CN103412488A - Small-sized unmanned rotary-wing aircraft high-precision control method based on adaptive neural network - Google Patents

Abstract

The invention discloses a small-sized unmanned rotary-wing aircraft high-precision control method based on an adaptive neural network, and relates to the design of a composite controller with the combination of the construction and optimization of the feedback control and non-sample training adaptive neural network of a small-sized unmanned rotary-wing aircraft. Firstly, as for a small-sized unmanned rotary-wing aircraft dynamical model, a feedback control coefficient matrix is constructed through a pole assignment method to ensure the preliminary stability of a system. Secondly, the adaptive neural network with independent updating weight features is designed, an adaptive network weight updating matrix is constructed based on error messages to update a weight matrix of the neural network in an online mode, and estimation and restraint of disturbance are achieved. An adaptive threshold value optimizing strategy is designed, online updating is carried out on a control residual error upper limit threshold value of the adaptive neural network on the basis of the mean square error between the actual position and the expectation position in a time window, and the small-sized unmanned rotary-wing aircraft high-precision attitude control under the complex environment is achieved. The small-sized unmanned rotary-wing aircraft high-precision control method has the advantages of being good in real-time performance, fast in dynamic parameter response, strong in multi-source interference adaptability and the like, can be used for the high-precision control over the small-sized unmanned aircraft under the complex multi-source interference environment and the like.

Description

A kind of giro of miniature self-service based on adaptive neural network high-accuracy control method

Technical field

The present invention relates to a kind of giro of miniature self-service based on adaptive neural network high-accuracy control method, the skyborne unmanned robot autonomous control field of the work that is applicable to.

Background technology

The characteristics such as the miniature self-service giro has vertical takeoff and landing, hover, by self-contained various kinds of sensors can be in hazardous location or the narrow space such as drive carry out the tasks such as observation, information, be with a wide range of applications.Along with the expansion of application, the working environment of miniature self-service giro is also complicated and changeable, and the miniature self-service giro high precision that vulnerability to jamming is strong, stability is high is controlled the focus that becomes research.

As complicated multi-input multi-output control system, that the miniature self-service giro has is non-linear, strong coupling, control difficulty characteristic.And the miniature self-service giro exists multiclass to disturb in flight course, as wind disturb, atmospheric turbulence, ground interferences, system electromagnetic interference (EMI) etc., therefore, the high precision control of miniature self-service giro under disturbance is one of gordian technique of flight control system.

For improving performance, all kinds of control methods such as intelligent PID control method, robust control, intelligent control method are used to the flight of miniature self-service giro and control.Intelligent PID controller is simple in structure, but poor anti jamming capability, the control performance of miniature self-service giro is easy to be subject to the external interference impact and reduces.Robust control can be eliminated the miniature self-service giro exists in flight course model parameter out of true and external interference problem preferably, and real-time is poor, the characteristic of dynamic parameter low-response but robust control has.By a large amount of sample trainings, neural network can realize nonlinear adaptive control, overcome the model uncertainty that the miniature self-service giro has, and there are problems such as multi-source interference, realize high-precision attitude control, but traditional neural network needs a large amount of sample datas to train, and has the poor shortcoming of real-time.

Summary of the invention

Technology of the present invention is dealt with problems and is: for miniature self-service giro control performance when executing the task easily be subject to external interference impact problem, a kind of composite control method combined based on adaptive neural network and pole-assignment is proposed, to the miniature self-service giro awing suffered multi-source disturb and estimate and suppress, realize the high precision control of large envelope scope.

Technical solution of the present invention is: for miniature self-service giro kinetic model, build by pole-assignment the primary stability that the feedback control coefficient matrix guarantees system; The adaptive neural network that design has autonomous renewal weights characteristic, build based on control information the weight matrix that adaptive network right value update matrix carrys out the online updating neural network, realizes estimation and inhibition to disturbance; And design adaptive threshold optimisation strategy, physical location in the time-based window and the error mean square of desired locations are poor, control residual error upper limit threshold to adaptive neural network carries out online updating, realizes that the miniature self-service giro high-precision attitude under complex environment is controlled.Implementation step is as follows:

(1) for miniature self-service giro kinetic model, build by pole-assignment the primary stability that the feedback control coefficient matrix guarantees system;

(2) multi-source existed is in-flight disturbed, the adaptive neural network that design has autonomous renewal weights characteristic, based on control information, build the weight matrix of adaptive network right value update matrix online updating neural network, realize that On-line Estimation and inhibition are carried out in suffered multi-source interference awing to the miniature self-service giro;

(3) design adaptive threshold optimisation strategy, physical location in the time-based window and the error mean square of desired locations are poor, control residual error upper limit threshold to adaptive neural network carries out online updating, realizes that the miniature self-service giro high-precision attitude under complex environment is controlled.

Miniature self-service based on adaptive neural network giro high-accuracy control method of the present invention, wherein said step (2) structure adaptive neural network right value update matrix and disturbance controlled quentity controlled variable expression formula are as follows:

${\stackrel{\·}{\hat{W}}}_{*i}={-\mathrm{\Γ}}_{i}s\left(e\right)e^T{P}_{*i}$

$\widehat{\mathrm{\ω}}=B^T{\left({\mathrm{BB}}^T\right)}^{-l}{\hat{W}}^{T}s\left(e\right)-{\mathrm{\αB}}^T{\left({\mathrm{BB}}^T\right)}^{-l}\mathrm{sign}\left(e^T{P}_{*i}\right)$

Wherein,

For the weight matrix of adaptive neural network,

Disturbance controlled quentity controlled variable for adaptive neural network; Γ _i, P is symmetric positive definite matrix, the input e=x-x of adaptive neural network _dFor expectation state variable x _dAnd the error between virtual condition variable x, B is miniature self-service giro state of a control transition matrix, α is the control residual error upper limit threshold of adaptive neural network, i* is i row vector of corresponding matrix, * i is i column vector of corresponding matrix, s (e) is the node function of adaptive neural network hidden layer, is defined as Gaussian function, and the node function expression of its corresponding j hidden layer is as follows:

$s_j\left(e\right)=\exp \left(\frac{{-(e-{\mathrm{\μ}}_j)}^T(e-{\mathrm{\μ}}_j)}{{\mathrm{\σ}}_j^2}\right),j=\mathrm{1,2},\·\·\·l$

Wherein, μ _j,

Be respectively central value and the width of adaptive neural network hidden layer Gaussian function, l is the implicit nodes of adaptive neural network hidden layer;

Miniature self-service based on adaptive neural network giro high-accuracy control method of the present invention, wherein said step (3) builds the adaptive threshold optimisation strategy and is defined as follows

${\mathrm{\&alpha;}}_w=\left\{\begin{array}{cc}\frac{{\mathrm{\&alpha;}}_{w-1}}{{k_1}^2}& \underset{i=k-t}{\overset{k}{\max }}\left|{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_1\mathrm{and}\left|\frac{1}{t}\underset{i=k-t}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_2\mathrm{and}{\mathrm{\&chi;}}_k>{\mathrm{\&chi;}}_{k-t}\mathrm{andflag}=1\\ k_1{\mathrm{\&alpha;}}_{w-1}& \underset{i=k-t}{\overset{k}{\max }}\left|{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_1\mathrm{and}\left|\frac{1}{t}\underset{i=k-t}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_2\mathrm{and}{\mathrm{\&chi;}}_k<{\mathrm{\&chi;}}_{k-t}\mathrm{andflag}=1\\ k_1{\mathrm{\&alpha;}}_{w-1}& \underset{i=k-t}{\overset{k}{\max }}\left|{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_1\mathrm{and}\left|\frac{1}{t}\underset{i=k-t}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_2\\ {\mathrm{\&alpha;}}_{w-1}& \mathrm{otherwise}\end{array}\right.$

$\mathrm{flag}=\left\{\begin{array}{cc}1& {\mathrm{\&alpha;}}_{w-1}\&NotEqual;{\mathrm{\&alpha;}}_{w-2}\\ 0& {\mathrm{\&alpha;}}_{w-1}={\mathrm{\&alpha;}}_{w-2}\end{array}\right.$

${\mathrm{\&chi;}}_k=\sqrt{\underset{m=k-t}{\overset{k}{\mathrm{\&Sigma;}}}{({\mathrm{xx}}_m-{\mathrm{xx}}_d)}^2/t}$

α in formula _wControl residual error upper limit threshold for current time window sample adaptive neural network in the cycle; α _W-1Control residual error upper limit threshold for upper time window sample adaptive neural network in the cycle; α _W-2Control residual error upper limit threshold for upper two time windows adaptive neural network in the sampling period; χ _kFor upper time window sample miniature self-service giro physical location xx in the cycle _mWith desired locations xx _dMean square deviation; χ _k-tFor upper two time windows miniature self-service giro physical location xx in the sampling period _mWith desired locations xx _dMean square deviation; Ee _iFor time window miniature self-service giro physical location xx in the sampling period _mWith desired locations xx _dDifference; T is sampling number in the time window sampling period; k ₁For controlling parameter, η ₁, η ₂Be respectively miniature self-service giro physical location xx in the time window sampling period _mWith desired locations xx _dMaximum absolute error value and average error amount.

The present invention's advantage compared with prior art is:

(1) the present invention guarantees on the basis of primary stability of system by pole-assignment, building the feedback control coefficient matrix, the adaptive neural network of training by no specimen, suffered disturbance in flight course is estimated and is suppressed to the miniature self-service giro, has advantages of strong interference immunity and is convenient to design;

(2) the present invention is in the situation that traditional FEEDBACK CONTROL guarantees system stability, further utilize adaptive neural network to estimate and suppress the suffered disturbance in flight course of miniature self-service giro, can according to the status information of aircraft, adjust the rudder amount in real time, not only have simple in structure and control characteristic easily,, dynamic parameter good with the period control method real-time responds fast, can meet high precision demand for control under miniature self-service giro load environment;

(3) the present invention only needs the status information gathered according in miniature self-service giro practical flight process, based on resolving the site error obtained, weights that just can the online updating adaptive neural network, without any need for sample training, have data acquisition convenient, calculate simple advantage.

The accompanying drawing explanation

Fig. 1 is the autonomous control flow of miniature self-service giro;

Fig. 2 is that 3.2m/s wind is disturbed under environment and utilized miniature self-service giro of the present invention to carry out four destination cruising flight effects.

Embodiment

As shown in Figure 1, concrete methods of realizing of the present invention is as follows:

(1) based on the FEEDBACK CONTROL of POLE PLACEMENT USING

Based on linearization technique, miniature self-service giro kinetics equation is expressed as

$\stackrel{\&CenterDot;}{x}\left(t\right)=\mathrm{Ax}\left(t\right)+\mathrm{Bu}\left(t\right)+d\left(t\right)$

Wherein, state variable x ∈ R ⁿMean the corresponding speed of miniature self-service giro system, angle and angular velocity information.Control variable u ∈ R ^mRepresent respectively miniature self-service giro side direction feathering, vertically feathering, always apart from control signal and course control signal; A ∈ R ^{N * n}With B ∈ R ^{N * m}Be respectively state-transition matrix and the controls metastasis matrix of state variable and control variable; D ∈ R ^mExpression by wind disturb, atmospheric turbulence, ground interference, system electromagnetic interference (EMI), sensor measurement error, and bounded composite interference uncertain by miniature self-service giro systematic parameter and that bring without factors such as mode dynamicss.

The controller input of miniature self-service giro consists of two parts, and a part is for being feedback of status input Kx (t), and another part is adaptive neural network output Be

$u\left(t\right)=\widehat{\mathrm{\&omega;}}\left(t\right)+\mathrm{Kx}\left(t\right)$

Wherein, feedback factor K obtains according to POLE PLACEMENT USING is theoretical, in order to guarantee the primary stability of system;

(2) build adaptive neural network

The multi-source existed is in-flight disturbed, design the control accuracy that the adaptive neural network with autonomous renewal weight matrix improves system, realize that On-line Estimation and inhibition are carried out in suffered multi-source interference awing to the miniature self-service giro.

Adaptive neural network consists of input layer, hidden layer and output layer; The input layer of adaptive neural network be input as expectation state variable x _dAnd the error between virtual condition variable x, i.e. e=x-x _d

Hidden layer consists of a plurality of Gaussian functions, is defined as s (e), and the node function expression of its corresponding j hidden layer is as follows:

$s_j\left(e\right)=\exp \left(\frac{{-(e-{\mathrm{\&mu;}}_j)}^T(e-{\mathrm{\&mu;}}_j)}{{\mathrm{\&sigma;}}_j^2}\right),j=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;l$

Wherein, μ _j,

Be respectively central value and the width of adaptive neural network hidden layer Gaussian function, l is the implicit nodes of adaptive neural network hidden layer, the central value μ of self-adaptation radial base neural net node _j, width

By the user, determined.

The estimated value that adaptive output layer disturbs multi-source

$\widehat{\mathrm{\&omega;}}=B^T{\left({\mathrm{BB}}^T\right)}^{-l}{\hat{W}}^{T}s\left(e\right)-{\mathrm{\&alpha;B}}^T{\left({\mathrm{BB}}^T\right)}^{-l}\mathrm{sign}\left(e^T{P}_{*i}\right)$

Wherein,

Weight matrix for adaptive neural network; Γ _i, P is symmetric positive definite matrix, α is the control residual error upper limit threshold of adaptive neural network, i* is i row vector of corresponding matrix, * i is i column vector of corresponding matrix, wherein the weight matrix of adaptive neural network independently upgrades according to following rule

${\stackrel{\&CenterDot;}{\hat{W}}}_{*i}={-\mathrm{\&Gamma;}}_{i}s\left(e\right)e^T{P}_{*i}$

Wherein P is following formula equation symmetric positive definite solution

(A+BK) ^TP+P(A+BK)＝-Q

Symmetric positive definite matrix Q=I wherein.

(3) design adaptive threshold optimisation strategy

Physical location in the time-based window is poor with the error mean square of desired locations, and the control residual error upper limit threshold of adaptive neural network is carried out to online updating, realizes the miniature self-service giro high-precision attitude control under complex environment.

The adaptive threshold optimisation strategy is defined as follows:

${\mathrm{\&alpha;}}_w=\left\{\begin{array}{cc}\frac{{\mathrm{\&alpha;}}_{w-1}}{{k_1}^2}& \underset{i=k-t}{\overset{k}{\max }}\left|{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_1\mathrm{and}\left|\frac{1}{t}\underset{i=k-t}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_2\mathrm{and}{\mathrm{\&chi;}}_k>{\mathrm{\&chi;}}_{k-t}\mathrm{andflag}=1\\ k_1{\mathrm{\&alpha;}}_{w-1}& \underset{i=k-t}{\overset{k}{\max }}\left|{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_1\mathrm{and}\left|\frac{1}{t}\underset{i=k-t}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_2\mathrm{and}{\mathrm{\&chi;}}_k<{\mathrm{\&chi;}}_{k-t}\mathrm{andflag}=1\\ k_1{\mathrm{\&alpha;}}_{w-1}& \underset{i=k-t}{\overset{k}{\max }}\left|{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_1\mathrm{and}\left|\frac{1}{t}\underset{i=k-t}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{ee}}_i\right|>{\mathrm{\&eta;}}_2\\ {\mathrm{\&alpha;}}_{w-1}& \mathrm{otherwise}\end{array}\right.$

$\mathrm{flag}=\left\{\begin{array}{cc}1& {\mathrm{\&alpha;}}_{w-1}\&NotEqual;{\mathrm{\&alpha;}}_{w-2}\\ 0& {\mathrm{\&alpha;}}_{w-1}={\mathrm{\&alpha;}}_{w-2}\end{array}\right.$

${\mathrm{\&chi;}}_k=\sqrt{\underset{m=k-t}{\overset{k}{\mathrm{\&Sigma;}}}{({\mathrm{xx}}_m-{\mathrm{xx}}_d)}^2/t}$

α in formula _wControl residual error upper limit threshold for current time window sample adaptive neural network in the cycle; α _W-1Control residual error upper limit threshold for upper time window sample adaptive neural network in the cycle; α _W-2Control residual error upper limit threshold for upper two time windows adaptive neural network in the sampling period; χ _kFor upper time window sample miniature self-service giro physical location xx in the cycle _mWith desired locations xx _dMean square deviation; χ _k-tFor upper two time windows miniature self-service giro physical location xx in the sampling period _mWith desired locations xx _dMean square deviation; Ee _iFor time window miniature self-service giro physical location xx in the sampling period _mWith desired locations xx _dDifference; T is sampling number in the time window sampling period; k ₁For controlling parameter, η ₁, η ₂Be respectively miniature self-service giro physical location xx in the time window sampling period _mWith desired locations xx _dMaximum absolute error value and average error amount.

(5) flight example

Based on the miniature self-service giro, carry out four destination flight experiment checkings.Four destination patrol flights are fixed high 20 meters, are starting point with destination (10 ,-20,20), pass through respectively destination (40 ,-20,20), and (40,0,20) and (10,0,20), finally hover over starting point.Based on the comparing result of the feedback of identical feedback control matrix parameter and adaptive neural network control method as shown in Figure 2, strong wind at 3.2m/s is disturbed under environment, line ball precision based on the miniature self-service giro of adaptive neural network the four destination line walking task of execution the time is 1.56m, line ball precision at hovering phase is 0.83m, all is better than traditional feedback.

The miniature self-service giro high-accuracy control method that the present invention is based on adaptive neural network has overcome the deficiency of existing control method, can realize that miniature self-service giro complexity disturbs high precision flight under environment more and control etc.

The content be not described in detail in instructions of the present invention belongs to the known prior art of professional and technical personnel in the field.

Claims (3)

1. the giro of the miniature self-service based on adaptive neural network high-accuracy control method is characterized in that realizing following steps:

(1) for miniature self-service giro kinetic model, build by pole-assignment the primary stability that the feedback control coefficient matrix guarantees system;

(2) multi-source existed is in-flight disturbed, the adaptive neural network that design has autonomous renewal weights characteristic, based on control information, build the weight matrix that adaptive network right value update matrix carrys out the online updating neural network, realize that On-line Estimation and inhibition are carried out in suffered multi-source interference awing to the miniature self-service giro;

(3) design adaptive threshold optimisation strategy, physical location in the time-based window and the error mean square of desired locations are poor, control residual error upper limit threshold to adaptive neural network carries out online updating, realizes that the miniature self-service giro high-precision attitude under complex environment is controlled.

2. the giro of the miniature self-service based on adaptive neural network high-accuracy control method according to claim 1 is characterized in that: described step (2) build adaptive neural network right value update matrix and disturbance controlled quentity controlled variable expression formula as follows:

Wherein, For the weight matrix of adaptive neural network,

Disturbance controlled quentity controlled variable for adaptive neural network; Γ _i, P is symmetric positive definite matrix, the input e=x-x of adaptive neural network _dFor expectation state variable x _dAnd the error between virtual condition variable x, B is miniature self-service giro state of a control transition matrix, α is the control residual error upper limit threshold of adaptive neural network, i* is i row vector of corresponding matrix, * i is i column vector of corresponding matrix, s (e) is the node function of adaptive neural network hidden layer, is defined as Gaussian function, and the node function expression of its corresponding j hidden layer is as follows:

Wherein, μ _j,

Be respectively central value and the width of adaptive neural network hidden layer Gaussian function, l is the implicit nodes of adaptive neural network hidden layer.

3. the giro of the miniature self-service based on adaptive neural network high-accuracy control method according to claim 1 is characterized in that: described step (3) builds the adaptive threshold optimisation strategy and is defined as follows:

α in formula _wControl residual error upper limit threshold for current time window sample adaptive neural network in the cycle; α _W-1Control residual error upper limit threshold for upper time window sample adaptive neural network in the cycle; α _W-2Control residual error upper limit threshold for upper two time windows adaptive neural network in the sampling period; χ _kFor upper time window sample miniature self-service giro physical location xx in the cycle _mWith desired locations xx _dMean square deviation; χ _k-tFor upper two time windows miniature self-service giro physical location xx in the sampling period _mWith desired locations xx _dMean square deviation; Ee _iFor time window miniature self-service giro physical location xx in the sampling period _mWith desired locations xx _dDifference; T is sampling number in the time window sampling period; k ₁For controlling parameter, η ₁, η ₂Be respectively miniature self-service giro physical location xx in the time window sampling period _mWith desired locations xx _dMaximum absolute error value and average error amount.

Images