CN104298213A - Index time varying gain type iterative learning control algorithm based on reference batch - Google Patents

Abstract

The invention relates to an iterative learning control algorithm in an intermittent manufacturing and producing process and belongs to the field of industrial production automatic control technology. According to the method, in the machining, manufacturing and producing process of a majority of batches, an iterative learning law can be designed by introducing a proper 'reference batch' into an iterative learning control process and utilizing input and output data of the reference batch and the current batch under the condition that specific knowledge on objects is little, and therefore effective tracking on the track of an output target is achieved. Meanwhile, index time varying gain is introduced, so that the performance of a track tracking transitional process is improved. The method is small in calculation amount, simple in optimizing process, wide in application range, ingenious in concept, simple and practical and can be widely applied to high-accuracy track curved line tracking control in the intermittent producing process of an industrial production line.

Description

A kind of based on reference to batch index time-varying gain type Iterative Learning Control Algorithm

Technical field

The present invention relates to the Iterative Learning Control Algorithm that a kind of batch (-type) manufactures production run, particularly object concrete knowledge know little in situation a kind of with reference to batch index time-varying gain type Iterative Learning Control Algorithm, belong to industrial production automation control technology field.

Background technology

Batch production process be a kind of production run that manufactures by batch carrying out repetitive operation, and batch between there is the mode of production of certain quiescent interval.At industrial circles such as biological products, pharmaceutical production, fine chemistry industry, SIC (semiconductor integrated circuit), batch process is produced to have and is applied very widely.Batch process has the features such as intermittence, repeatability, time-variant nonlinear, high-purity, multi items and short run.In order to the high-quality of the stability and final products that ensure product quality, quality control seems particularly important.But because batch process interior change mechanism is usually very complicated, different manufacturing installation form is different again, and the accurate model thus setting up batch process is very difficult.Because current product quality Advanced Control Techniques is mainly based on the method that linear model controls, is directly applied in batch process and also has certain difficulty.In batch process is produced, when manufacturing formula and being constant, production run reruns substantially, and within each batch of cycle of operation, control variable and product quality run along certain operation variation track, thus has stronger repeatability.

Iterative learning controls (Iterative Learning Control, ILC) controlled device that a class has periodicity, the characteristic that reruns is suitable for, its task finds control inputs, make the actual output trajectory of controlled device perfect tracking along whole desired output path implementation zero error in finite time interval, and whole control procedure requires to complete fast.For existing most of Iterative Learning Control Algorithm, when designing its iterative learning control law, more or less all need certain object concrete knowledge.That is, the system model structure of control object should be set up, and all or part of parameter of system model also should be known.In control practice, object concrete knowledge obtain not a duck soup, sometimes or even impossible.Such as, for a batch process, when being in: 2. 1. the operation of process is batch limited is under initially some batches of stages, 3. process had the stronger too high situation of non-linear 4. modeling cost, object concrete knowledge obtain all comparatively difficulty or expensive.Therefore, in these cases, the problem that object concrete knowledge is known little is needed to face.But, for existing most of Iterative Learning Control Algorithm, more or less all need certain object concrete knowledge to apply; That is, when object concrete knowledge is known little, existing most of Iterative Learning Control Algorithm is difficult to direct application.

When object concrete knowledge is known little, process during design iteration control learning algorithm, can be utilized at inputoutput data in the past batch.In conjunction with the feature that batch process reruns, the information of first previous batch or multiple batches can be used for improving the operation of next batch.Inputoutput data in the past batch, extract useful information from process and carry out design iteration control learning algorithm, mainly contain two kinds of thinkings: a kind of is directly according to inputoutput data in the past batch, calculate the input vector on new lot, such as, based on the Iterative Learning Control Algorithm of neuroid; Another kind is according to inputoutput data in the past batch, after calculating law of learning matrix, follows the general type of Iterative Learning Control Algorithm, calculates the input vector on new lot, such as nonparametric adaptive iterative learning control algorithm.

In general, when object concrete knowledge is known little, the existing achievement in research how designing suitable Iterative Learning Control Algorithm is little.

Summary of the invention

The object of the invention is propose a kind of based on reference to batch index time-varying gain type Iterative Learning Control Algorithm, the method manufactures production run for most of lots processed, can when object concrete knowledge be known little, by introducing in iterative learning control procedure suitable " reference batch ", and to reference to batch and the inputoutput data of present lot be used, carry out design iteration law of learning, thus the effective tracking realized exporting target trajectory, also introduce index time-varying gain simultaneously, thus improve the performance of track following transient process.

The present invention propose based on reference to batch index time-varying gain type Iterative Learning Control Algorithm, comprise the following steps:

Step 1) gather the history lot data of industrial installation, and obtain input and output sample after carrying out the conventional pre-service such as rejecting data bad value, be designated as U _k=[u _k(0), u _k(1) ..., u _k(N-1)] ^t∈ R ⁿ, Y _k=[y _k(1), y _k(2) ..., y _k(N)] ^t∈ R ⁿ, wherein k representative batch, N is the sample number of each batch, represents each batch of N number of data point collecting input and output amount, for history lot data calculating mean value as forming initial reference batch geometric locus U _s, Y _s; A note batch number k is initially 0, and note U _k=U _s, Y _k=Y _s; According to target trajectory curve Y _d, calculate the tracking error E after a kth batch production _k=Y _d-Y _k=[e _k(1), e _k(2) ..., e _k(N)] ^t; Specification error follows the tracks of the parameter such as index ε (ε is less arithmetic number), input minor shifts δ;

Step 2) utilize the production run that the collects input data after kth batch is produced, the input quantity U of generating reference batch _rk, wherein first judge the initial difference moment t in this batch _rs, then get t _rsinput quantity U after moment _kminor shifts δ (δ is non-zero real), obtain with reference to batch input vector U _rk, wherein each moment value u _rk(t) be:

$\left\{\begin{array}{cc}{u}_{\mathrm{rk}}\left(t\right)=u_k\left(t\right),& 0\&le;t<t_{\mathrm{rs}}\\ u_{\mathrm{rk}}\left(t\right)=u_k\left(t\right)+\mathrm{\&delta;},& t_{\mathrm{rs}}\&le;t<N\end{array}\right.$

Step 3) adopt step 2) in the input vector U of the reference that obtains batch _rkas the input of process units, process units is implemented, and gather this batch of corresponding output quantity data Y _rk;

Step 4) take out the production data (U of kth batch _k, Y _k) and step 3) middle with reference to a batch production data (U for input _rk, Y _rk) calculate, the input vector U of kth+1 batch is calculated according to iterative learning control law _k+1, wherein each moment value u _k+1(t) be

$\left\{\begin{array}{cc}{u}_{k+1}\left(t\right)=u_k\left(t\right)& 0\&le;t<t_{\mathrm{rs}}\\ u_{k+1}\left(t\right)=u_k\left(t\right)+\frac{u_k\left(t\right)-u_{\mathrm{rk}}\left(t\right)}{y_k(t+1)-y_{\mathrm{rk}}(t+1)}{e}^{-(t-t_{\mathrm{rs}})\mathrm{\&alpha;}}{e}_k(t+1)& t_{\mathrm{rs}}\&le;t<N\end{array}\right.$

Wherein with kth batch with reference to batch at the difference of the input in corresponding moment and the ratio of the difference of output long-pending with index time-varying gain as the learning gains of tracking error, α is learning gains parameter;

Step 5) adopt step 4) in the input quantity U of kth+1 batch that obtains _k+1as the input quantity of process units, process units is implemented, and gather this batch of corresponding output quantity data Y _rk, calculate tracking error E _k+1;

When judging after this batch of end of run, the convergence conditions of checking tracking error | E _k+1|≤| E _k|, if do not meet, regularized learning algorithm gain parameter α and the new reference locus (U of adjustment _s, Y _s), to follow the tracks of the change of process units adaptively, accelerating the speed of tracking target track, realizing the effective tracking to exporting target trajectory;

Step 6) if batch number k+1 does not reach the largest production batch number k of setting _max, then step 2 is returned), proceed iteration.If more than k _max, then show the maximum batch of number reaching the production of this Product processing of process units, can terminate.If new production restarts, then put k=0, algorithm returns step 1) restart.

Above-mentioned based on reference to batch index time-varying gain type Iterative Learning Control Algorithm in, described step 2) in, as follows with reference to the generation step of batch input quantity:

1. initial difference moment t is determined _rs, according to tracking convergence index ε (ε is less arithmetic number) of setting, judge whether present lot is less than this convergence index in the tracking error of moment t, thus obtain reference input vector U _rkwith present lot input vector U _kinitial difference moment t _rs:

$\left\{\begin{array}{cc}{t}_{\mathrm{rs}}=\min \left(t\right),& \mathrm{if}\left|e_k(t+1)\right|>\mathrm{\&epsiv;}\\ t_{\mathrm{rs}}=N,& \mathrm{if}\left|e_k(t+1)\right|<\mathrm{\&epsiv;for\; all}0\&le;t<N\end{array}\right.---\left(1\right)$

2. generating reference input vector U is calculated _rk, as 0≤t < t _rstime, keep present lot input vector U _kconstant; Work as t _rsduring≤t < N, get U _kminor shifts δ (δ is non-zero real), obtain reference input vector U _rk, wherein each moment value u _rk(t) be

$\left\{\begin{array}{cc}{u}_{\mathrm{rk}}\left(t\right)=u_k\left(t\right),& 0\&le;t<t_{\mathrm{rs}}\\ u_{\mathrm{rk}}\left(t\right)=u_k\left(t\right)+\mathrm{\&delta;},& t_{\mathrm{rs}}\&le;t<N\end{array}\right.---\left(2\right)$

If t _rs=N, show that tracking error that present lot engraves when all is all less than the convergence index of setting, namely present lot has achieved the requirement that target trajectory is followed the tracks of, and algorithm can not upgrade input quantity again.Under normal conditions, then t can be met _rs< N, the reference input that therefore will be calculated by formula (4) vector U _rkas system input, process units is implemented, obtains system output vector Y _rk, wherein batch U _rkbe called with reference to batch (reference batch); Batch U _kbe called basic batch (basic batch).

Above-mentioned based on reference to batch index time-varying gain type Iterative Learning Control Algorithm in, described step 4) Iterative Learning Control Algorithm Gain generating step as follows:

1. basic batch of U is taken out _kwith corresponding with reference to batch U _rkproduction data, and calculate this batch of U _ktracking error E _k+1;

2. basic batch of new input vector U is calculated _k+1, wherein as 0≤t < t _rstime, due to reference batch input quantity u _rkt () is constant, therefore basic batch also keeps input u _k+1t () is constant; Work as t _rsduring≤t < N, Iterative Learning Control Algorithm then with kth batch with reference to batch at the difference of the input in corresponding moment and the ratio of the difference of output long-pending with index time-varying gain as the learning gains of tracking error, α is learning gains parameter, is shown below:

$\left\{\begin{array}{cc}{u}_{k+1}\left(t\right)=u_k\left(t\right)& 0\&le;t<t_{\mathrm{rs}}\\ u_{k+1}\left(t\right)=u_k\left(t\right)+\frac{u_k\left(t\right)-u_{\mathrm{rk}}\left(t\right)}{y_k(t+1)-y_{\mathrm{rk}}(t+1)}{e}^{-(t-t_{\mathrm{rs}})\mathrm{\&alpha;}}{e}_k(t+1)& t_{\mathrm{rs}}\&le;t<N\end{array}\right.---\left(5\right)$

Wherein, for at t _rsthe index time-varying gain that≤t < N adopts, α is index time-varying gain coefficient, from formula (5), when calculating new batch of input quantity, only use only the inputoutput data collected, and do not need the priori understanding object, do not need the mathematical model setting up object yet.

Above-mentioned based on reference to batch index time-varying gain type Iterative Learning Control Algorithm, wherein said step 5) in, self-adaptative adjustment reference locus curve (U _s, Y _s) method be when kth+1 batch terminate after Gather and input output trajectory curve (U _k+1, Y _k+1), use output trajectory Y _k+1with reference locus Y _scompare calculating, obtain trajector deviation Δ Y _k+1=Y _k+1-Y _s, calculate the root-mean-square value of this deviation, then with the threshold xi preset, (ξ is non-zero arithmetic number) compares, if exceed this threshold xi, judges that production there occurs change, needs to upgrade reference locus, gets (U _s, Y _s) be the input and output geometric locus (U of current kth+1 batch _k+1, Y _k+1), to follow the tracks of the change of production adaptively, wherein threshold xi can provide in advance in conjunction with the operation situation of change of actual production device.

The present invention propose based on reference to batch index time-varying gain type Iterative Learning Control Algorithm, have the following advantages:

1, the present invention utilize proposition based on reference to batch carrying out design iteration study control rule, thus can to use when object specifying information is known little about it, thus greatly to have widened range of application.

2, the present invention passes through to introduce index time-varying gain thus the transiting performance changing Iterative Learning Control Algorithm.This method is skillfully constructed, simple and practical, can be widely used in the high precision geometric locus tracing control of industrial production line batch production process.

Embodiment

The present invention propose a kind of with reference to batch index time-varying gain type Iterative Learning Control Algorithm, comprise the following steps:

1) gather the history lot data of industrial installation, and obtain input and output sample after carrying out the conventional pre-service such as rejecting data bad value, be designated as U _k=[u _k(0), u _k(1) ..., u _k(N-1)] ^t∈ R ⁿ, Y _k=[y _k(1), y _k(2) ..., y _k(N)] ^t∈ R ⁿ, wherein k representative batch, N is the sample number of each batch, represents each batch of N number of data point collecting input and output amount.For history lot data calculating mean value as forming initial reference batch geometric locus U _s, Y _s; A note batch number k is initially 0, and note U _k=U _s, Y _k=Y _s; According to target trajectory curve Y _d, calculate the tracking error E after a kth batch production _k=Y _d-Y _k=[e _k(1), e _k(2) ..., e _k(N)] ^t; Specification error follows the tracks of the parameter such as index ε (ε is less arithmetic number), input minor shifts δ;

2) production run that the collects input data after kth batch is produced are utilized, the input quantity U of generating reference batch _rk, wherein first judge the initial difference moment t in this batch _rs, then get t _rsinput quantity U after moment _kminor shifts δ (δ is non-zero real), obtain with reference to batch input vector U _rk, wherein each moment value u _rk(t) be

$\left\{\begin{array}{cc}{u}_{\mathrm{rk}}\left(t\right)=u_k\left(t\right),& 0\&le;t<t_{\mathrm{rs}}\\ u_{\mathrm{rk}}\left(t\right)=u_k\left(t\right)+\mathrm{\&delta;},& t_{\mathrm{rs}}\&le;t<N\end{array}\right.---\left(1\right)$

3) adopt step 2) in the input vector U of the reference that obtains batch _rkas the input of process units, process units is implemented, and gather this batch of corresponding output quantity data Y _rk;

4) production data (U of individual batch of kth is taken out _k, Y _k) and step 3) middle with reference to a batch production data (U for input _rk, Y _rk) calculate, the input vector U of kth+1 batch is calculated according to iterative learning control law _k+1, wherein each moment value u _k+1(t) be

$\left\{\begin{array}{cc}{u}_{k+1}\left(t\right)=u_k\left(t\right)& 0\&le;t<t_{\mathrm{rs}}\\ u_{k+1}\left(t\right)=u_k\left(t\right)+\frac{u_k\left(t\right)-u_{\mathrm{rk}}\left(t\right)}{y_k(t+1)-y_{\mathrm{rk}}(t+1)}{e}^{-(t-t_{\mathrm{rs}})\mathrm{\&alpha;}}{e}_k(t+1)& t_{\mathrm{rs}}\&le;t<N\end{array}\right.---\left(2\right)$

Wherein with kth batch with reference to batch at the difference of the input in corresponding moment and the ratio of the difference of output long-pending with index time-varying gain as the learning gains of tracking error, α is learning gains parameter.

5) adopt step 4) in the input quantity U of kth+1 batch that obtains _k+1as the input quantity of process units, process units is implemented, and gather this batch of corresponding output quantity data Y _rk, calculate tracking error E _k+1.

When judging after this batch of end of run, the convergence conditions of checking tracking error | E _k+1|≤| E _k|, if do not meet, regularized learning algorithm gain parameter α and the new reference locus (U of adjustment _s, Y _s), to follow the tracks of the change of process units adaptively, accelerating the speed of tracking target track, realizing the effective tracking to exporting target trajectory.

6) if batch number k+1 does not reach the largest production batch number k of setting _max, then step 2 is returned), proceed iteration.If more than k _max, then show the maximum batch of number reaching the production of this Product processing of process units, can terminate.If new production restarts, then put k=0, algorithm returns step 1) restart.

Described step 2) in, the generation step with reference to batch input quantity is as follows: 1. determine initial difference moment t _rs.According to tracking convergence index ε (ε is less arithmetic number) of setting, judge whether present lot is less than this convergence index in the tracking error of moment t, thus obtain reference input vector U _rkwith present lot input vector U _kinitial difference moment t _rs:

$\left\{\begin{array}{cc}{t}_{\mathrm{rs}}=\min \left(t\right),& \mathrm{if}\left|e_k(t+1)\right|>\mathrm{\&epsiv;}\\ t_{\mathrm{rs}}=N,& \mathrm{if}\left|e_k(t+1)\right|<\mathrm{\&epsiv;for\; all}0\&le;t<N\end{array}\right.---\left(3\right)$

2. generating reference input vector U is calculated _rk.As 0≤t < t _rstime, keep present lot input vector U _kconstant; Work as t _rsduring≤t < N, get U _kminor shifts δ (δ is non-zero real), obtain reference input vector U _rk, wherein each moment value u _rk(t) be

$\left\{\begin{array}{cc}{u}_{\mathrm{rk}}\left(t\right)=u_k\left(t\right),& 0\&le;t<t_{\mathrm{rs}}\\ u_{\mathrm{rk}}\left(t\right)=u_k\left(t\right)+\mathrm{\&delta;},& t_{\mathrm{rs}}\&le;t<N\end{array}\right.---\left(4\right)$

If t _rs=N, show that tracking error that present lot engraves when all is all less than the convergence index of setting, namely present lot has achieved the requirement that target trajectory is followed the tracks of, and algorithm can not upgrade input quantity again.Under normal conditions, then t can be met _rs< N, the reference input that therefore will be calculated by formula (4) vector U _rkas system input, process units is implemented, obtains system output vector Y _rk.This operation batch is called with reference to batch (reference batch) U _rk; Accordingly, batch basic batch of (basic batch) U is called in order to generating reference input vector _k.

Described step 4) in, Iterative Learning Control Algorithm Gain generating step is as follows: 1. take out basic batch of U _kwith corresponding with reference to batch U _rkproduction data, and calculate this batch of U _ktracking error E _k+1; 2. basic batch of new input vector U is calculated _k+1, wherein as 0≤t < t _rstime, due to reference batch input quantity u _rkt () is constant, therefore basic batch also keeps input u _k+1t () is constant; Work as t _rsduring≤t < N, Iterative Learning Control Algorithm then with kth batch with reference to batch at the difference of the input in corresponding moment and the ratio of the difference of output long-pending with index time-varying gain as the learning gains of tracking error, α is learning gains parameter, is shown below:

$\left\{\begin{array}{cc}{u}_{k+1}\left(t\right)=u_k\left(t\right)& 0\&le;t<t_{\mathrm{rs}}\\ u_{k+1}\left(t\right)=u_k\left(t\right)+\frac{u_k\left(t\right)-u_{\mathrm{rk}}\left(t\right)}{y_k(t+1)-y_{\mathrm{rk}}(t+1)}{e}^{-(t-t_{\mathrm{rs}})\mathrm{\&alpha;}}{e}_k(t+1)& t_{\mathrm{rs}}\&le;t<N\end{array}\right.---\left(5\right)$

Wherein, for at t _rsthe index time-varying gain that≤t < N adopts, α is index time-varying gain coefficient.From formula (5), when calculating new batch of input quantity, only use only the inputoutput data collected, and do not need the priori understanding object, also do not need the mathematical model setting up object.Method is very effectively simple.

Described step 5) in, in order to follow the tracks of the change of process units, self-adaptative adjustment reference locus curve (U _s, Y _s) method be when kth+1 batch terminate after Gather and input output trajectory curve (U _k+1, Y _k+1), use output trajectory Y _k+1with reference locus Y _scompare calculating, obtain trajector deviation Δ Y _k+1=Y _k+1-Y _s, calculate the root-mean-square value (RMSE) of this deviation, then compare with the threshold xi preset (ξ is non-zero arithmetic number), if exceed this threshold xi, judge that production there occurs change, need to upgrade reference locus, get (U _s, Y _s) be the input and output geometric locus (U of current kth+1 batch _k+1, Y _k+1), to follow the tracks of the change of production adaptively.Wherein threshold xi can provide in advance in conjunction with the operation situation of change of actual production device.

Claims (4)

1. based on reference to batch an index time-varying gain type Iterative Learning Control Algorithm, it is characterized in that comprising the following steps:

Step 1) gather the history lot data of industrial installation, and obtain input and output sample after carrying out the conventional pre-service such as rejecting data bad value, be designated as U _k=[u _k(0), u _k(1) ..., u _k(N-1)] ^t∈ R ⁿ, Y _k=[y _k(1), y _k(2) ..., y _k(N)] ^t∈ R ⁿ, wherein k representative batch, N is the sample number of each batch, represents each batch of N number of data point collecting input and output amount, for history lot data calculating mean value as forming initial reference batch geometric locus U _s, Y _s; A note batch number k is initially 0, and note U _k=U _s, Y _k=Y _s; According to target trajectory curve Y _d, calculate the tracking error E after a kth batch production _k=Y _d-Y _k=[e _k(1), e _k(2) ..., e _k(N)] ^t; Specification error follows the tracks of the parameter such as index ε (ε is less arithmetic number), input minor shifts δ;

Step 2) utilize the production run that the collects input data after kth batch is produced, the input quantity U of generating reference batch _rk, wherein first judge the initial difference moment t in this batch _rs, then get t _rsinput quantity U after moment _kminor shifts δ (δ is non-zero real), obtain with reference to batch input vector U _rk, wherein each moment value u _rk(t) be:

$\left\{\begin{array}{cc}{u}_{\mathrm{rk}}\left(t\right)=u_k\left(t\right),& 0\&le;t<t_{\mathrm{rs}}\\ u_{\mathrm{rk}}\left(t\right)=u_k\left(t\right)+\mathrm{\&delta;},& t_{\mathrm{rs}}\&le;t<N\end{array}\right.$

Step 3) adopt step 2) in the input vector U of the reference that obtains batch _rkas the input of process units, process units is implemented, and gather this batch of corresponding output quantity data Y _rk;

Step 4) take out the production data (U of kth batch _k, Y _k) and step 3) middle with reference to a batch production data (U for input _rk, Y _rk) calculate, the input vector U of kth+1 batch is calculated according to iterative learning control law _k+1, wherein each moment value u _k+1(t) be

$\left\{\begin{array}{cc}{u}_{k+1}\left(t\right)=u_k\left(t\right)& 0\&le;t<t_{\mathrm{rs}}\\ u_{k+1}\left(t\right)=u_k\left(t\right)+\frac{u_k\left(t\right)-u_{\mathrm{rk}}\left(t\right)}{y_k(t+1)-y_{\mathrm{rk}}(t+1)}{e}^{-(t-t_{\mathrm{rs}})\mathrm{\&alpha;}}{e}_k(t+1)& t_{\mathrm{rs}}\&le;t<N\end{array}\right.$

Wherein with kth batch with reference to batch at the difference of the input in corresponding moment and the ratio of the difference of output e and index time-varying gain long-pending as the learning gains of tracking error, α is learning gains parameter;

Step 5) adopt step 4) in the input quantity U of kth+1 batch that obtains _k+1as the input quantity of process units, process units is implemented, and gather this batch of corresponding output quantity data Y _rk, calculate tracking error E _k+1;

When judging after this batch of end of run, the convergence conditions of checking tracking error | E _k+1|≤| E _k|, if do not meet, regularized learning algorithm gain parameter α and the new reference locus (U of adjustment _s, Y _s), to follow the tracks of the change of process units adaptively, accelerating the speed of tracking target track, realizing the effective tracking to exporting target trajectory;

Step 6) if batch number k+1 does not reach the largest production batch number k of setting _max, then step 2 is returned), proceed iteration.If more than k _max, then show the maximum batch of number reaching the production of this Product processing of process units, can terminate.If new production restarts, then put k=0, algorithm returns step 1) restart.

2. as claimed in claim 1 based on reference to batch index time-varying gain type Iterative Learning Control Algorithm, it is characterized in that: described step 2) in, the generation step with reference to batch input quantity is as follows:

1. initial difference moment t is determined _rs, according to tracking convergence index ε (ε is less arithmetic number) of setting, judge whether present lot is less than this convergence index in the tracking error of moment t, thus obtain reference input vector U _rkwith present lot input vector U _kinitial difference moment t _rs:

$\left\{\begin{array}{cc}{t}_{\mathrm{rs}}=\min \left(t\right),& \mathrm{if}\left|e_k(t+1)\right|>\mathrm{\&epsiv;}\\ t_{\mathrm{rs}}=N,& \mathrm{if}\left|e_k(t+1)\right|<\mathrm{\&epsiv;for\; all}0<t<N\end{array}\right.---\left(1\right)$

2. generating reference input vector U is calculated _rk, as 0≤t < t _rstime, keep present lot input vector U _kconstant; Work as t _rsduring≤t < N, get U _kminor shifts δ (δ is non-zero real), obtain reference input vector U _rk, wherein each moment value u _rk(t) be

$\left\{\begin{array}{cc}{u}_{\mathrm{rk}}\left(t\right)=u_k\left(t\right),& 0\&le;t<t_{\mathrm{rs}}\\ u_{\mathrm{rk}}\left(t\right)=u_k\left(t\right)+\mathrm{\&delta;},& t_{\mathrm{rs}}\&le;t<N\end{array}\right.---\left(2\right)$

If t _rs=N, show that tracking error that present lot engraves when all is all less than the convergence index of setting, namely present lot has achieved the requirement that target trajectory is followed the tracks of, and algorithm can not upgrade input quantity again.Under normal conditions, then t can be met _rs< N, the reference input that therefore will be calculated by formula (4) vector U _rkas system input, process units is implemented, obtains system output vector Y _rk, wherein batch U _rkbe called with reference to batch (reference batch); Batch U _kbe called basic batch (basic batch).

3. as claimed in claim 1 based on reference to batch index time-varying gain type Iterative Learning Control Algorithm, it is characterized in that: described step 4) in, Iterative Learning Control Algorithm Gain generating step is as follows:

1. basic batch of U is taken out _kwith corresponding with reference to batch U _rkproduction data, and calculate this batch of U _ktracking error E _k+1;

2. basic batch of new input vector U is calculated _k+1, wherein as 0≤t < t _rstime, due to reference batch input quantity u _rkt () is constant, therefore basic batch also keeps input u _k+1t () is constant; Work as t _rsduring≤t < N, Iterative Learning Control Algorithm then with kth batch with reference to batch at the difference of the input in corresponding moment and the ratio of the difference of output long-pending with index time-varying gain as the learning gains of tracking error, α is learning gains parameter, is shown below:

$\left\{\begin{array}{cc}{u}_{k+1}\left(t\right)=u_k\left(t\right)& 0\&le;t<t_{\mathrm{rs}}\\ u_{k+1}\left(t\right)=u_k\left(t\right)+\frac{u_k\left(t\right)-u_{\mathrm{rk}}\left(t\right)}{y_k(t+1)-y_{\mathrm{rk}}(t+1)}{e}^{-(t-t_{\mathrm{rs}})\mathrm{\&alpha;}}{e}_k(t+1)& t_{\mathrm{rs}}\&le;t<N\end{array}\right.---\left(5\right)$

Wherein, for at t _rsthe index time-varying gain that≤t < N adopts, α is index time-varying gain coefficient, from formula (5), when calculating new batch of input quantity, only use only the inputoutput data collected, and do not need the priori understanding object, do not need the mathematical model setting up object yet.

4. as claimed in claim 1 based on reference to batch index time-varying gain type Iterative Learning Control Algorithm, it is characterized in that: described step 5) in, self-adaptative adjustment reference locus curve (U _s, Y _s) method be when kth+1 batch terminate after Gather and input output trajectory curve (U _k+1, Y _k+1), use output trajectory Y _k+1compare calculating with reference locus Ys, obtain trajector deviation Δ Y _k+1=Y _k+1-Y _s, calculate the root-mean-square value of this deviation, then with the threshold xi preset, (ξ is non-zero arithmetic number) compares, if exceed this threshold xi, judges that production there occurs change, needs to upgrade reference locus, gets (U _s, Y _s) be the input and output geometric locus (U of current kth+1 batch _k+1, Y _k+1), to follow the tracks of the change of production adaptively, wherein threshold xi can provide in advance in conjunction with the operation situation of change of actual production device.