetting factor is constructed. The parameter update law is determined based on the cost function using an approximate nonlinear least squares algorithm.
[0050] Specifically, to facilitate the explanation of the parameter update law construction process, the system model of the controlled system is represented as follows: , , , , , in, Indicates system output, Indicates system control variables. Represents the Laplace operator. Indicates the damping ratio. Indicates the natural frequency. Represents the time constant. Indicates aerodynamic gain. Indicates time delay. Indicates the start time. Represents the Laplace operator The polynomial composed of , Indicates adoption effect The following signal, Indicates adoption effect The following signal, this action indicates that for any signal , ,have , .
[0051] Furthermore, considering and This may not be readily available; therefore, this embodiment employs a stable filter to obtain the derivative information of the system output. Specifically, calculations are performed on both sides of the system model, i.e., the system model is rewritten as: , , in, Indicates a stable filter. and Represents a stable transfer function. Indicates adoption Operator role The following signal, Indicates adoption effect The signal after.
[0052] Furthermore, a regression model containing time-delay nonlinear terms can be constructed using the calculated system model. This regression model can be expressed as: , , , in, Indicates system output, Indicates system control variables. Represents a vector of unknown parameters. Represents the regression vector. Indicates time, Indicates adoption Operator role The following signal, Indicates adoption effect The following signal, Indicates adoption effect The following signal, Indicates adoption effect The signal after.
[0053] The regression model shows that if the time delay is known, then If it is measurable, then linear estimation algorithms (such as gradient method, least squares method, etc.) can be used to estimate it. However, the time delay considered in the embodiments of this application is unknown, making the linear estimation algorithm unsuitable for estimation. .
[0054] Therefore, the embodiments of this application employ an approximate nonlinear least squares algorithm for estimation. Specifically, let for The estimated value, for The estimated value can determine the estimation error of the system output, which is expressed as: , , in, Indicates the estimation error. express The estimated value.
[0055] Furthermore, the and It can be represented as a vector of generalized parameters to be estimated. ,Should It can be represented as: , in, express The estimated value, express The estimated value, express The estimated value, express The estimated value.
[0056] Then, the gradient of the estimation error relative to the generalized vector to be estimated is obtained, which can be expressed as: .
[0057] Similar to linear least squares, nonlinear least squares also involves minimizing a cost function, which is: , in, Represents the cost function, This represents the weight matrix, used to prevent estimation errors. Deviation from initial value Too many, and , Represents the normalization factor. This represents the forgetting factor, used to introduce an exponential time-weighted mechanism into the cost function.
[0058] This cost function assigns greater importance to recent estimation errors by distributing increasing weights over time, while gradually reducing the impact of historical errors. Then, a parameter update law can be constructed with the goal of minimizing this cost function. This constructed parameter update law has time-varying tracking capabilities with adjustable memory depth, thereby promoting adaptive behavior over time.
[0059] In one embodiment, when determining the estimated system parameters and time delay of the aircraft using the parameter update law, a parameter boundary constraint compensation term is configured for the system parameters and time delay to impose boundary restrictions on the estimated system parameters and time delay. This parameter boundary constraint compensation term is: , , in, The first vector to be estimated represents the generalized vector to be estimated. Boundary constraint functions for the parameters to be estimated. Indicates the first One parameter to be estimated, Indicates the first The lower limit value of each parameter to be estimated. Indicates the first The upper limit value of each parameter to be estimated. Indicates the first The update speed of each parameter to be estimated.
[0060] The parameter boundary constraint compensation term is used to constrain the update range of the parameter to be estimated, so as to avoid the update result from exceeding the preset reasonable value range during the parameter estimation process, prevent the estimation result from diverging and the controller from becoming unstable due to parameter overflow, and at the same time accelerate the convergence speed of parameter estimation and improve the stability and reliability of online estimation.
[0061] Accordingly, the parameter update law can be updated as follows: , .
[0062] S20. Determine the approximate controller parameters of the aircraft based on the estimated system parameters, the estimated time delay, and the preset expected stability margin.
[0063] It is understandable that in practical engineering applications, the system parameters of the controlled system are generally unknown or dynamically changing with the system state. Relying solely on residual learning networks obtained through offline training may result in low control accuracy. Therefore, this application, when using residual learning networks to achieve high-precision mapping from the desired stability margin to the PI controller parameters, combines online parameter estimation and the residual learning network to construct an adaptive control framework, thereby achieving adaptive high-precision stability margin configuration under unknown system parameters. Specifically, this adaptive control framework uses the residual learning network as the nominal control law, then uses the estimated system parameters and time delay estimates to determine the approximate controller parameters of the aircraft, and provides prior knowledge to the residual learning network through these approximate controller parameters.
[0064] The approximate controller parameters can be obtained using a function approximation method, and the analytical expression for these controller parameters is: , , in, Indicates the desired phase margin. This represents the expected magnitude margin.
[0065] Therefore, after obtaining the estimated system parameters and time delay, the phase margin configured for the controlled system is read. and gain margin (i.e., the preset expected stability margin) allows us to use the analytical expressions of the controller parameters mentioned above to determine the approximate controller parameters of the aircraft.
[0066] It should be noted that, after introducing the generalized estimate vector, the process of identifying the system parameter estimates and time delay estimates can be transformed into an update process for the generalized estimate vector. The analytical expression for the controller parameters can then be updated as follows: , .
[0067] Then, after obtaining the updated generalized estimated vector using the parameter update law, the updated generalized estimated vector can be directly substituted into the analytical expression of the updated controller parameters, and the approximate controller parameters can be obtained by solving the analytical expression of the controller parameters.
[0068] S30. Using the approximate controller parameters as prior knowledge, the controller parameters are determined through a residual learning network based on the estimated system parameters, the estimated time delay, and the expected stability margin to compensate for the residuals.
[0069] Specifically, the residual learning network is pre-trained and used to determine controller parameters to compensate for residuals. The input data of this residual learning network includes system parameters, time delay, desired stability margin, and approximate controller parameters. These approximate controller parameters provide prior knowledge to the residual learning network to achieve adaptive control. The input data and the controller parameter compensation residuals can be expressed as follows: , , , , , , , in, Indicates input data, Represents the first in the generalized vector to be estimated. One parameter to be estimated, This indicates that the controller parameters compensate for the residual. This represents a residual learning network.
[0070] It can be understood that the process of solving for the controller parameters can be described as a mapping from system parameters, time delay, and desired stability margin to controller parameters, which can be expressed as: , in, Indicates the mapping relationship. Represented by 7-dimensional feature vectors Zhang Cheng's parametric space, This represents the 2D controller parameter vector to be solved. Zhang Cheng's output space.
[0071] In this embodiment, approximate controller parameters are used as prior knowledge to provide physical prior data for the residual learning network, and these approximate controller parameters are input into the residual learning network. Therefore, the input space of this residual learning network... For the reason Zhang Cheng's 9-dimensional space allows for the explicit injection of approximate controller parameters obtained through analytical models into the residual learning network. This provides the residual learning network with a clear physical benchmark reference, enabling it to accurately output the corresponding compensation residuals based directly on the information from the approximate calculation results, without having to explore the global distribution of controller parameters from scratch. This significantly reduces the complexity of the learning task and further enhances the physical interpretability of the model output.
[0072] Furthermore, from the analytical expression of the controller parameters, the integral time constant can be determined. The solution is obtained by calculating its reciprocal. Achieved, and compared to , The relationship with system parameters is simpler. Therefore, the embodiments of this application focus on the proportional gain coefficient. With integration time constant A differentiated spatial transformation strategy is adopted, that is, the parameter compensation residual includes the compensation residual of the proportional gain coefficient and the compensation residual of the reciprocal of the integral time constant.
[0073] Specifically, for the proportional gain coefficient direct learning For the integration time constant The learning task is mapped from the integral time constant space to the integral gain space. This not only facilitates the fitting of the mapping relationship between input and output by residual learning networks, but also overcomes numerical singularities at physical boundaries. This is because there may be... as well as Combination makes The value is 0, then the corresponding It will be infinite; even a tiny change in the system parameters near these singularities will lead to... This will produce enormous changes. If we predict directly... Data points that are extremely sensitive to input parameters can have a significant negative impact on the learning process of residual learning networks, while the integral gain coefficient... In the middle, leading to The divergent critical conditions are smoothly mapped to a continuous region near the zero point, thus overcoming the numerical singularity at the physical boundary.
[0074] Therefore, the learning objective of residual learning networks is no longer to directly output controller parameters. and Instead, it predicts the true controller parameters and the approximate controller parameters. and The residual between them (i.e., the controller parameter compensation residual) is denoted as the actual controller parameter. and Then the controller parameter compensation residual can be expressed as: , , in, This represents the compensation residual of the proportional gain coefficient. This represents the compensation residual of the integral gain coefficient.
[0075] Accordingly, the learning objective of the residual learning network is a mapping from the input space to the controller parameter compensation residuals, which can be represented as: , in, This represents the residual space, whose numerical distribution range is much smaller than that of the original controller parameter space. The target controller parameters can be obtained by superimposing the approximate controller parameters and the controller parameter compensation residuals. These target controller parameters can be: , , , in, Indicates the target controller parameters. Represents a residual learning network. This represents the input data for the residual learning network.
[0076] Furthermore, before using the residual learning network to determine the controller parameter compensation residual, the residual learning network will be trained offline to obtain the mapping relationship between the input data (i.e., approximate controller parameters, system parameters, time delay and expected stability margin) and the output data (i.e. the compensation residual of the proportional gain coefficient and the compensation residual of the inverse of the integral time constant).
[0077] In offline training, a training dataset is first acquired. The process of acquiring this training dataset can be as follows: First, for the controlled system, the effective engineering envelope range of the system parameters and time delay, as well as the effective engineering envelope range of the expected stability margin, are determined. Then, a preset number of training data are randomly sampled within their respective effective engineering envelope ranges to avoid the curse of dimensionality in uniform grid traversal. Each training data includes system parameters, time delay, and expected stability margin, and these are known.
[0078] Secondly, based on the training data, the stability margin configuration problem is transformed into a system of frequency domain transcendental equations. First, approximate controller parameters are obtained using the analytical expressions for the controller parameters. Then, using these approximate controller parameters as initial values, a numerical solver is used to iteratively find the roots to solve for the true values of the controller parameters that satisfy the margin constraints. Among them, the numerical solver can utilize nonlinear equation solvers such as fsolve in MATLAB.
[0079] Finally, considering that high-order nonlinear frequency domain equations may have multiple roots or non-physical solutions, the true values of the solved controller parameters are physically verified to remove invalid samples, thereby obtaining the training dataset. ,in, The training data used as input data, For the true values of the controller parameters, This represents the number of samples in the training dataset.
[0080] It should be noted that, in order to improve the reliability of the training dataset, after obtaining the training data and its corresponding true values of controller parameters, the true values of controller parameters can be substituted into the transfer function of the controlled system, and the actual phase margin and gain margin can be calculated back based on methods such as Bode plots to verify the accuracy of the configuration.
[0081] Furthermore, in the aforementioned offline training, after obtaining the training dataset, the initial residual learning network is trained using this dataset. This training process can be as follows: First, to eliminate the bias caused by inconsistencies in the physical dimensions and numerical magnitudes of various features in updating network weights, and to avoid features of large magnitude dominating the gradient descent process, compensation residuals for both input data and output controller parameters are implemented. Standardization processes map the input data and controller parameter compensation residuals to a distribution with a mean of 0 and a standard deviation of 1. Specifically, for and The standardized formula is: , , in, This represents the mean of the input data on the training dataset. This represents the mean of the controller parameter compensation residuals over the training dataset. This represents the standard deviation of the input data on the training dataset. This represents the standard deviation of the controller parameter compensation residuals on the training dataset. This represents the input data after standardization. This represents the controller parameter compensation residual after standardization.
[0082] It should be noted that since the residual learning network is used to learn the mapping from the input space to the controller parameter compensation residual, after obtaining the training dataset, for each training data in the training dataset, the true value of the controller parameter corresponding to the training data can be subtracted from the approximate controller parameter used as the initial value of the iteration for the true value of the controller parameter to obtain the true value of the parameter compensation residual.
[0083] Secondly, considering that the trained residual learning network needs to be deployed on platforms with limited computing resources (such as aircraft), the network structure of the residual learning network needs to be simplified based on accuracy requirements. In the embodiments of this application, the residual network model can use a multi-layer perceptron (MLP) as the basic regression network structure.
[0084] In one embodiment, the topology of the residual learning network can be: 1. Input Layer: A total of 9 neurons are used to receive signals from the input layer. The standardized 9-dimensional input data includes 7-dimensional system parameters, time delay, and desired stability margin of the controlled system, as well as 2-dimensional analytical approximation controller parameters. and This provides prior physical knowledge for residual learning.
[0085] 2. Hidden Layers: A two-layer fully connected hidden layer architecture is adopted. The first fully connected hidden layer has 128 neurons, and the second fully connected hidden layer has 64 neurons. The activation function of the fully connected hidden layers is ReLU (Rectified Linear Unit). This activation function can effectively alleviate the gradient vanishing problem during backpropagation in deep networks, and at the same time accelerate the feature extraction and fitting of the model for strongly nonlinear boundaries such as statically unstable critical conditions.
[0086] 3. Regularization: Considering the strong nonlinearity and sparse distribution of samples at the statically unstable boundary, a Dropout layer is introduced between each hidden layer to avoid overfitting of the residual learning network to extreme conditions. The dropout rate of this Dropout layer is 0.05. This mechanism randomly masks some neuron connections during the training phase, forcing the residual learning network to learn more robust general features and improving the model's generalization ability under unknown flight conditions.
[0087] 4. Output layer: Two neurons are set up, and a linear activation function is used to directly output continuous residual prediction values that are not limited by the value range. No additional nonlinear compression is required, which ensures the flexibility of residual compensation.
[0088] In conventional regression tasks, mean squared error (MSE) loss is the most commonly used loss function. However, in the stability margin configuration scenario of this application embodiment, even through... While the inverse transformation significantly reduces numerical singularity, the true residuals may still exhibit outlier characteristics and a long-tailed distribution under critical boundary conditions. If MSE loss is directly applied, large-biased outliers will generate extremely strong squared penalty terms, causing the residual learning network to over-apply to extreme boundaries during optimization, thereby sacrificing the overall fitting accuracy near the nominal parameters.
[0089] Therefore, in order to balance global fitting accuracy and robustness to boundary conditions, this embodiment of the application adopts the Huber loss function as the optimization objective for training the residual learning network. The Huber loss combines the smooth convergence of MSE with the noise resistance of MAE, is insensitive to outliers, and is very suitable for scenarios with strong nonlinearity in the control parameter boundaries. The Huber loss can be: , in, The true residual vector obtained from the true dataset , The predicted residual vector is the output of the neural network. Threshold hyperparameters are used to distinguish between the quadratic and linear domains. Near the nominal parameter, the neural network error is small, and the Huber loss function maintains the characteristics of a quadratic function, with smooth and sensitive gradients, which is conducive to high-precision model convergence. Near the critical value of the parameter, the neural network error increases, and the loss function degenerates into a linear function, which can significantly suppress the gradient impact caused by outliers and avoid training divergence.
[0090] Furthermore, this application employs an Adaptive Moment Estimation (Adam) optimizer, which combines the advantages of momentum terms and adaptive learning rates, resulting in faster convergence and greater stability than traditional stochastic gradient descent (SGD). The initial learning rate in this paper is set to... To further ensure the stability of the training process under conditions of high latency and strong nonlinearity, a gradient pruning strategy is introduced, and an upper limit is set for the gradient norm. This forces the backpropagation gradient to be constrained within a safe range, thus preventing gradient explosion from a fundamental mechanism.
[0091] It should be noted that in practical applications, other network structures (such as convolutional neural networks, recurrent neural networks, or lightweight Transformer-based network structures) can also be used as residual learning networks according to actual needs. It is only necessary to ensure that the number of features corresponds to the input dimension and the number of residuals compensated by the controller parameters corresponds to the output dimension. This application does not impose any restrictions on the specific structure of the residual learning network. Furthermore, the estimation optimizer and learning rate used during the training process of the residual learning network can also be adjusted according to actual needs, and no specific restrictions are imposed here.
[0092] S40. Determine the target controller parameters based on the controller parameter compensation residual and the approximate controller parameters, and control the aircraft based on the target controller parameters.
[0093] Specifically, after obtaining the controller parameter compensation residual, the controller parameter compensation residual is superimposed with the approximate controller parameters to obtain the target controller parameters, which can be expressed as: , .
[0094] S50. Re-execute the steps of identifying the estimated system parameters and time delay values of the aircraft based on the observation data according to the preset conditions, so as to realize the adaptive stability margin configuration during the flight process of the aircraft.
[0095] Specifically, the preset conditions can be preset time intervals. At each interval, the current observation data of the aircraft is re-acquired, the current system parameters and time delay are re-estimated, and the target controller parameters adapted to the current operating conditions are recalculated to complete the online adaptive update of the controller parameters. Alternatively, the preset conditions can be trigger conditions for aircraft maneuver switching or significant changes in operating conditions. When a significant change in flight status is detected, a new round of stability margin configuration process is triggered to ensure that the pitch channel can meet the expected stability margin requirements in different flight phases, taking into account both flight stability and handling quality.
[0096] To demonstrate the effectiveness of the residual learning-based adaptive stability margin configuration method for aircraft in the embodiments of this application, simulation verification is performed. Specifically, the response of the aircraft's pitch rate loop is characterized as a second-order system with time delay, i.e., the Low-Order Equivalent System (LOES) method is adopted. An aircraft's pitch rate control system is selected as the controlled system for experimental verification: .
[0097] To ensure sufficient robustness of the controlled system when facing complex external environmental disturbances or executing highly maneuverable flight commands, adequate stability margins are reserved. According to general aviation industry standards, the phase margin of the controlled system... Not less than And magnitude margin The minimum value should be no less than 6 dB. However, considering the high degree of operational uncertainty, mechanical wear, and complex aerodynamic nonlinearities in actual flight, and the potential need to balance the time-domain transient performance of the controlled system with the stability margin, the stability margin can be relaxed to 50% of the standard value (i.e., phase margin reduced to 22.5 degrees and amplitude margin reduced to 3 dB) under extreme or severe conditions. To fully verify the effectiveness of the method proposed in this application, a set of representative stability margin target values were selected to demonstrate the impact of selecting the desired value for precise stability margin configuration on the system's frequency domain response performance. Simultaneously, the amplitude margin is measured using a ratio. This allows for a more intuitive measurement of configuration errors by assessing the configuration effectiveness.
[0098] To simulate the physical characteristics that are difficult to model accurately in real flight, this embodiment assumes that the system structural parameters and critical time delays are unknown. The initial parameter configuration of the parameter estimator is as follows: the covariance matrix is initialized to... =diag(10000,10000,10000,10000,10000,1000), where the forgetting factor is set to... The normalization factor is set to To provide sufficient excitation signals for the online parameter estimation process, the pitch rate command is a square wave signal with an amplitude of 0.5 and a period of 6 seconds. After this excitation phase, the command signal is switched and maintained at a steady-state constant value to examine the steady-state regulation performance of the system. Furthermore, to rigorously test the convergence speed and robustness of the parameter estimator, this embodiment sets the initial estimated values of the system parameters to 1.5 times the nominal values and the initial estimated values of the time delay to a deviation value containing 1.2 times the disturbance. This setting with significant initial deviation aims to simulate the harsh conditions of parameter mismatch in actual engineering, thereby verifying the algorithm's self-correction capability.
[0099] It is important to emphasize that there is an inherent coupling relationship between the time-domain dynamic response characteristics and the frequency-domain stability margin index of the controlled system. Specifically, once the stability margin of the controlled system... Once configured within a preset range, the time-domain parameters of the closed-loop system, including overshoot and damping characteristics, are determined accordingly. According to classical control theory, the system's damping characteristics are highly correlated with its phase margin. Generally, a larger phase margin means the system has a higher damping ratio, which can effectively suppress oscillations during transients and improve the system's stability.
[0100] First, based on the full envelope high-fidelity labeled dataset, the dataset is divided into three parts: 80% training set, 10% validation set, and 10% test set. The training set is used for model parameter updates, the validation set for monitoring the training process and hyperparameter tuning, and the test set for final model generalization performance evaluation. The residual learning network in this embodiment is trained offline using the TensorFlow deep learning framework. The core training hyperparameters are configured as follows: batch size (BatchSize) is set to 64, and the maximum number of iterations (Epochs) is set to 200. During training, the Huber loss function curves of the training and validation sets are recorded in real time, and the results are as follows: Figure 2 As shown, the model convergence process can be divided into three typical stages, demonstrating excellent training stability and fitting performance. Within the first 20 iterations of training, both training and validation losses exhibit an exponential and rapid decrease. This is attributed to the adaptive optimization capability of the Adam optimizer and the physical prior guidance provided by the analytical approximate solutions in the input features. Furthermore, the residual learning network framework reduces the burden on the model to fit physical laws. As the number of iterations increases, the loss function remains stable and continuously decreases in logarithmic coordinates. The validation set loss curve is slightly lower than the training set loss in some iteration intervals. This phenomenon is due to the introduction of the Dropout regularization mechanism, which prevents overfitting under statically unstable boundaries and extreme conditions, ensuring excellent generalization ability. Towards the end of training, the model loss value converges to approximately... The extremely low magnitude indicates that the residual learning network has accurately learned the implicit nonlinear mapping relationship between the approximate controller parameters and the true values of the controller parameters, demonstrating that residual compensation achieves high-precision control law parameter configuration. The Huber loss function characterizes the overall loss function change of the neural network. Figure 3 and Figure 4 The proportional gain coefficients are shown respectively. With integral gain coefficient The distribution of prediction accuracy.
[0101] To evaluate the effectiveness of the actual stability margin configuration, the error distributions of phase margin and gain margin were statistically analyzed. The results are as follows: Figure 5 and Figure 6 As shown. From Figure 5 It can be seen that the phase margin error of most data points is within 1%, and the error of almost all data points is controlled within 3%. From Figure 6 It can be seen that the magnitude margin error of almost all data points was controlled within 1%. These results demonstrate that, given the system parameters, the trained neural network can ensure the precise configuration of the system's stability margin over a large parameter space with extremely high accuracy.
[0102] The following will demonstrate the effectiveness of the adaptive stability margin configuration algorithm under unknown system parameters and latency locations, with the stability margin configured as follows: With the simulation parameters set unchanged, the simulation results are as follows: Figure 7 , Figure 8 and Figure 9 As shown.
[0103] Experimental data confirms that the residual learning-based adaptive stability margin configuration method for aircraft provided in this application can accurately complete the stability margin configuration task for different design specifications. Compared with the nominal controller, this method can still maintain excellent control performance when facing parameter uncertainties, thereby ensuring that the closed-loop system has stronger robustness and environmental adaptability, and can achieve high-precision parameter identification and margin adjustment.
[0104] Based on the above-mentioned residual learning-based adaptive stability margin configuration method for aircraft, this embodiment provides a residual learning-based adaptive stability margin configuration device for aircraft, such as... Figure 10 As shown, the aircraft adaptive stability margin configuration device based on residual learning specifically includes: The identification module 100 is used to identify the estimated system parameters and time delay of the aircraft based on observation data; The determination module 200 is used to determine the approximate controller parameters of the aircraft based on the estimated system parameters, the estimated time delay, and the preset expected stability margin. The residual compensation module 300 is used to determine the controller parameters to compensate for residuals based on the system parameter estimates, the time delay estimates, and the expected stability margin through a residual learning network, using the approximate controller parameters as prior knowledge. The control module 400 is used to determine the target controller parameters based on the controller parameter compensation residual and the approximate controller parameters, control the aircraft based on the target controller parameters, and re-execute the steps of identifying the system parameter estimates and time delay estimates of the aircraft based on observation data according to preset conditions, so as to realize the adaptive stability margin configuration during the flight process of the aircraft.
[0105] Based on the above-described residual learning-based adaptive stability margin configuration method for aircraft, this embodiment provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the residual learning-based adaptive stability margin configuration method for aircraft as described in the above embodiment.
[0106] Based on the above-mentioned residual learning-based adaptive stability margin configuration method for aircraft, this application also provides a terminal device, such as... Figure 11As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communications interface 23 and a bus 24. The processor 20, display screen 21, memory 22, and communications interface 23 can communicate with each other via the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setup mode. The communications interface 23 can transmit information. The processor 20 can invoke logical instructions in the memory 22 to execute the methods described in the above embodiments.
[0107] Furthermore, the logical instructions in the aforementioned memory 22 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.
[0108] The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, thereby implementing the methods in the above embodiments.
[0109] The memory 22 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 22 may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as transient storage media.
[0110] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.
[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for adaptive stability margin configuration of an aircraft based on residual learning, characterized in that, The aforementioned residual learning-based adaptive stability margin configuration method for aircraft specifically includes: Based on observation data, the estimated values of system parameters and time delay of the aircraft are identified, including elevator deflection angle and pitch rate. The approximate controller parameters of the aircraft are determined based on the system parameter estimates, the time delay estimates, and the preset expected stability margin. Using the approximate controller parameters as prior knowledge, the controller parameters are determined through a residual learning network based on the estimated system parameters, the estimated time delay, and the expected stability margin to compensate for the residuals. The target controller parameters are determined based on the controller parameter compensation residuals and the approximate controller parameters, and the aircraft is controlled based on the target controller parameters. The steps of identifying the estimated system parameters and time delay of the aircraft based on the observation data are re-executed according to preset conditions to achieve adaptive stability margin configuration during the flight process of the aircraft.
2. The adaptive stability margin configuration method for aircraft based on residual learning according to claim 1, characterized in that, The controller of the aircraft is: , in, This represents the transfer function of the controller. Represents the proportional gain coefficient. Represents the integration time constant. This represents the Laplace operator.
3. The adaptive stability margin configuration method for aircraft based on residual learning according to claim 1, characterized in that, The parameter compensation residuals include the compensation residuals of the proportional gain coefficient and the compensation residuals of the reciprocal of the integral time constant.
4. The adaptive stability margin configuration method for aircraft based on residual learning according to claim 1 or 3, characterized in that, The residual learning network uses a multilayer perceptron as the basic regression network structure and a two-layer fully connected hidden layer architecture. Dropout layers are introduced between each hidden layer. During the training process of the residual learning network, the Huber loss function is used as the optimization objective for training the residual learning network.
5. The adaptive stability margin configuration method for aircraft based on residual learning according to claim 1, characterized in that, The system parameter estimates and time delay estimates for identifying the aircraft based on observation data specifically include: Obtain system parameters and the parameter update law corresponding to time delay; Based on the observation data, the system parameter estimates and time delay estimates of the aircraft are determined using the parameter update law; The parameter update law is expressed as follows: , , in, The derivative of the generalized vector to be estimated is... Represents the covariance matrix. This represents the estimation error of the pitch angular velocity. express Relative to the generalized vector to be estimated gradient, Represents the normalization factor. The derivative of the covariance matrix is... Indicates the forgetting factor, Indicates time.
6. The adaptive stability margin configuration method for aircraft based on residual learning according to claim 1 or 5, characterized in that, The process of constructing the system parameters and the parameter update law corresponding to the time delay specifically includes: A stable filter is used to calculate the system model of the aircraft, and a regression model containing time delay nonlinear terms is constructed using the calculated system model. The estimation error of the system output is constructed based on the regression model. Introduce a generalized vector to be estimated with unknown parameters and time delay, and obtain the gradient of the estimation error relative to the generalized vector to be estimated; Based on the estimation error, the generalized vector to be estimated, and the gradient, a cost function with a forgetting factor is constructed. The parameter update law is determined based on the cost function using an approximate nonlinear least squares algorithm. The regression model is expressed as follows: , , , , in, Indicates system output, Indicates system control variables. Represents a vector of unknown parameters. Represents the regression vector. Indicates time, Indicates a stable filter. Indicates adoption Operator role The following signal, Indicates adoption effect The following signal, Indicates adoption effect The following signal, Indicates adoption effect The signal after.
7. The adaptive stability margin configuration method for aircraft based on residual learning according to claim 5, characterized in that, When determining the estimated system parameters and time delay of the aircraft using the parameter update law, parameter boundary constraint compensation terms are configured for the system parameters and time delay to impose boundary restrictions on the estimated system parameters and time delay. The parameter boundary constraint compensation terms are: , , in, The first vector to be estimated represents the generalized vector to be estimated. Boundary constraint functions for the parameters to be estimated. Indicates the first One parameter to be estimated, Indicates the first The lower limit value of each parameter to be estimated. Indicates the first The upper limit value of each parameter to be estimated. Indicates the first The update speed of each parameter to be estimated.
8. An adaptive stability margin configuration device for aircraft based on residual learning, characterized in that, The aforementioned residual learning-based adaptive stability margin configuration device for aircraft specifically includes: The identification module is used to identify the estimated system parameters and time delay of the aircraft based on observation data; The determination module is used to determine the approximate controller parameters of the aircraft based on the estimated system parameters, the estimated time delay, and the preset expected stability margin. The residual compensation module is used to determine the controller parameters to compensate for residuals based on the approximate controller parameters as prior knowledge, the estimated system parameters, the estimated time delay, and the expected stability margin through a residual learning network. The control module is used to determine the target controller parameters based on the controller parameter compensation residuals and the approximate controller parameters, control the aircraft based on the target controller parameters, and re-execute the steps of identifying the system parameter estimates and time delay estimates of the aircraft based on observation data according to preset conditions, so as to realize the adaptive stability margin configuration during the flight process of the aircraft.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the residual learning-based adaptive stability margin configuration method for aircraft as described in any one of claims 1-7.
10. A terminal device, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the residual learning-based adaptive stability margin configuration method for aircraft as described in any one of claims 1-7.
