CN102053628A - Neural network-based servo control system and method - Google Patents

Abstract

The invention aims to improve the control accuracy of a servo system and provides a neural network model reference adaptive control method applied to the servo system. The nonlinearity of the servo system is effectively compensated, the interference is suppressed, and the tracking accuracy and robustness of the servo system are improved. In addition, the control of the servo system is not needed to be constructed on the basis of accurately modeling an object, the modeling cost is saved, the method is easily implemented in engineering, and the cost is reduced.

Description

Servo-control system and method based on neural network

Technical field

The present invention relates to a kind of servo-control system and method based on neural network.

Technical background

Servo-drive system is complicated Mechatronic control system, and its essence can be considered as one by motor-driven position closed loop control system, and it plays an important role in national product and national defense construction.Because it occupies crucial status in each field, thus its performance demands is also improved constantly, especially in leading-edge fields such as national defense and military and Aero-Space.From the total development trend of current domestic and international servo-drive system as can be seen, " high frequency sound, Ultra-Low Speed, high precision " is its main developing direction.Wherein, " high frequency sound " is the ability that the reflection servo-drive system is followed the tracks of high-frequency signal, the i.e. tracking power of system when position command signal constantly changes." Ultra-Low Speed " is the low-speed stability of reflection system, and the principal element that influences low-speed characteristic is mechanical friction, must adopt certain control method that friction is compensated." high precision " is meant the order of accuarcy of system keeps track command signal.

Be present in non-perpendicularity or factor such as many nonlinear, uncertainties such as the imbalance, mechanical hook-up insufficient rigidity of degree of friendship and the system load moment that causes and the mechanically deform that causes, the fluctuation of load and the slot effect of motor itself etc. not between drift, the moment coupling between axle system, environmental interference and the axle system of mechanical friction in the servo-drive system, circuit parameter, caused a lot of difficulties for the control of servo-drive system, very big to the precision influence of system.Therefore, eliminate the disturbance that these interference sources cause and to overcome the influence that various non-linear factors bring system be the key that realizes the servo-drive system High Accuracy Control.

" three rings " structure PID control method (referring to Fig. 3) that the general employing of classical Servo System Design is traditional is electric current loop, speed ring and position ring from inside to outside.The effect of electric current loop and speed ring is the non-linear and external disturbance that the rigidity of raising system suppresses system, and the precision of control system is guaranteed by position ring.But the bad adaptability of this traditional control method, control accuracy is low under the disturbed situation of system, is not suitable for the occasion of High Accuracy Control.And the present invention can be good at suppressing the disturbance that parameter perturbation, friction interference and the load variations of system are brought; Under the non-linear and uncertain stronger situation of object, also can normally move, improve the control accuracy of servo-drive system greatly.

Summary of the invention

In order to improve the control accuracy of servo-drive system, particularly improve the control accuracy of servo-drive system under the disturbed conditions such as parameter perturbation, friction interference and load variations that have non-linear and uncertainty and system, the present invention proposes a kind of control method and system based on neural network.

The present invention has realized the nonlinear compensation of servo-drive system and has disturbed and suppress, improved the tracking accuracy of servo-drive system.The present invention is the method that has added Neural Network Adaptive Control on the basis of traditional control method, has improved the robustness of system greatly, and system can both be suppressed fast and effectively to various interference, reaches high control accuracy.And guaranteed the stability of this method based on the adaptive control algorithm of Lyapunov stability theory.In addition, on the basis of speed ring accurate reference model, also become very simple of the design of position ring controller makes the design effort of total system become quite convenient, is easy to implement in engineering reality.

According to an aspect of the present invention, provide a kind of positional servosystem, having comprised:

A position ring controller is used to receive a site error, and produces a speed command;

An adaptive controller is used to receive described speed command, a model error and a target servo rate signal, and produces an adaptive controller output;

A reference model is used to receive described speed command and produces a reference model output;

A nerve network controller is used to receive the differential of described model error and described model error, and exports a nerve network controller output;

A robust item parts is used to receive described model error signal, and produces a robust item output;

A differentiator is used to receive described model error, and produces the differential of described model error;

A servo actuating unit is used for carrying out servo operation under the control of described control signal.

According to a specific embodiment of the present invention, above-mentioned positional servosystem further comprises:

A position detecting device is used to detect the position of the target servo of described positional servosystem, and generates a controlling object position output;

A speed detector is used to detect the speed of the target servo of described positional servosystem, thereby generates described target servo rate signal;

A first adder is used for an output of described controlling object position and a received position command of described positional servosystem are subtracted each other, thereby generates described site error;

A second adder is used for described adaptive controller output, the output of described nerve network controller and described robust item output addition, thereby produces described control signal;

One the 3rd totalizer is used for described target servo speed and the output of described reference model are subtracted each other, thereby generates described model error.

According to a further aspect of the present invention, above-mentioned nerve network controller further comprises: the normalization part, be used for differential to described model error and described model error, and obtain corresponding normalization x1 as a result, x2...xn; With the output x1 of normalization part, x2...xn handles through the gaussian basis function of choosing by input matrix V computing again and obtains corresponding hidden node q _iThe weighted sum part is used for hidden node q _i(being the gaussian basis function) multiply by corresponding weights and summation, is the output of nerve network controller thereby obtain the weighted sum result as the output of neural network.Wherein, the update algorithm of described weights is

Wherein, q _iBe the correspondence input (i.e. the hidden node of this neural network) of these weights, γ ₁Be a default neural network learning speed, its value is the real number greater than zero, the system that this numerical basis is different make corresponding adjustment (need to prove this value generally 01 between value, but, for all control system, this value is not to be the bigger the better, and is not necessarily the smaller the better yet, should rule of thumb make rational selection according to system).e _ωPromptly be described model error, its expression formula and neural network weight update algorithm see for details hereinafter derives.

According to another aspect of the present invention, provide a kind of position servo method, comprised

Receive a site error and produce a speed command with a position ring controller;

Receive described speed command, a model error and a target servo rate signal with an adaptive controller, and produce an adaptive controller output;

Receive described speed command and produce a reference model output with a reference model;

Receive the differential of described model error and described model error with a nerve network controller, and export a nerve network controller output;

Receive described model error signal and produce a robust item output with a robust item parts;

Receive described model error and produce the differential of described model error with a differentiator.

According to a specific embodiment of the present invention, above-mentioned position servo method further comprises:

By a servo actuating unit, under the control of described control signal, carry out servo operation;

By a position detecting device, detect the position of the target servo of described positional servosystem, and generate a controlling object position output;

By a speed detector, detect the speed of the target servo of described positional servosystem, thereby generate described target servo rate signal;

A first adder is used for an output of described controlling object position and a received position command of described positional servosystem are subtracted each other, thereby generates described site error;

A second adder is used for the output addition of described adaptive controller output, the output of described nerve network controller and described robust item, thereby produces described control signal;

One the 3rd totalizer is used for described target servo speed and the output of described reference model are subtracted each other, thereby generates described model error.

Description of drawings

Fig. 1 is the detailed structure synoptic diagram of nerve network controller according to an embodiment of the invention;

Fig. 2 is the synoptic diagram of positional servosystem according to an embodiment of the invention;

Fig. 3 is the synoptic diagram of traditional servo-drive system of classical " three rings " structure.

Embodiment

The theoretical construct of servo-control system according to an embodiment of the invention as shown in Figure 2.The control section of this system comprises position ring controller 101, adaptive controller 102, nerve network controller 104 and robust item parts 105.

The controlling object of label 107 these servo-control systems of expression among Fig. 2.The essential part of controlling object is servo operating part 1071.In the actual conditions, controlling object 107 also can comprise current feedback part 1072 and power amplifying part 1073 usually.

The servo-control system of embodiment shown in Figure 2 also comprises position detecting device 108, speed detector 109 and differentiator 106.

1) position ring controller 101

The difference that is input as position command and position output of position ring controller 101, it is output as speed ring instruction r.Polytype controller can be used as position ring controller, for example PID controller (ratio, integration, derivative controller) and P controller (proportional controller), but be not limited only to this two kinds of controllers.Recommend to select simple P controller for use in this programme, its input/output relation can be written as r=K _p(θ _d-θ), K _pBe artificial selected constant (span?).This controller can be realized by computer software, also can realize with hardware circuit.

2) the controlled device module 107

Label is that 107 module is the controlled device of system among Fig. 2, i.e. the whole controlling system object.

The controlled device of servo-drive system is generally motor, and the power amplifier apparatus that is used for drive motor, can also include the electric current loop of certain form simultaneously.Controlled device 107 shown in Fig. 2 is a kind of possible structure of this module, but is not limited only to this structure.This module is input as controller output u, is output as motor speed ω.Its input/output relation can be by following differential equation:

$\left\{\begin{array}{c}J\frac{\mathrm{d\ω}}{\mathrm{dt}}+\mathrm{B\ω}+T_d={\mathrm{Ki}}_a\\ \mathrm{K\ω}+R_a{i}_a+L_a\frac{{\mathrm{di}}_a}{\mathrm{dt}}=u_a\\ u_a=K_{m}u\\ \mathrm{\θ}=\∫\mathrm{\ωdt}\end{array}\right.$

Wherein J, B represent the moment of inertia and the viscous friction coefficient of motor.T _dTorque is disturbed in expression equivalence, for example moment of friction, since mechanical deformation and the moment of elasticity that produces on the transmission shaft etc.i _a, u _a, L _a, θ represents armature supply respectively, armature voltage, armature inductance and motor corner.K represents moment coefficient.K _mAmplification coefficient for power amplifier device.

On meaning more generally, click rotational speed omega and the rotational angle theta of embodiment belongs to " object parameters " among Fig. 2.

3) the adaptive controller module 102

Being numbered 102 part among Fig. 2 is the adaptive controller module, and it both can be realized that also available hardware realized by computer software.

Adaptive controller 102 modules be input as speed command r, speed output ω and model error e _ω, be output as adaptive controller output u _MracIn order to guarantee system stability,, following relation is arranged between the input and output of this module according to theoretical derivation (seeing for details hereinafter):

$u_{\mathrm{mrac}}=\mathrm{\γ}\left[{\∫}_0^t{e}_{\mathrm{\ω}}\left(\mathrm{\τ}\right)r\left(\mathrm{\τ}\right)\mathrm{d\τ}\right]r\left(t\right)+\mathrm{\γ}\left[{\∫}_0^t{e}_{\mathrm{\ω}}\left(\mathrm{\τ}\right)\mathrm{\ω}\left(\mathrm{\τ}\right)\mathrm{d\τ}\right]\mathrm{\ω}\left(t\right)$

Wherein, γ for the people be the neural network learning speed chosen (this value generally 01 between value, still, in general, this value is not to be the bigger the better, and is not necessarily the smaller the better yet, should rule of thumb make rational selection according to system).

4) the reference model module 103

Being numbered 103 part among Fig. 2 is the reference model module of system, both can realize by computer software, also can utilize hardware circuit to realize.The output that is input as position ring controller of this module, it is output as reference model output.The input/output relation of this module is

$\frac{d{\mathrm{\ω}}_m}{\mathrm{dt}}+a_m{\mathrm{\ω}}_m=b_{m}r$

This relation is to draft according to desirable dc motor model.The object that its purpose is a reference is provided is to controller, and adaptive controller and nerve network controller will be as standards, and on purpose Adjustment System makes velocity output signal consistent with reference model output.ω in the formula _mBe the output of reference model.a _mAnd b _mBe model parameter, can utilize the method for System Discrimination commonly used to record.

Characteristics of the present invention do not need to be the accurate model of controlled device, and the object model that simple discrimination method commonly used records just can use in system, and this is the important difference of this method and other model reference methods.

5) the nerve network controller module 104

Label is that 104 part is a nerve network controller among Fig. 2.

According to the concrete structure of a kind of embodiment of nerve network controller of the present invention referring to Fig. 1.In control system according to the present invention, nerve network controller 104 both can be realized with computer software, also can realize with hardware circuit.This module be input as described model error e _ωAnd the differential of model error Be output as nerve network controller output u _n

As shown in Figure 1, in nerve network controller 104 of the present invention, at first to e _ω, The value of current sampling instant carry out normalization.Method for normalizing can use various method for normalizing commonly used, and the method for normalizing that the present invention recommends to use is to import the absolute value of historical peak value with the currency of certain input divided by this.Then with the x1 after the normalization, be brought into after the computing of x2...xn through input matrix V in the gaussian basis function and obtain the value of hidden node Q.Hidden node is by multiply by the output u that corresponding weights are sued for peace and can be obtained nerve network controller _nWherein, the update algorithm of neural network weight w is

${\stackrel{\·}{w}}_i=-{\mathrm{\γ}}_1{e}_{\mathrm{\ω}}{q}_i$

Q wherein _iBe the correspondence input (also being the hidden node of this neural network) of these weights, γ ₁Be neural network learning speed, its value is the real number greater than zero; In the actual servo control system, can select suitable γ according to experience ₁Value, and make the performance of system reach optimum; Parameter e _ωBe described model error.

Described above is multiple a kind of in several neural network structure of the present invention of being suitable for, the cerebellum model artificial neural network that this area is commonly used, and several networks such as wavelet neural network also can be used as according to the neural network in the technical scheme of the present invention.

6) robust item module 105

Label is that 105 part is the robust item among Fig. 2.

Robust item module 105 can realize by computer programming in servo-control system of the present invention, also can have hardware circuit to realize.This module be input as described model error e _ω, be output as robust item output u _r, the output expression formula of robust item module is

Wherein

γ ₂It is a default robust item coefficient, value is the real number greater than zero, in the control system of reality, can choose suitable numerical value according to experience, make the performance of control system reach optimum (need to prove this value generally 01 between value, still, in general, this value is not to be the bigger the better, also not necessarily the smaller the better, should rule of thumb make rational selection according to system)

Sign () is-symbol function.

The relation of its input and output will be made detailed derivation hereinafter.

The theoretical foundation of Neural Network Adaptive Control device design

Control system is except accuracy requirement, and the very important requirement of another one is the stability requirement of system.Realize the automatic control of a system, just necessary assurance system stablizes.Otherwise unsettled system contingent out-of-control condition in actual production will be unacceptable.In case generation systems is out of control in the actual production, can cause any property loss usually, sometimes or even casualties.Thereby, the analysis of system stability or prove the indispensable part of perfect control system design.

Technical scheme of the present invention has solid theory and strict stability to prove that this can guarantee the actual production that is applied to that the present invention can be very safe.

The stability proof of technical scheme of the present invention is as follows:

The dynamic perfromance of servo-drive system can be described by following equation

$J\frac{\mathrm{d\ω}}{\mathrm{dt}}+\mathrm{B\ω}+T_d{\mathrm{Ki}}_a---\left(1\right)$

$u_a=\mathrm{K\ω}+R_a{i}_a+L_a\frac{d{i}_a}{\mathrm{dt}}---\left(2\right)$

θ＝∫ωdt (3)

Wherein J, B represent the moment of inertia and the viscous friction coefficient of motor.ω represents the angle speed of motor.T _dTorque is disturbed in expression equivalence, for example moment of friction, since mechanical deformation and the moment of elasticity that produces on the transmission shaft etc.i _a, u _a, L _aRepresent armature supply respectively, armature voltage and armature inductance.K represents moment coefficient.θ is the position, angle of motor corner.Because armature inductance is very little, i.e. L _a≈ 0, usually it ignored in practice.Can obtain following dynamic equation from equation (1) to (3),

$J\frac{\mathrm{d\ω}}{\mathrm{dt}}+\mathrm{B\ω}+T_d=\frac{K}{R_a}{u}_a-\frac{K^2}{R_a}\mathrm{\ω}---\left(4\right)$

$\frac{{\mathrm{JR}}_a}{K}\frac{\mathrm{d\ω}}{\mathrm{dt}}\mathrm{\ω}+\frac{R_a}{K}(B+\frac{K^2}{R_a})\mathrm{\ω}+\frac{R_a}{K}{T}_d=u_a---\left(5\right)$

Order

Equation (1) can be write as again so

$A\frac{\mathrm{d\ω}}{\mathrm{dt}}\mathrm{\ω}+\mathrm{C\ω}+u_d=u_a---\left(6\right)$

Further, order

Can obtain

$\frac{\mathrm{d\ω}}{\mathrm{dt}}=-\mathrm{a\ω}+\mathrm{bu}-u_{\mathrm{fr}}---\left(7\right)$

Can select the speed ring reference model as follows according to following formula

$\frac{d{\mathrm{\ω}}_m}{\mathrm{dt}}=-a_m{\mathrm{\ω}}_m+b_{m}r---\left(8\right)$

Wherein, a _m, b _mIt is the motor model estimated parameter that records according to least-squares algorithm.

Definition model error e _m=ω-ω _m.

The definition controller architecture is as follows,

u＝u _mrac+u _n-u _r (9)

U wherein _Mrac=θ ₁R-θ ₂ω is the output of adaptive controller, u _nBe the output of nerve network controller, u _sBe the output of robust item.

Model error to the derivative of time is so,

$\frac{{\mathrm{de}}_{\mathrm{\ω}}}{\mathrm{dt}}=\frac{\mathrm{d\ω}}{\mathrm{dt}}-\frac{{\mathrm{d\ω}}_m}{\mathrm{dt}}$

$=-a{\mathrm{\ω}}_{\mathrm{\ω}}-(a+b{\mathrm{\θ}}_2-a_m)\mathrm{\ω}+(b{\mathrm{\θ}}_1-b_m)r+b{u}_n-b{u}_r-u_{\mathrm{fr}}$ (10)

$=-a_m{\mathrm{\ω}}_{\mathrm{\ω}}-(a+b{\mathrm{\θ}}_2-a_m)\mathrm{\ω}+(b{\mathrm{\θ}}_1-b_m)r+b{u}_n^{*}-u_{\mathrm{fr}}+{\mathrm{bu}}_n-{\mathrm{bu}}_n^{*}-{\mathrm{bu}}_r$

$=-a_m{\mathrm{\ω}}_{\mathrm{\ω}}-(a+b{\mathrm{\θ}}_2-a_m)\mathrm{\ω}+(b{\mathrm{\θ}}_1-b_m)r+b(u_n^{*}-\frac{u_{\mathrm{fr}}}{b})+b(u_n-u_n^{*})-{\mathrm{bu}}_r$

Order Be distracter

Then $\frac{{\mathrm{de}}_{\mathrm{\ω}}}{\mathrm{dt}}=-a_m{\mathrm{\ω}}_{\mathrm{\ω}}-(a+b{\mathrm{\θ}}_2-a_m)\mathrm{\ω}+(b{\mathrm{\θ}}_1-b_m)r+b(u_n^{*}-f\left(x\right))+b(u_n-u_n^{*})-b{u}_r---\left(11\right)$

Definition, u _n=WQ, W are the weight vectors of neural network, W=[w ₁, w ₂... w _k] ∈ R ^{1 * k}, X is the input of neural network, X=[x ₁x ₂... x _n] ∈ R ^{N * 1}, u _nBe the output of nerve network controller.Wherein X is through normalized.Neural network structure figure such as Fig. 1. wherein, hidden node is

Gaussian basis function parameters m=[m ₁, m ₂... m _k] ^T∈ R ^{K * 1},

S=[s ₁, s ₂... s _k] ^T∈ R ^{K * 1}, input matrix is

Being input as of gaussian basis function

I _n＝VX＝[I _n1，I _n2，...I _nk] ^T∈R ^k×1

For real system, desirable

But be not limited to this two parameter.

Wherein, as gaussian basis function parameters m, when the s covering scope was enough big, its parameter needn't iteration, and only needs the output parameter of iteration neural network to get final product.

As fixing V, m, during s, an iteration output weights W, its optimal value is W ^*

The evaluated error of power is defined as

Then,

Order

| E|＜ε.ε is the boundary of approximate error.Estimated value to these error bound is

Estimated bias is

Then have:

$\frac{{\mathrm{de}}_{\mathrm{\ω}}}{\mathrm{dt}}=-a_m{\mathrm{\ω}}_{\mathrm{\ω}}-(a+b{\mathrm{\θ}}_2-a_m)\mathrm{\ω}+(b{\mathrm{\θ}}_1-b_m)r+\mathrm{bE}+b\tilde{W}Q-{\mathrm{bu}}_r---\left(12\right)$

Definition Lyapunov function is as follows,

$V({\mathrm{\θ}}_1,{\mathrm{\θ}}_2,\tilde{W},\widetilde{\mathrm{\ϵ}})=\frac{1}{2}[{e_{\mathrm{\ω}}}^2+\frac{1}{\mathrm{b\γ}}{(a+b{\mathrm{\θ}}_2-a_m)}^2+\frac{1}{\mathrm{b\γ}}{(b{\mathrm{\θ}}_1-b_m)}^2+\frac{b}{\mathrm{\γ}}\tilde{W}{\tilde{W}}^T+\frac{b}{\mathrm{\γ}}{\widetilde{\mathrm{\ϵ}}}^2]---\left(13\right)$

V to the derivative of time t is

Substitution

:

Make the turnover rate of parameter be:

$\left\{\begin{array}{c}\frac{d{\mathrm{\θ}}_1}{\mathrm{dt}}=-\mathrm{\γ}{e}_{\mathrm{\ω}}r\\ \frac{d{\mathrm{\θ}}_2}{\mathrm{dt}}=\mathrm{\γ\ω}{e}_{\mathrm{\ω}}\\ \stackrel{\·}{W}=-\mathrm{\γ}{e}_{\mathrm{\ω}}Q\end{array}\right.---\left(15\right)$

The manually selected learning rate of γ wherein; The value of γ generally 01 between value, but in general, this value is not to be the bigger the better, and is not necessarily the smaller the better yet, should rule of thumb make rational selection according to system,

Then have, when

(a, b, a _m, b _mBe constant greater than 0),

Only need to guarantee

Then

This moment, error equation was by exponential convergence.

Because

(16)

Get

Then

Get again

Then

Promptly

Know that according to the Lyapunov stability principle system can guarantee to stablize.

By above analytical derivation, the actual dynamic perfromance of speed ring is with the dynamic perfromance of track reference model.Position ring controller can design according to the speed ring reference model.Like this, be easy to guarantee the stability of whole position closed loop system.

The present invention utilize adaptive algorithm can be online, the characteristics of real-time regulated parameter, reach purpose by change to system stability control to parameter values.Neural network algorithm has characteristics such as speed of convergence, the nonlinear function that can approach complexity and self-learning capability faster simultaneously, distributed parallel is handled, Nonlinear Mapping, characteristics such as robust Fault-Tolerant and generalization ability are strong make it realize squelch effect and nonlinear compensation to servo-drive system in learning process.

Advantage of the present invention comprises:

The control method that-employing self adaptation and neutral net combine has overcome the impact of nonlinearity erron to the servo-drive system control accuracy effectively;

-the present invention adopts adaptive algorithm, can regulate online parameter control system, has characteristics simple to operate, with low cost;

-the present invention does not need to be based upon on the basis of object Accurate Model to the control of servo-drive system, has saved the expense of modeling.

Claims (10)

1. a positional servosystem comprises

A position ring controller (101) is used to receive a site error, and produces a speed command (r);

An adaptive controller (102) is used to receive described speed command (r), a model error (e _ω) and a target servo rate signal (ω), and produce an adaptive controller output (u _Mrac);

A reference model (103) is used to receive described speed command (r) and produces a reference model output (ω _m);

A nerve network controller (104) is used to receive described model error (e _ω) and the differential of described model error

And export a nerve network controller and export (u _n);

A robust item parts (105) is used to receive described model error signal (e _ω), and produce a robust item output (u _r);

A differentiator (106) is used to receive described model error (e _ω), and produce the differential of described model error

A servo actuating unit (107) is used for carrying out servo operation under the control of described control signal (u).

2. positional servosystem according to claim 1 is characterized in that further comprising:

A position detecting device (108) is used to detect the position of the target servo of described positional servosystem, and generates a controlling object position output;

A speed detector (109) is used to detect the speed of the target servo of described positional servosystem, thereby generates described target servo rate signal (ω);

A first adder (110) is used for an output of described controlling object position and a received position command of described positional servosystem are subtracted each other, thereby generates described site error;

A second adder (111) is used for described adaptive controller output (u _Mrac), described nerve network controller output (u _n) and described robust item output (u _r) addition, thereby produce described control signal (u);

One the 3rd totalizer (112) is used for described target servo speed (ω) and described reference model output (ω _m) subtract each other, thereby generate described model error (e _ω).

3. positional servosystem according to claim 2 is characterized in that described nerve network controller further comprises:

The normalization part is used for described model error (e _ω) and the differential of model error

Carry out normalization, obtain corresponding normalization result (x ₁, x ₂... x _n);

The weighted sum part is used for hidden node (q _i) multiply by corresponding weights (w _i) and summation, thereby obtain weighted sum result as the output of neural network,

Gaussian basis function processing section is used for the output (x with the normalization part ₁, x ₂... x _n) handle through a gaussian basis function of choosing again through the computing of input matrix (V), thereby obtain corresponding described hidden node (q _i).

4. positional servosystem according to claim 3 is characterized in that described weights (w _i) update algorithm be

${\stackrel{\·}{w}}_i=-{\mathrm{\γ}}_1{e}_{\mathrm{\ω}}{q}_i$

Wherein, q _iBe the correspondence input of these weights, γ ₁Be a predefined neural network learning speed, e _ωBe described model error,

γ wherein ₁Value be real number greater than zero, its numerical value can be adjusted according to the different qualities of control system.

5. positional servosystem according to claim 1 is characterized in that the output expression formula of described robust item parts (105) is

Wherein

e _ωBe described model error,

γ ₂Be a default robust item coefficient, its value is the real number greater than zero, and described value can adjust according to the difference of system,

Sign () is-symbol function.

6. a position servo method comprises

Receive a site error and produce a speed command (r) with a position ring controller (101);

Receive described speed command (r), a model error (e with an adaptive controller (102) _ω) and a target servo rate signal (ω), and produce an adaptive controller output (u _Mrac);

Receive described speed command (r) and produce a reference model output (ω with a reference model (103) _m);

Receive described model error (e with a nerve network controller (104) _ω) and the differential of described model error

And export a nerve network controller and export (u _n);

Receive described model error signal (e with a robust item parts (105) _ω) and produce a robust item output (u _r);

Receive described model error (e with a differentiator (106) _ω) and produce the differential of described model error

7. position servo method according to claim 6 is characterized in that further comprising:

By a servo actuating unit (107), under the control of described control signal (u), carry out servo operation;

By a position detecting device (108), detect the position of the target servo of described positional servosystem, and generate a controlling object position output;

By a speed detector (109), detect the speed of the target servo of described positional servosystem, thereby generate described target servo rate signal (ω);

A first adder (110) is used for an output of described controlling object position and a received position command of described positional servosystem are subtracted each other, thereby generates described site error;

A second adder (111) is used for described adaptive controller output (u _Mrac), described nerve network controller output (u _n) and the output (u of described robust item _r) addition, thereby produce described control signal (u);

One the 3rd totalizer (112) is used for described target servo speed (ω) and described reference model output (ω _m) subtract each other, thereby generate described model error (e _ω).

8. position servo method according to claim 7 is characterized in that receiving described model error (e with described nerve network controller _ω) and the differential of model error

And the described step of exporting described nerve network controller output further comprises:

Differential to described model error and model error carries out normalization, obtains corresponding normalization result (x ₁, x ₂... x _n);

Output (x with the normalization part ₁, x ₂... x _n) through the processing of input matrix (V) computing and selected gaussian basis function, thus corresponding hidden node (q obtained _i);

With hidden node (q _i) multiply by corresponding weights and summation, thus weighted sum result obtained as the output of neural network.

9. position servo method according to claim 8 is characterized in that the update algorithm of described weights is

${\stackrel{\·}{w}}_i=-{\mathrm{\γ}}_1{e}_{\mathrm{\ω}}{q}_i$

Wherein, q _iBe the correspondence input of these weights,

γ ₁Be a predefined neural network learning speed, its value also can adjust according to the different qualities of control system for the real number greater than zero;

e _ωIt is described model error.

10. position servo method according to claim 6 is characterized in that with the expression formula of described robust item output being

Wherein

e _ωBe described model error,

γ ₂Be a default robust item coefficient, its value can be adjusted according to the difference of system for real number and this value greater than zero,

Sign () is-symbol function.

Images