Search

CN-122018297-A - Aircraft element learning robust control method and system considering actuator dynamics

CN122018297ACN 122018297 ACN122018297 ACN 122018297ACN-122018297-A

Abstract

A method for controlling the element learning robustness of aircraft includes such steps as creating a controlled object model containing high-order actuator dynamic, incorporating the actuator dynamic into inverse dynamic decoupling process by nonlinear dynamic inverse calculation, eliminating modeling deviation and closed-loop stability risk caused by neglecting the response delay of actuator, off-line creating a multi-task scene by meta-learning technique, extracting the general structural feature basic function matrix capable of representing the uncertainty and commonality features of system, and real-time estimating and compensating the lumped error caused by the inconsistency between nominal model and real object.

Inventors

  • Hong Haichao
  • LIU YUE
  • HU SHIQIANG

Assignees

  • 上海交通大学

Dates

Publication Date
20260512
Application Date
20260212

Claims (10)

  1. 1. A method for controlling the robust learning of aircraft elements by considering actuator dynamics includes such steps as creating a generalized aircraft nonlinear system model containing explicit actuator dynamics, high-order differentiating the model output, deriving the ANDI control law based on nominal parameters, incorporating the actuator dynamics into inverse control design to compensate the response delay caused by limited bandwidth of actuator, creating a multi-task data set, training neural network to extract general structural feature matrix, on-line flying, creating a composite error variable containing prediction error and tracking error by off-line training and fixed structural feature matrix, real-time updating the uncertainty parameters by adaptive law of leakage terms to estimate the lumped uncertainty caused by the mismatch of actuator parameters and pneumatic perturbation, and substituting the uncertainty compensation term obtained by on-line estimation into the ANDI control law based on nominal parameters to generate final robust control instruction and output it to the actuator.
  2. 2. The method for robust control of aircraft element learning taking actuator dynamics into account according to claim 1, characterized in that it comprises in particular: The method comprises the steps of 1) establishing a generalized aircraft input affine nonlinear system model containing explicit actuator dynamics and a generalized high-order actuator dynamics model, obtaining each derivative signal of a controlled object by adopting a high-order state estimation and reconstruction technology, and constructing a nonlinear dynamic inverse control law taking the actuator dynamics into consideration nominally based on an inverse dynamics principle, namely an ANDI control law based on nominal parameters; Step 2) offline element learning and feature priori construction, namely constructing a task data set containing multiple working conditions, designing an MLP neural network fitting regression matrix, establishing an element learning joint loss function, and reversely propagating and updating network weights and parameters; and 3) constructing an online composite adaptive law, namely estimating uncertainty parameters in real time through the composite adaptive law by utilizing the regression matrix which is trained and fixed in the step 2 and the task parameter statistical characteristics obtained in the step 2, wherein the method specifically comprises the following steps of: Step 4) robust control instruction generation, namely substituting the updated online self-adaptive parameter estimation value as a compensation term into an ANDI control law based on nominal parameters to generate a final control instruction; And 5) judging the task state, namely judging whether the current flight task reaches a preset end point condition in real time, if not, jumping to the step 3 to continue to execute the online circulation, and if the end condition is met, ending the control process.
  3. 3. The method for learning and robust controlling actuator dynamics-considered aircraft according to claim 2, wherein the input affine nonlinear system model is a generalized differential equation system describing rigid dynamics and kinematics of the aircraft, and is constructed based on Newton-Euler equations and can be constructed through nonlinear functions Characterizing complex aerodynamic properties of an aircraft as a function of state in wide envelope flight: Wherein: Is that Maintaining a system state vector comprising an angle and an angular rate of the aircraft; Is that The actual control input vector is maintained, namely the actual rudder deflection angle or thrust vector deflection angle generated by the actuator; Is that The maintenance system outputs a vector, namely a controlled variable to be tracked; And The functions are smooth vector field functions respectively, and the inherent aerodynamic characteristics and inertial coupling characteristics of the aircraft are represented; The generalized high-order actuator dynamics model is characterized in that a physical hysteresis of a description executing mechanism aiming at dynamic response between an input instruction and actual output is established, and a high-order model is adopted: Wherein: representing actual control inputs An order time derivative; For the control command to be designed, the controller designs the nominal parameter matrix based on To characterize the physical characteristics of actuator bandwidth, damping ratio, etc.
  4. 4. -A method of learning a robust control for an aircraft element taking actuator dynamics into account according to claim 2 or 3, characterized in that said nominal parameter-based ANDI control law is implemented by a flight control computer performing the following calculations: i) State sensing and processing, namely collecting missile state vector in real time through a sensor Actual rudder deflection angle The second order sliding mode differentiator is adopted to conduct anti-noise filtering processing on the original data so as to smooth the high-frequency burr signals returned by the sensor, and the internal discretization recursive operation is adopted to ensure that the high-frequency burr signals are input to a subsequent control module And (3) with Equal signal and augmented state vector The signal has a higher signal-to-noise ratio; ii) nonlinear dynamic inverse calculation, namely mapping by algebraic analysis by utilizing a pneumatic parameter model preset by the missile Deployed to contain control instructions Control item and system inherent dynamic item of (2) Is defined by the analytical equation: wherein, control the efficiency matrix Gain matrix of actuator Pre-storing the pneumatic derivative in a non-volatile memory in a multi-dimensional lookup table form based on wind tunnel test data or a numerical simulation result, and determining the nominal bandwidth and damping coefficient based on the missile second-order steering engine physical identification parameter; iii) Error feedback adjustment and instruction calculation are carried out according to a reference model Synthesizing virtual control quantity with expected error attenuation characteristic Wherein: The parameter distribution of the constant matrix is preset based on the expected characteristic root position of the closed-loop system, so that the closed-loop error polynomial meets the stability requirement, and the constant matrix is hard coded in a memory in advance, and then the nominal control instruction is solved by utilizing an inverse dynamics operator through an analytic solution formula 。
  5. 5. The method for robust control of aircraft element learning taking actuator dynamics into consideration as set forth in claim 1, wherein the element learning is joint representation learning based on multi-task feature extraction, namely by constructing a multi-task data set covering multiple working conditions, extracting a cross-task commonality mapping relation by offline training a neural network, and taking the cross-task commonality mapping relation as a general structural feature basis function matrix Preset in the system so that the model mismatch difference between different tasks is only reflected in low-dimensional linear parameter vector In this way, the system is provided with the capability of rapidly adapting to new tasks in an online stage, in particular, the lumped uncertainty between the real system dynamics and the nominal model Expressed as regression matrix And unknown parameter vector Is the product of: wherein the state vector is augmented Including system state and its derivatives, control inputs and its derivatives; is a bounded approximation error.
  6. 6. The method for robust control of aircraft element learning taking actuator dynamics into account according to claim 2, wherein said step 2 specifically comprises: 2.1 Construction of a multitasking dataset, setting of aerodynamic parameter scaling factors And actuator characteristic parameters Is of the variation range of (1) Different flight task environments are acquired for the length of time of Is a sequence of flight states of (a) Corresponding true uncertainty tags Wherein: Constructing a multi-tasking dataset ; 2.2 Training the characteristic structural basis function network by constructing an MLP fully connected neural network as a structural characteristic basis function matrix Wherein: the network weight to be optimized; 2.3 Building a meta-learning joint loss function: Wherein: is the first Specific parameter vectors corresponding to the tasks; 2.4 Structured prior extraction and parameter storage, adopting a back propagation method and being based on joint loss function Updating network weights And each task parameter After training convergence, performing physical extraction operation, deriving weight parameters of a hidden layer of the MLP fully-connected neural network, and storing the weight parameters in a static storage area of a flight control computer in a static constant matrix form to form a structural feature basis function matrix capable of being directly invoked in an online stage Meanwhile, a specific parameter vector set after optimization of each training task is saved and extracted And calculating an arithmetic average value of the set, defining the arithmetic average value as a task parameter statistical characteristic, and storing the task parameter statistical characteristic into a data area of the flight control system for parameter hot start initialization in an online stage.
  7. 7. The method for robust control of aircraft element learning taking actuator dynamics into account according to claim 6, wherein said step 3 specifically comprises: 3.1 Parameter initialization-the aircraft operating mechanism directly calls the task parameter statistics defined in step 2.4 from the static data store as initial values for the online adaptive parameter estimates : The statistical characteristic is utilized to guide the initialization mapping of the parameter vector in the estimation space, the searching range of the parameter in the convergence process is reduced, and the control input buffeting caused by the initial estimation deviation is restrained; 3.2 Structured feature basis function matrix mapping, real-time acquisition of flight status and construction of augmentation vectors Inputting the structural characteristic basis function matrix trained in the step 2 In the method, a structural characteristic basis function matrix under the current working condition is obtained through function mapping ; 3.3 Compound error calculation, namely real-time calculation of compound error based on higher derivative deviation obtained by signal reconstruction in the step 1 and virtual control quantity deviation The solving result is used as a feedback adjustment item of the self-adaptive law and used for quantitatively representing transient deviation of the real physical dynamic response of the missile relative to the dynamic state of an ideal reference instruction; 3.4 Parameter online integral update, calling adaptive law with leakage term to calculate update rate of parameter estimated value, and at initial value Is updated on the basis of integration: Wherein: is a leakage coefficient for preventing parameter drift in the absence of excitation; An adaptive gain matrix is positively defined; the term corrects the parameter estimate using the current error direction and the structured feature basis function matrix.
  8. 8. The method for learning robust control of an aircraft element in consideration of actuator dynamics according to claim 7, wherein said step 4 specifically comprises: 4.1 Uncertainty real-time reconstruction, wherein the flight control computer performs online self-adaptive parameter estimation value updated according to the step 3.4 And stored structured feature basis function matrices By performing linear weighted product operation of vectors and matrixes, real-time reconstruction is performed to obtain a lumped uncertainty estimated value representing modeling deviation and external disturbance under the current flight working condition ; 4.2 Based on inverse dynamics principle, analyzing and substituting the lumped uncertainty estimated value in the nominal resolving architecture obtained in the step 1 And synthesizing a robust control instruction finally sent to the missile execution mechanism by counteracting mismatch errors between the nominal model and the real physical object: The instruction is dynamically compensated for model mismatch errors, so that high-precision track tracking of the missile under the perturbation working condition is realized.
  9. 9. The method for learning and robust control of the aircraft element taking actuator dynamics into consideration as claimed in claim 2, wherein the online circulation is realized by a real-time task system of an aircraft flight control computer, the execution period of the real-time task system is synchronous with the flight control main control law, and in each control period, a processor sequentially calls a stored neural network weight matrix, namely a structured characteristic basis function matrix and real-time adaptive law calculation logic, and the calculation cost of the real-time task system is required to meet the hard real-time constraint of the flight control system.
  10. 10. An aircraft element learning robust control system taking actuator dynamics into consideration for implementing the method of any one of claims 1-9, comprising a state sensing and processing module, a nonlinear dynamic inverse resolving module, a structural feature basis function matrix reasoning module, an on-line adaptive estimation module and a comprehensive instruction generating module, wherein the state sensing and processing module collects flight state data in real time, performs anti-noise filtering processing to suppress measurement noise, completes high-order differential calculation, and generates an augmented state vector comprising a system state, a control input and various derivatives thereof; the system comprises a nonlinear dynamic inverse calculation module, a structural characteristic base function matrix reasoning module, an online self-adaptive estimation module, a comprehensive instruction generation module and an integrated instruction generation module, wherein the nonlinear dynamic inverse calculation module calculates virtual control quantity and an ANDI control law based on nominal parameters according to preset nominal actuator parameters, a pneumatic model and expected tracks, the structural characteristic base function matrix reasoning module internally stores a neural network model which is trained through offline element learning and is fixed, the structural characteristic base function matrix which covers the mismatch and pneumatic perturbation characteristics of the actuator is output in real time according to an augmented state vector, the online self-adaptive estimation module constructs a composite error variable according to a high-order derivative, expected tracks and various tracking errors of the system output, the self-adaptive law with leakage items is utilized to update an uncertainty parameter vector in real time, the structural characteristic base function matrix is combined with the ANDI control law to calculate a real-time uncertainty compensation item, and the comprehensive instruction generation module is combined with the ANDI control law and the real-time uncertainty compensation item to solve a final control input instruction to drive an aircraft actuating mechanism through inverse operation, so that the real dynamic response capability of the actuator is matched when the control instruction is compensated for model mismatch.

Description

Aircraft element learning robust control method and system considering actuator dynamics Technical Field The invention relates to a technology in the field of flight control, in particular to an aircraft element learning robust control method and system considering actuator dynamics. Background The traditional nonlinear dynamic inverse control is highly dependent on the nominal model precision, ignores the dynamic characteristics of an actuator, and is easy to induce model mismatch and phase lag under the working conditions of aerodynamic perturbation and bandwidth limitation. The existing meta-learning method has the defects that the multi-emphasis black box compensation is slow in adaptive convergence aiming at parameter drift, and in addition, the structural characterization is lacking, so that the transient performance and the steady-state precision under the multi-task switching are difficult to be considered. Disclosure of Invention Aiming at the defects of phase lag caused by increment nonlinear dynamic inverse dependence time scale separation hypothesis, high sensitivity of inverse control (ANDI) with actuator dynamic to model precision, slow traditional self-adaptive convergence and the like in the prior art, the invention provides an aircraft element learning robust control method and system taking actuator dynamic into consideration, which overcome the problem of phase lag caused by time scale separation hypothesis failure through a high-order actuator dynamic model, and secondly, for actuator mismatch and pneumatic perturbation, a multi-task scene is built offline by utilizing element learning technology, a structural feature basis function matrix capable of representing the uncertainty common feature of the system is extracted and used as a structural priori for online estimation, and the lumped error caused by the inconsistency of a nominal model and a real object is estimated and compensated in real time by combining with a self-adaptive law based on a composite error. The invention is realized by the following technical scheme: The invention relates to an aircraft element learning robust control method considering actuator dynamics, which comprises the steps of firstly establishing a generalized aircraft nonlinear system model containing the actuator dynamics, carrying out high-order differential processing on model output, deriving an ANDI control law based on nominal parameters based on a nonlinear dynamic inverse principle, incorporating the actuator dynamics into an inverse control design to compensate response hysteresis caused by the bandwidth limitation of the actuator, constructing a multi-task data set in an offline stage, training a neural network by using a meta learning method to extract a universal structural feature basis function matrix, constructing a composite error variable containing a prediction error and a tracking error by using the offline training and fixed structural feature basis function matrix in an online flight stage, updating the uncertainty parameters in real time by adopting the adaptive law of a leakage term to estimate lumped uncertainty caused by the actuator parameter mismatch and pneumatic perturbation, and finally substituting the uncertainty compensation term obtained by online estimation into the ANDI control law based on the nominal parameters to generate a final robust control instruction and outputting the final robust control instruction to an aircraft actuator, thereby realizing stable control and robust tracking track of the aircraft under the conditions of uncertain parameters and limitation. The data set is generated by randomly setting aerodynamic parameter scaling factors (simulation environment or configuration change) and actuator characteristic parameters in a simulation environment and is used for representing various mismatch situations between a nominal model and a real aircraft. Technical effects In the nominal control law construction stage, the invention systematically eliminates modeling errors adopted by the traditional method due to ignoring actuator response lag from a control law layer in a nonlinear dynamic inverse control framework which is adopted by high-order actuator dynamics explicit, thereby effectively avoiding closed loop stability risks caused by time scale separation assumption failure. In uncertainty compensation, the invention adopts a general uncertainty regression matrix extracted by a meta-learning mechanism in an offline stage as a structured priori knowledge, and further designs a composite self-adaptive update law driven by the cooperation of a prediction error and a tracking error, thereby not only overcoming the problems of slow on-line convergence speed and poor transient performance of parameters in the traditional self-adaptive control, but also aiming at the key engineering problem of parameter mismatch of an actuator, realizing the rapid identification and on-line correction of dynamic response differ