CN106200383A - A kind of three axle Inertially-stabilizeplatform platform control method based on model reference adaptive neutral net - Google Patents

Abstract

A kind of three axle Inertially-stabilizeplatform platform control method based on model reference adaptive neutral net, relate to the Disturbance Rejection design of control parameter On-line Estimation based on adaptive neural network and expansion state state observer, first, according to three axle inertially stabilized platform kinetic models, for the three uncertain feedback control parameters time-varying characteristics caused of axle inertially stabilized platform model parameter, build adaptive neural network and feedback control matrix parameter is carried out On-line Estimation, make three axle inertially stabilized platform control accuracies approach expectational model control accuracy；Secondly, the interference compensation controlled quentity controlled variable that the upper bound disturbed for adaptive neural network estimation difference and three axle inertially stabilized platforms the builds impact on control accuracy, build extended state observer interference is estimated and suppresses, it is achieved three axle inertially stabilized platform high accuracy under complex environment control.The present invention has that real-time is good, dynamic parameter response is fast, to advantages such as multi-source interference strong adaptabilities, can be used for the high accuracy control etc. under complicated multi-source interference environment of the three axle inertially stabilized platforms.

Description

Three-axis inertially stabilized platform control method based on model reference adaptive neural network

Technical Field

The invention relates to a three-axis inertial stabilization platform control method based on a model reference adaptive neural network, which is suitable for the field of high-precision control of an aerial surveying and mapping stabilization platform.

Background

The three-axis pod platform is fixedly connected to the flight carrier through the base, supports and stabilizes the remote sensing imaging load, isolates the influence of multi-source interference on the visual axis of the remote sensing imaging load, improves the pointing accuracy of the remote sensing load to the ground, and has wide application prospect.

The three-axis inertially stabilized platform has various interference types in the working process, unpredictable random wind interference and angular motion interference caused by the vibration of an airplane body exist, unbalanced moment caused by misalignment of the center of mass of a load system and the origin of the system also exists, and internal interference caused by measurement error of a platform sensor device exists, so that a high-precision control method in a complex multi-interference environment becomes one of key technologies for research of the three-axis inertially stabilized platform.

In order to improve the performance, a PID control method, a robust control method and an intelligent control method are used for high-precision control of the triaxial inertially-stabilized platform. The PID controller has a simple structure, but has poor anti-interference capability, and the stability and the precision of the triaxial inertia platform under the action of multi-source interference are difficult to ensure. The robust control can inhibit the influence of uncertainty of system model parameters and multi-source interference on the control precision, but the control precision has larger conservative property. Through a large amount of sample training, the neural network can approach a nonlinear system infinitely, so that the nonlinearity of a triaxial inertially stabilized platform model is solved, high-precision attitude control is realized, but the traditional neural network needs a large amount of sample data for training and has the defect of poor real-time performance.

Disclosure of Invention

The technical problem solved by the invention is as follows: the control precision of the triaxial inertially stabilized platform is influenced by the nonlinearity of a system model and multisource interference, the nonlinearity of the model is solved by estimating feedback control parameters in real time through a self-adaptive neural network, an extended observer is constructed to improve the anti-interference capability of the system, and the high-precision control of the triaxial inertially stabilized platform in a complex environment is realized.

The technical solution of the invention is as follows: firstly, establishing a dynamic model for a triaxial inertial stabilization platform, establishing a triaxial inertial stabilization platform reference dynamic model according to expected system performance indexes, and enabling the triaxial inertial stabilization platform dynamic model to approach a reference model through an adaptive network online track feedback control parameter; and secondly, constructing an extended observer to reduce the influence of the estimation error of the adaptive neural network and the symbolic function gain constructed by the upper bound of external disturbance on the control precision of the system. The method comprises the following implementation steps:

(1) designing a model reference adaptive neural network control method according to a three-axis inertially stabilized platform dynamic model, and constructing an adaptive neural network to perform online estimation on feedback control matrix parameters aiming at the time-varying characteristic of the feedback control parameters caused by uncertain parameters of the three-axis inertially stabilized platform model so as to enable the control precision of the three-axis inertially stabilized platform to approach the control precision of an expected model;

(2) aiming at the influence of the interference compensation control quantity constructed by the self-adaptive neural network estimation error and the upper bound of the triaxial inertially stabilized platform interference on the system control precision, an extended state observer is constructed to estimate and inhibit the interference, so that the triaxial inertially stabilized platform high-precision control under the complex environment is realized;

the invention discloses a triaxial inertially stabilized platform control method based on a model reference adaptive neural network, wherein the step (1) is based on a control input u controlled by the model reference adaptive neural network_jAdaptive neural network update lawAnd a disturbance compensation control amount r_jAre respectively expressed as

$u_j={\hat{W}}_{j*}\·\mathrm{\Θ}\left(x\right)x+u_j^{mm}+r_j-\frac{g_j}{b_j},j=1,2,3$

${\stackrel{\·}{\hat{W}}}_{j*}=-e^T{\mathrm{PB}}_{*j}{x}^T{\mathrm{\Θ}}^T\left(x\right)\mathrm{\Γ}$

r_j＝-_jsgn(e^TPB_*j)

Wherein, j equals 1 to represent a roll frame, j equals 2 to represent a pitch frame, j equals 3 to represent an azimuth frame,is a weight matrix of the adaptive neural network,is an ideal weight matrix W of the adaptive neural network^*The real-time estimate of the jth line of (1),is the basis function of the adaptive neural network, n is the dimension of the system model, m is the dimension of the system input, l is the hidden dimension of the adaptive neural networkThe number of nodes comprising a layer is,is a state variable of the system, b_jThe input coefficient matrix B for the frame corresponds to the ideal control coefficient of j columns, g_jIs an estimate of the inertial inertia of the system j frame, u_j ^mmIs the expected input to the framework of a reference model j of the system, the reference model being

${\stackrel{\·}{x}}^{mm}=A^{mm}{x}^{mm}+{\mathrm{Bu}}^{mm}$

Wherein,to refer to the matrix of state equations of the model,andis a matrix artificially designed according to the dynamic performance requirement of the triaxial inertially stabilized platform, wherein k is 1,2 and 3 respectively represent a corresponding roll frame, a pitch frame and an azimuth frame, and e is a current state quantity x and a state variable x of a reference model^mmIs a positive definite symmetric solution of the equation of state,

PA^mm+A^mmTP＝-Q

whereinIs a positive definite symmetrical array, and the array is,is the adaptive law gain matrix, l is the number of nodes in the hidden layer of the adaptive neural network, B_*jIs the jth column of the coefficient matrix B of the frame input, and the upper bound of the estimation error and the external disturbance of the adaptive neural network is

_j>|_jx+d_j/b_j|

Wherein,_jis the approximation error of the adaptive neural network, d_jInterference to the system j framework;

the invention relates to a triaxial inertia stable platform control method based on a model reference adaptive neural network, wherein the extended state observer constructed in the step (2) and a corrected interference compensation control quantity r_jAre respectively expressed as

${\stackrel{\·}{\hat{z}}}_1=A^m{z}_1-B\stackrel{\·}{\hat{W}}\mathrm{\Θ}\left(x\right)x+Br+{\hat{z}}_2-2w_0({\hat{z}}_1-z_1)$

${\stackrel{\·}{\hat{z}}}_2=-w_0^2({\hat{z}}_1-z_1)$

$r_j=-\frac{{\left({\hat{z}}_2\right)}_j}{b_j}-\frac{{\mathrm{\ξ}}_j\mathrm{sgn}\left(e^T{\mathrm{PB}}_{*j}\right)}{b_j}$

Wherein z is₁＝e，z₂＝Δ＝Bx+Dd，Andare respectively a state variable z₁And z₂Estimated value of, w₀>0 is a design variable which can beAndclosely tracking e and delta within a limited time, modified disturbance compensation control quantity r_jIs expressed as

$r_j=-\frac{{\left({\hat{z}}_2\right)}_{j+3}}{b_j}-\frac{{\mathrm{\ξ}}_j\mathrm{sgn}\left(e^T{\mathrm{PB}}_{*j}\right)}{b_j}$

Wherein,estimating residual Bx, system interference Dd and disturbance estimate z for a system adaptive neural network₂Residual error of

The corresponding system input is

$u_j={({\hat{W}}_j^T\·{\mathrm{\Θ}}_j)}^{T}x+u_j^m-\frac{g_j}{b_j}-\frac{{\left({\hat{z}}_2\right)}_{j+3}}{b_j}-\frac{{\mathrm{\δ}}_j\mathrm{sgn}\left(e^T{\mathrm{PB}}_{*j}\right)}{b_j},j=1,2,3,$

Wherein,is thatThe j +3 th component.

Compared with the prior art, the invention has the advantages that:

(1) according to the method, the feedback control parameters of the system are estimated in real time through the infinite generalization approximation capability of the adaptive neural network, the problem of time varying nonlinearity of the feedback control parameters caused by nonlinearity and uncertainty of the model parameters is solved, the actual model approximates to the reference model, and the control method is simple in structure and high in anti-interference capability;

(2) under the condition that the self-adaptive neural network ensures the stability of the system, the disturbance borne by the triaxial inertially stabilized platform in the working process is further estimated and suppressed by using the extended observer, the control precision is high, and the high-precision control requirement of the triaxial inertially stabilized platform can be met;

(3) according to the method, the weight value of the adaptive neural network can be updated on line only by designing the weight value updating matrix of the adaptive neural network by utilizing the Lyapunov function according to the state information of the triaxial inertially stabilized platform in the working process, and the method does not need any sample training and has the advantages of convenience in data acquisition and simplicity in calculation.

Drawings

FIG. 1 is a three-axis inertially stabilized platform control flow;

FIG. 2 shows the control effect of the pitching channel of the three-axis inertially stabilized platform in a flight experiment;

FIG. 3 shows the control effect of the roll channel of the three-axis inertially stabilized platform in a flight experiment;

FIG. 4 shows the control effect of the three-axis inertially stabilized platform azimuth channel in the flight experiment.

Detailed Description

As shown in FIG. 1, the present invention is embodied as follows

(1) Construction of model-based reference adaptive neural network

Based on the Newton-Euler equation, the dynamic equation of the triaxial inertially stabilized platform is expressed as

$\stackrel{\·}{x}=Ax+Bu+D(g+d)$

Wherein x is [ theta ]_jω_j]^T，H＝0_3×3，F＝(f_jk),j,k＝1,2,3，

u＝[u₁u₂u₃]^T，g＝[g₁g₂g₃]^T，d＝[d₁d₂d₃]^T，

Wherein, j equals 1 to represent a roll frame, j equals 2 to represent a pitch frame, j equals 3 to represent an azimuth frame,for the state variable of the system, n-6 is the dimension of the state variable, θ_jFor corresponding j frame angle, ω_jFor the corresponding j-frame angular velocity,is a matrix of coefficients for the state variables,coefficient matrix input for frame, m-3 is dimension of system input, u_jFor the corresponding j frame voltage input, g is the estimated value of the frame moment of inertia, d is the system frame disturbance, perturbed value △ g of the system disturbance, frame moment of inertia_jAnd control input perturbation Δ b caused by measurement noise_jConstitution b_jThe coefficient matrix B input for the frame corresponds to the ideal control coefficients of j columns, and F is the corresponding ideal state variable coefficient matrix, wherein

$f_{11}=-\frac{K_t{K}_e{N}^2}{J_2{R}_m},f_{12}=-\frac{(J_{pz}-J_{py}-J_{ay})({\mathrm{\ω}}_{irz}^{r}cos2{\mathrm{\θ}}_p-{\mathrm{\ω}}_{iry}^r{\mathrm{sin\θ}}_p{\mathrm{cos\θ}}_p)}{J_2},$

$f_{13}=\frac{J_{az}{\mathrm{\ω}}_{ipy}^p}{J_2},f_{21}=-\frac{(J_{ax}+J_{px}){\mathrm{\ω}}_{irz}^r}{J_3},f_{23}=\frac{J_{az}{\mathrm{\ω}}_{irx}^r{\mathrm{cos\θ}}_p}{J_3},$

$f_{22}=-(\frac{K_t{K}_e{N}^2}{J_3{R}_m}+\frac{2J_{az}{\stackrel{\·}{\mathrm{\θ}}}_p{\mathrm{cos\θ}}_p{\mathrm{sin\θ}}_p+(J_{pz}-J_{ay}-J_{py})({\mathrm{\ω}}_{irx}^r+2{\stackrel{\·}{\mathrm{\θ}}}_p){\mathrm{cos\θ}}_p{\mathrm{sin\θ}}_p}{J_3}),$

$f_{31}=f_{32}=0,f_{33}=-\frac{K_t{K}_e{N}^2}{J_1{R}_m};$

${\stackrel{\‾}{b}}_1=\frac{{\mathrm{NK}}_t}{J_2{R}_m},{\stackrel{\‾}{b}}_2=\frac{{\mathrm{NK}}_t}{J_3{R}_m},{\stackrel{\‾}{b}}_3=\frac{{\mathrm{NK}}_t}{J_1{R}_m};$

${\stackrel{\‾}{g}}_1=\frac{K_t{K}_e{N}^2{\mathrm{\ω}}_{ibx}^p+N(N-1)R_m{J}_m{\stackrel{\·}{\mathrm{\ω}}}_{ibx}^p}{J_2{R}_m},$

${\stackrel{\‾}{g}}_2=\frac{K_t{K}_e{N}^2{\mathrm{\ω}}_{iby}^r+N(N-1)R_m{J}_m{\stackrel{\·}{\mathrm{\ω}}}_{iby}^t}{J_3{R}_m},{\stackrel{\‾}{g}}_3=\frac{K_t{K}_e{N}^2{\mathrm{\ω}}_{ibz}^a+N(N-1)R_m{J}_m{\stackrel{\·}{\mathrm{\ω}}}_{ibz}^a}{J_1{R}_m};$

${\stackrel{\‾}{d}}_1=\frac{-(J_{pz}-J_{py}-J_{ay}){{\mathrm{\ω}}_{irz}^r}^2{\mathrm{sin\θ}}_p{\mathrm{cos\θ}}_p+{\mathrm{NT}}_{dm}+T_{dp}}{J_2},$

$\begin{array}{c}{\stackrel{\‾}{d}}_2=\frac{(J_{ay}+J_{py}){\sin }^2{\mathrm{\θ}}_p+J_{pz}{\cos }^2{\mathrm{\θ}}_p-J_{rx}+J_{rz}}{J_3}{\mathrm{\ω}}_{irz}^r{\mathrm{\ω}}_{irx}^r\\ +\frac{{\mathrm{NT}}_{dm}+T_{dr}-{\[(J_{ay}+J_{py}-J_{az}-J_{pz}){\mathrm{\ω}}_{irz}^r{\mathrm{cos\θ}}_p{\mathrm{sin\θ}}_p-{\mathrm{sin\θ}}_p{J}_{az}{\stackrel{\·}{\mathrm{\θ}}}_a\]}^{\′}}{J_3}\end{array},$

${\stackrel{\‾}{d}}_3=\frac{{\mathrm{NT}}_{dm}+T_{da}}{J_1}$

Wherein N represents the motor transmission ratio, K_eRepresents the back electromotive force constant, K_tRepresenting the motor moment coefficient, R_mDenotes the motor resistance, J_mRepresenting the moment of inertia of the motor, J_a＝diag(J_ax,J_ay,J_az) Is the projection of the moment of inertia of the orientation frame in the x, y, z directions of the orientation frame coordinate system, J_p＝diag(J_px,J_py,J_pz) Is the projection of the moment of inertia of the pitch frame in the x, y, z direction of the pitch frame coordinate system, J_r＝diag(J_rx,J_ry,J_rz) Is the projection of the rotational inertia of the roll frame on the x, y and z directions of the roll frame in a roll coordinate system,k is a mapping of the angular velocity of k relative to the inertial space in the directions of k x, y and z;respectively, the roll frame is in roll coordinate system relative to the base, and the pitch frame is in pitch relative to the roll frameCoordinate system, angular velocity of azimuth frame relative to pitch frame in azimuth coordinate system, theta_rIndicating the angle of rotation of the carriage relative to the base, theta_pRepresenting the angle of rotation, theta, of the pitch frame relative to the roll frame_aIndicating the angle of rotation of the azimuth frame relative to the pitch frame, measured by a code wheel mounted on the frame shaft, T_dmFor disturbing torque acting on the motor, T_djFor disturbing moments acting on the j-frame, the moment of inertia is

J₁＝J_az+N²J_m，

J₂＝J_px+J_ax+N²J_m，

J₃＝J_ry+(J_ay+J_py)cos²θ_p+(J_az+J_pz)sin²θ_p+N²J_m

Designing a reference model

${\stackrel{\·}{x}}^{mm}=A^{mm}{x}^{mm}+{\mathrm{Bu}}^{mm}$

WhereinTo refer to the matrix of state equations of the model,andis a matrix artificially designed based on the pole allocation principle according to the dynamic performance requirement of the system,is the control input to the ideal reference model,

designing a feedback controller

$u_j^{*}=K_j^{*T}x+u_j^{mm}+r_j-\frac{g_j}{b_j},j=1,2,3$

Wherein r ═ r₁r₂r₃]^TIs a control quantity designed for compensating the disturbance d, and satisfies Is feedback gain, then the system

$\stackrel{\·}{x}=(A+BK)x+{\mathrm{Bu}}^{mm}+Br+Dd$

If A + BK is equal to A^mmThen the actual system approaches the ideal reference model infinitely;

thus, the feedback gain

$K_{j*}^{*}=\frac{1}{b_j}(\left[\begin{array}{cc}{h}_{j*}^{mm}& f_{j*}^{mm}\end{array}\right]-\left[\begin{array}{cc}{h}_{j*}& f_{j*}\end{array}\right])$

Due to h_jk,f_jkK, j is 1,2,3 is a time-varying nonlinear function,is also a time-varying nonlinear function, therefore, the infinite generalization approximation capability of the adaptive neural network is utilized to estimate the time-varying nonlinear function on line

$K_{j*}^{*}=W_{j*}^{*}\·\mathrm{\Θ}\left(x\right)+{\mathrm{\ϵ}}_j,j=1,2,3$

Wherein,is a weight matrix of the adaptive neural network,is an ideal weight matrix of the adaptive neural networkThe real-time estimate of the jth line of (1),is the basis function of the adaptive neural network, is the number of nodes of the hidden layer of the adaptive neural network,_jis the approximation error of the adaptive neural network;

thus, the control input u based on model reference adaptive neural network control_jAdaptive neural network update lawAnd a disturbance compensation control amount r_jAre respectively expressed as

$u_j={\hat{W}}_{j*}\·\mathrm{\Θ}\left(x\right)x+u_j^{mm}+r_j-\frac{g_j}{b_j}$

${\stackrel{\·}{\hat{W}}}_{j*}=-e^T{\mathrm{PB}}_{*j}{x}^T{\mathrm{\Θ}}^T\left(x\right)\mathrm{\Γ}$

r_j＝-_jsgn(e^TPB_*j)

Wherein e is the current state quantity x and the state variable x of the expected model^mmIs a positive definite symmetric solution of the equation of state,

PA^mm+A^mmTP＝-Q

whereinIs a positive definite symmetrical array, and the array is,is an adaptive law gain matrix, B_*jIs the jth column of the B matrix, and the upper bound of the estimation error and the external disturbance of the adaptive neural network is

_j>|_jx+d_j/b_j|

Wherein,_jis the approximation error of the adaptive neural network, d_jInterference to the system j framework;

(2) constructing an adaptive neural network

Interference compensation control quantity r constructed aiming at upper bound of adaptive neural network estimation error and external disturbance_jThe influence on the control precision of the system, the extended state observer is constructed to estimate and inhibit the interference, the high-precision control of the triaxial inertially stabilized platform under the complex environment is realized,

defining a state variable z₁＝e，z₂＝Δ＝Bx+Dd，

Then the three-axis inertially stabilized platform error state equation is

${\stackrel{\·}{z}}_1=A^m{z}_1-B\stackrel{\·}{\hat{W}}\·\mathrm{\Θ}\left(x\right)x+Br+z_2$

Will z₂Introduces an error equation of state, an

Wherein s is the rate of change of Δ, and the extended state observer is constructed as

${\stackrel{\·}{\hat{z}}}_1=A^m{z}_1-B\stackrel{\·}{\hat{W}}\mathrm{\Θ}\left(x\right)x+Br+{\hat{z}}_2-2w_0({\hat{z}}_1-z_1)$

${\stackrel{\·}{\hat{z}}}_2=-w_0^2({\hat{z}}_1-z_1)$

Wherein,andare respectively a state variable z₁And z₂Estimated value of, w₀>0 is a design variable;

by selecting the parameter w appropriately₀Can make it possible toAndclosely tracking e and delta within a limited time,

corrected disturbance compensation control amount r_jIs expressed as

$r_j=-\frac{{\left({\hat{z}}_2\right)}_j}{b_j}-\frac{{\mathrm{\ξ}}_j\mathrm{sgn}\left(e^T{\mathrm{PB}}_{*j}\right)}{b_j}$

Wherein,estimating residual Bx, system interference Dd and disturbance estimate z for a system adaptive neural network₂Residual error of

The corresponding system input is

$u_j={({\hat{W}}_j^T\·{\mathrm{\Θ}}_j)}^{T}x+u_j^m-\frac{g_j}{b_j}-\frac{{\left({\hat{z}}_2\right)}_{j+3}}{b_j}-\frac{{\mathrm{\δ}}_j\mathrm{sgn}\left(e^T{\mathrm{PB}}_{*j}\right)}{b_j},j=1,2,3$

Wherein,is thatThe j +3 th component.

(3) Flight example

In the flight process, according to the angle information of the high-precision attitude measurement unit, the three-axis inertially stabilized platform frame system is correspondingly adjusted to ensure that the remote sensing load visual axis is vertical to the ground, and the flight result of a certain experiment is shown in fig. 2,3 and 4.

The triaxial inertially stabilized platform realizes high-precision control, the standard deviation of a pitching channel is 0.0183 degrees, the standard deviation of a rolling channel is 0.0157, and the standard deviation of an azimuth channel is 0.0214.

The three-axis inertially stabilized platform control method based on the model reference adaptive neural network overcomes the defects of the existing control method, and can realize high-precision control of the three-axis inertially stabilized platform in a complex and multi-disturbance environment.

Those skilled in the art will appreciate that the invention may be practiced without these specific details.

Claims (3)

1. A three-axis inertial stabilization platform control method based on a model reference adaptive neural network is characterized by comprising the following steps:

(1) designing a model reference adaptive neural network control method according to a three-axis inertially stabilized platform dynamic model, and constructing an adaptive neural network to perform online estimation on feedback control matrix parameters aiming at the time-varying characteristic of the feedback control parameters caused by uncertain parameters of the three-axis inertially stabilized platform model so as to enable the control precision of the three-axis inertially stabilized platform to approach the control precision of an expected model;

(2) aiming at the influence of the interference compensation control quantity constructed by the adaptive neural network estimation error and the upper bound of the triaxial inertially stabilized platform interference on the system control precision, an extended state observer is constructed to estimate and inhibit the interference, and the triaxial inertially stabilized platform high-precision control under the complex environment is realized.

2. The method for controlling the three-axis inertially stabilized platform based on the model-reference adaptive neural network according to claim 1, wherein: the step (1) is based on the control input u of model reference adaptive neural network control_jAdaptive neural network update lawAnd a disturbance compensation control amount r_jAre respectively expressed as

r_j＝-_jsgn(e^TPB_*j)

Wherein, j equals 1 to represent a roll frame, j equals 2 to represent a pitch frame, j equals 3 to represent an azimuth frame,is a weight matrix of the adaptive neural network,is an ideal weight matrix W of the adaptive neural network^*The real-time estimate of the jth line of (1),is the basis function of the adaptive neural network, m is the dimension of the system input, l is the hidden layer of the adaptive neural networkN is the dimension of the system model,is a state variable of the system, b_jThe input coefficient matrix B for the frame corresponds to the ideal control coefficient of j columns, g_jIs an estimate of the inertial inertia of the system j frame, u_j ^mmIs the expected input to the framework of a reference model j of the system, the reference model being

Wherein,to refer to the matrix of state equations of the model,andis a matrix artificially designed according to the dynamic performance requirement of the triaxial inertially stabilized platform, wherein k is 1,2 and 3 respectively represent a corresponding roll frame, a pitch frame and an azimuth frame, and e is a current state quantity x and a state variable x of a reference model^mmIs a positive definite symmetric solution of the equation of state,

PA^mm+A^mmTP＝-Q

whereinIs a positive definite symmetrical array, and the array is,is the adaptive law gain matrix, l is the number of nodes in the hidden layer of the adaptive neural network, B_*jIs the jth column of the coefficient matrix B of the frame input, and the upper bound of the estimation error and the external disturbance of the adaptive neural network is

_j>|_jx+d_j/b_j|

Wherein,_jis the approximation error of the adaptive neural network, d_jIs a disturbance of the system j framework.

3. The method for controlling the three-axis inertially stabilized platform based on the model-reference adaptive neural network according to claim 1, wherein: the extended state observer constructed in the step (2) and the corrected interference compensation control quantity r_jAre respectively expressed as

Wherein z is₁＝e，z₂＝Δ＝Bx+Dd，Andare respectively a state variable z₁And z₂Estimated value of, w₀>0 is a design variable which can beAndclosely tracking e and delta within a limited time, modified disturbance compensation control quantity r_jIs expressed as

Wherein,estimating residual Bx, system interference Dd and disturbance estimate z for a system adaptive neural network₂Residual error of

The corresponding system input is

Wherein,is thatThe j +3 th component.