CN1178015A - Process controlling method and device - Google Patents

Abstract

In order to control a process, correcting variables for several actuators that influence the process are calculated in a regulating device from mesured starting values of the process. In order to optimise the regulation of the process by the regulating device, the effectiveness (W11,... Wnm) of various actuators is learned in a neutonal network. The effectiveness describes for each actuator the extent to which changes in the correcting variables (dul..., dun) cause changes in the starting values (dyl, ..., dym). The thus learned effectiveness is supplied to the regulating device to improve the calculation of the correcting rariables (u1, ...,un).

Description

Course control method for use and equipment

The present invention relates to a kind of course control method for use, wherein in a regulating device, calculate the regulated quantity of a plurality of governor motions that process is worked by the measured value of process output quantity.The invention still further relates to corresponding a kind of equipment.

A kind of like this method and a kind of like this equipment also are the themes of a border application PCT/DE94/00028 of old-national, and this applies for still unexposed before the application's priority date always.

When process control, be typically provided with a plurality of governor motions, so that definite output quantity of influence process.To calculate the regulated quantity of each governor motion in the regulating device according to controlled quentity controlled variable in this case, the output quantity of process should obtain the value of this controlled quentity controlled variable.For a kind of like this example of process control is to roll the band flatness in rolling equipment, wherein to the influence of rolling seam and thus to the influence of rolling the band flatness be according to the measured value that is rolled the band flatness that distributes on the width of band and in addition by the waving of roll, crooked, axial displacement and or local the cooling realize.

Has the parameter of providing of very big importance for the well-tuned of process output quantity, just the parameter that how output quantity is worked about each governor motion about the regulating device of governor motion performance.Here, there is a problem, promptly in commercial unit, often can not provide or provide the parameter of not enough associated regulating mechanism.

What know usually is, when metal band rolling, the flatness of band is rolled the governor motion of seam and is enhanced by operating a plurality of changes simultaneously.

By Hitachi's magazine (Hitachi Review) 41 (1992) 1, the the 31st to 38 page known in addition, in rolling machine frame in order to have disposed the typical sample cross section of band flatness respectively for each governor motion of rolling machine frame with the adjusting of flatness, by it and according to knowing, can be improved by operating corresponding governor motion about the experimental knowledge of the operation of rolling.In the back of rolling machine frame, measure the actual cross-section of band flatness, and the measured value that will obtain like this delivers to a neural network, response will provide this network as network: what kind of ratio measured band flatness cross section is made up of with each sample in cross section.The ratio that obtains is like this combined with the conditioning signal of regulating machinery in Fuzzy Controller.This known method is based on the experimental knowledge of rolling personnel about the governor motion function largely, but to control for optimizing process be relatively general and coarse for this.In addition, the experimental knowledge that is obtained in adjusting links mutually with corresponding apparatus, and can not directly be diverted to miscellaneous equipment.

The objective of the invention is to: provide a kind of method and apparatus that is used for process control, it can not rely on existing about process experimental knowledge and the Optimal Control of implementation procedure.

Purpose of the present invention will solve by the equipment that provides in the method that provides in the claim 1 and the claim 10.

Further favourable design according to the inventive method and equipment is provided by dependent claims.

Utilize the method according to this invention and corresponding device thereof, when the conventional control of process, to automatically learn the governor motion performance of each governor motion according to the variation of regulated quantity of measuring during the process operation and process output quantity, and the governor motion performance will be input to regulating device for the calculating that improves regulated quantity.And constantly to learning by the performance of that governor motion of regulating device control.If regulating device only is a part of having controlled governor motion, also only the performance of this part governor motion is learnt.Remaining governor motion performance will be learnt when affiliated governor motion input is used again.By the governor motion performance of such study, regulating device is supported that when it calculates regulated quantity regulating device itself can not intervened.The method according to this invention also can be applicable on the existing equipment, and wherein existing there regulating device is still worked in usual mode.

For the study of each governor motion performance is to change according to the variation of the regulated quantity of being carried out by regulating device during the process conventional operation and consequent output quantity to realize in a neural network.Like this under the constant situation in process working point, funtcional relationship between the variation of process output quantity and the variation of regulated quantity can be linearized in small signal region, nervous system can only have an input layer and an output layer thus, and simple structure is correspondingly arranged therewith.Regulated quantity to its input side input in this neural network changes du _iWill be by corresponding to the network parameter w that waits to learn the governor motion performance _IjAnd according to relational expression ${d\tilde{y}}_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}\·{\mathrm{du}}_i$ Interrelate with the network response, and change network parameter w according to the principle that reduces this deviation according to the deviation between network response and the output quantity variation _IjThe learning process of study governor motion performance can be followed the variation at a slow speed of process working point in neural network.Be the governor motion performance that learn a working point in addition, especially during quick or big variation, store in the working point; When process has the working point that the governor motion performance of being stored learnt again, these governor motion performances will be re-entered neural network as network parameter.

For the relation between the variable-operation point of considering governor motion performance to be learnt and process all sidedly, be preferably in the work support point of determining predetermined number in the changeable perform region, working point, wherein regulated quantity in neural network input side input being changed use respectively a weighting coefficient weighting, this weighting coefficient for each work fulcrum is the measuring of distance between real work point and the corresponding work point.In neural network, this is weighted coefficient O then _aWeighting regulated quantity change O _aDu _iWill be according to relational expression: ${d\tilde{y}}_j=\underset{a=1}{\overset{s}{\mathrm{\Σ}}}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{aij}}\·o_a\·{\mathrm{du}}_i$ Respond with network

Interrelate w in the formula _AijBe the network parameter of neural network may joint, and $\underset{a=1}{\overset{s}{\mathrm{\Σ}}}{o}_a\·w_{\mathrm{aij}}=w_{\mathrm{ij}}$ It is governor motion performance to be learnt.In order to make neural network adaptive, will change network parameter w according to the principle that reduces this deviation according to the deviation between network response and the variation of process output quantity _Aij

May have noise because be input to the regulated quantity variation of neural network as input quantity, preferably at first be entered in the wave filter, be suppressed by neural network so that be lower than the regulated quantity variation of predetermined threshold.

Better and the convergence faster for acquisition when learning the governor motion performance, best study step-length for each change network parameter is considered the data set of a plurality of regulated quantitys, network response and output quantity respectively.

Below the present invention is described in detail for the example that will regulate with band flatness in the rolling equipment, in the accompanying drawing:

Fig. 1 is an an embodiment of the apparatus according to the invention, and it has the regulating device of accommodation zone flatness in rolling equipment and is used to learn the neural network of governor motion performance;

Fig. 2 is an employed neural network example of working point study governor motion performance for process;

Fig. 3 is an example of study governor motion performance under the situation of using a plurality of data sets;

Fig. 4 is for employed another neural network example of working point study governor motion performance that changes;

Fig. 5 is used for illustrating the synoptic diagram of determining the weighting coefficient relevant with the study of governor motion performance in the neural network of Fig. 4.

Fig. 1 represents an example of present device, and it is used for process control, is used for here rolling with 1 flatness at a rolling equipment regulating.Rolling equipment can have a plurality of rolling machine frames, has described here wherein with the last frame 2 on 1 the direct of travel; It is provided with back up roll 3, central roll 4 and work roll 5. Roll 3,4 and 5 and known and being used to of therefore not being depicted as itself regulated, waves, device crooked and mobile roll and the predetermined roll zones of cooling is configured for influencing the band flatness together various governor motions.Be provided with measuring device for testing alignment 6 in the back of rolling machine frame 2, it is the flatness parameter y to reduce a vector y and to distribute on bandwidth continuously ₁..., y _mForm measure the actual flatness rolled with 1.Flatness measurement by measurement mechanism 6 for example can realize like this, promptly measures with the band pressure distribution on 1 width with unshowned measurement roller here.These flatness parameters constitute the output quantity of process, and this output quantity is transmitted to the input side of a regulating device 7.This regulating device is by m output quantity y ₁..., y _mCalculate regulated quantity u ₁..., u _n, come n governor motion of controlled rolling frame 2 by these regulated quantitys by a control corresponding equipment 8, so that influence band flatness y.

In order to support by the adjusting on 7 pairs of bands of regulating device flatness and to optimize, during the process of operation, study governor motion performance in a calculation element 9 W=(W ₁₁..., W _Nm) and send it to regulating device 7, this governor motion performance specification the output quantity variation Dy(dy ₁..., dy _m) depend on that regulated quantity changes Du=(du ₁..., du _n).In this case, only need the adjusting function of study constantly all the time at each by those governor motions of regulating device 7 or hand-guided.The governor motion performance W that these are to be calculated ₁₁..., W _NmWhat always depend on rolling equipment passes through the working point parameter b=(b ₁..., b _k) definition the real work point.Here related to and removed regulated quantity u ₁..., u _nOutward to the influential amount of influence of process.These amounts of influence are bandwidth in the shown herein present embodiment, the diameter of roll-force and roll.

Neural network 10 changes the regulated quantity of input side input DuAnd running parameter in case of necessity bRespond with network $\underset{-}{d\tilde{y}}=({d\tilde{y}}_1,...,{d\tilde{y}}_m)$ Interrelate, the latter changes with measured output quantity in a comparison means 11 DyCompare.Change according to the output quantity of trying to achieve like this DyRespond with network

Between deviation e=(e ₁..., e _m), and the network parameter of neural network 10 is changed according to the principle that reduces this deviation e.The network response of in neural network 10, being learnt The relative adjustment quantitative changeization DuRatio corresponding to governor motion performance to be learnt W, and this ratio is input to is used in the regulating device 7 improving to regulated quantity uCalculating.

Output quantity y _j, j=1 ..., m and regulated quantity u _i, i=1 ..., n, and working point parameter b ₁, l=1 ..., the funtcional relationship between the k can followingly be represented:

y _j＝f _j(u ₁，...，u _n，b ₁，...，b _k)。

Can linearization obtain like this with concerning in small signal region: ${\mathrm{dy}}_j=\frac{\∂\mathrm{fj}}{{\∂u}_1}\·{\mathrm{du}}_1+\frac{\∂\mathrm{fj}}{{\∂u}_2}\·{\mathrm{du}}_2+...+\frac{\∂\mathrm{fj}}{{\∂u}_n}\·{\mathrm{du}}_n$ In the formula $\frac{\∂\mathrm{fj}}{{\∂u}_i}=w_{\mathrm{ij}}$ Be the working point parameter b ₁..., b _kFunction.Adopt other literary style therefore to obtain: Dy= w Du, in the formula wBe a m * n matrix, its coefficient w _IjBe equivalent to the governor motion performance.

Be used in Fig. 2 presentation graphs 1 each steady job point study governor motion performance w of process _IjThe example of neural network 10.Because the working point is indeclinable, governor motion performance w to be learnt _IjWith the working point parameter b ₁..., b _kIrrelevant.

Neural network has an input level, and it is provided with n and is used for each regulated quantity variation du _iInput block 13.Input block 13 plays wave filter here, is lower than predetermined threshold X, and for example 2% of the maximal regulated stroke regulated quantity changes du _iTo be suppressed, so as from the output quantity of neural network the elimination noise component.This neural network also has an output level, and it is provided with m and process output quantity y _jThe corresponding output unit 14 of number.Filtered input quantity du _iFor each output unit 14 each by a network parameter w _IjWeighting, and follow total addition network response: ${d\tilde{y}}_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}\·{\mathrm{du}}_i$ Network parameter w wherein _IjCorresponding to governor motion performance w to be learnt _Ij

The training of neural network realizes according to contrary propagation algorithm, at this moment always seeks error function by the decline of gradient direction $E=0.5{e^2}_j=0.5{({\mathrm{dy}}_j-{d\tilde{y}}_j)}^2$ Minimum value.As the network response is described by Fig. 1

In comparison means 11, change dy with the output quantity of measuring _jCompare, wherein in subordinate's learning algorithm 12, reducing to make network parameter w on the direction of error E _Ijη changes suitablely with the study step-length.Obtain the adaptation step of following formula like this: ${\mathrm{dw}}_{\mathrm{ij}}=-\mathrm{\η}\frac{\∂E}{{\∂w}_{\mathrm{ij}}}=\mathrm{\η}({\mathrm{dy}}_j-{d\tilde{y}}_j)\·{\mathrm{du}}_i$

For governor motion performance w being provided for regulating device 7 _Ij, on the input block 13 of neural network 10, will import as the n of input quantity the vector model with each diverse location of value " 1 " (0 ..., 0,1,0, ..., 0) so that appearance is as the governor motion performance w that is learnt of network response on output unit 14 _Ij, and be used to send to regulating device 7.

To explain by Fig. 3 as following, at study governor motion performance w _IjIn the time, can obtain better and convergence faster, at this moment changes network parameter w for each _IjThe study step-length, will consider the data set of a plurality of regulated quantitys Du(t ₁) ..., Du(t _p), a plurality of network response data groups $\underset{-}{d\tilde{y}}\left(t_1\right),...,\underset{-}{d\tilde{y}}\left(t_p\right)$ And the data set of a plurality of process output quantities variations Dy(t ₁) ..., Dy(t _p).Here will obtain adaptation step to following formula: ${\mathrm{dw}}_{\mathrm{ij}}=-\mathrm{\η}\frac{\∂E}{{\∂w}_{\mathrm{ij}}}=\mathrm{\η}\·\underset{r=1}{\overset{p}{\mathrm{\Σ}}}({\mathrm{dy}}_j\left(t_r\right)-{d\tilde{y}}_j\left(t_r\right))\·{\mathrm{du}}_i\left(t_r\right)$

As stating, utilize the neural network shown in Fig. 2 to learn governor motion performance w to a steady job point of process _IjWhen this working point just slowly changed, learning process can be got caught up in this variation.When disturbances of power is fast or big, promptly above-mentioned working point parameter b ₁..., b _kVariation when fast or big, must be at study governor motion performance w _IjThe time consider it and working point parameter b ₁..., b _kRelation.

Fig. 4 represents to satisfy an example of the neural network 10 of this requirement.This neural network has an input level, and it is provided with n and is used for each regulated quantity variation du _i Input block 15 and k be used for each working point parameter b ₁Other input block 16.Input block 15 plays wave filter, and they suppress to be lower than the regulated quantity variation du of reservation threshold X _i, so that from the input quantity of neural network the elimination noise component.

Be provided with the overlayer with unit 17 and 18 in the input level back with input block 15 and 16, its function is at first explained by Fig. 5.The working point parameter b ₁, l=1 ..., k has determined a k dimension space, among Fig. 5 for for the purpose of simplifying description, express the situation of k=2.In this k dimension space, determined the work fulcrum P of predetermined number _a, a=1 ..., S.In illustrated example, work fulcrum P _aDetermine the working point parameter b by each q=5 _IDifferent value, therefore for work fulcrum P _aNumber S S=q is arranged ^kFor each work fulcrum P _aBe provided with relative each working point parameter b respectively ₁Partition function h _A1, this function is at relevant work fulcrum P _aThe position on have value " 1 ", and continue to drop to value " 0 " by this work fulcrum to direct neighbor.

As shown in Figure 5, each the real work point for process always has 2 ^kIndividual, be 4 work fulcrum Ps immediate here with it _a, express these fulcrums in the drawings emphatically.Working point P is to the work fulcrum P of direct neighbor _aThe degree of approach will be by partition function h under on the position of working point P _A1Value define, for example for illustrated work fulcrum P _aIt and a weighting coefficient O _a=f (h _a, h _A2) interrelate.So this weighting coefficient O _a, 0≤O _a≤ 1, be that working point P is to the fulcrum P that works _aMeasuring of distance.

Now, the overlayer of the neural network shown in Fig. 4 is for each work fulcrum P _aHave a unit 17 respectively, it is by the working point parameter b that inputs to it and define real work point P ₁Calculate weighting coefficient O _aFor S work fulcrum P _aIn each will use weighting coefficient O in ns unit 18 altogether _aRegulated quantity to filtering changes du _iWeighting.

This neural network also has an output level, and it is provided with and process output quantity y _jThe corresponding m of a number m output unit 19.For each output unit 19, the regulated quantity that is weighted changes O _aDu _iRespectively by a network parameter w _AijWeighting again, and then be summed into the network response: ${d\tilde{y}}_j=\underset{a=1}{\overset{s}{\mathrm{\Σ}}}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{aij}}\·o_a\·{\mathrm{du}}_i$ Network parameter w _AijTo adaptively realizing that real process changes in the mode identical with above-mentioned Fig. 2 neural network example, but amount of bias e wherein _jUse weighting coefficient O _aBe weighted to: O _aE _j

For governor motion performance w to be learnt _IjHave: $\underset{a=1}{\overset{s}{\mathrm{\Σ}}}{o}_a\·w_{\mathrm{aij}}=w_{\mathrm{ij}}$ It will be also as offering regulating device 7 by this way the example of the neural network among Fig. 2 situation, promptly to n the vector model (0 respectively having value " 1 " diverse location of input block 15 inputs as input quantity, ..., 0,1,0 ..., 0), on output unit 19, produces the governor motion performance w that is learnt that responds as network thus _Ij, it is transmitted to regulating device 7 then.

Same under the situation of neural network shown in Figure 4 about the performance w of study governor motion _IjAlso can reach better and convergence faster,, always consider the data set of a plurality of regulated quantitys at this moment as explaining according to Fig. 3 Du(t _p), a plurality of network response data groups $\underset{-}{d\tilde{y}}\left(t_1\right),...,\underset{-}{d\tilde{y}}\left(t_p\right)$ And the data set of a plurality of process output quantities variations Dy(t ₁) ..., Dy(t _p).

Claims (13)

1. course control method for use, wherein in regulating device (7) by the measured value (y of process output quantity _j) calculate the regulated quantity (u of a plurality of governor motions that process is worked _i), it is characterized in that: study governor motion performance (w in a neural network (10) _Ij), for each governor motion, this performance specification output quantity changes (dy _j) change (du with regulated quantity _i) relation; And with the governor motion performance (w that learns _Ij) be input to regulating device (7), to improve regulated quantity (u _i) calculating.

2. method according to claim 1 is characterized in that: neural network (10) by with the corresponding network parameter (w of governor motion performance to be learnt _Ij) and according to relational expression ${d\tilde{y}}_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}\·{\mathrm{du}}_i$ Make the regulated quantity that inputs to it at input side change (du _i) respond with network

Interrelate; And respond according to network

And output quantity changes (dy _j) between deviation (e _j) make network parameter (w _Ij) according to reducing this deviation (e _j) principle change.

3. method according to claim 1 is characterized in that: at governor motion performance (w to be learnt _Ij) and process working point parameter (b ₁) under the relevant situation, by working point parameter (b ₁) determine work fulcrum (P in the hyperspace of definition _a) preset number; For each work fulcrum (P _a) make each regulated quantity of on neural network (10) input side, importing change (du _i) usefulness weighting coefficient (O _a) weighting, this weighting coefficient is at the actual value (b by the working point parameter ₁) working point (P) and the corresponding work fulcrum (P of definition _a) between the measuring of distance; The regulated quantity that is weighted changes (O _aDu _i) according to relational expression ${d\tilde{y}}_j=\underset{a=1}{\overset{s}{\mathrm{\Σ}}}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{aij}}\·o_a\·{\mathrm{du}}_i$ Respond with network

Interrelate w in the formula _AijBe adjustable network parameter, and $\underset{a=1}{\overset{s}{\mathrm{\Σ}}}{o}_a\·w_{\mathrm{aij}}=w_{\mathrm{ij}}$ It is governor motion performance to be learnt; And respond according to network

Change (dy with output quantity _j) between deviation (e _j) according to reducing this deviation (e _j) principle change network parameter (w _Aij).

4. method according to claim 3 is characterized in that: the value (O of weighting coefficient _a) along with working point (P) to relevant work fulcrum (P _a) distance increase, constantly from 1 null value that drops on each direct neighbor work position of the fulcrum.

5. according to each described method in the claim 2 to 4, it is characterized in that: realize network parameter (w according to contrary propagation algorithm _Ij, w _Aij) change, error function at this moment $E=0.5\·{({\mathrm{dy}}_i-{d\stackrel{\″}{y}}_j)}^2$ Minimum value try to achieve by the decline on the gradient direction.

6. according to each described method in the claim 2 to 5, it is characterized in that: for the governor motion performance (w that will be learnt _Ij) send regulating device (7) to, to neural network (10) input conduct and regulated quantity (u _i) a plurality of model values of the corresponding input quantity of number, these model values have value 1 on each diverse location, and remaining is value zero, wherein produces governor motion performance (w as the network response at the outgoing side of neural network (10) _Ij).

7. according to the described method of above-mentioned each claim, it is characterized in that: the regulated quantity to neural network (10) input changes (du _i) at first import a wave filter, make the regulated quantity that is lower than a predetermined threshold (X) change (du _i) suppressed by neural network (10).

8. according to the described method of above-mentioned each claim, it is characterized in that: change network parameter (w for each _Ij, w _Aij) the study step pitch, always consider the data set (du of a plurality of regulated quantitys _i(t _r)), the data set of a plurality of network responses

And the data set (dy of a plurality of output quantities variations _j(t _r)).

9. according to the described method of above-mentioned each claim, it is characterized in that: in a rolling equipment, during the course the flatness or the longitudinal profile of milling material (1) are regulated, wherein output quantity (y _j) by the band flatness parameter that on the width of milling material (1), distributes, form as band pressure or band longitudinal profile parameter, reach wherein regulated quantity (u _i) comprise and change the adjusting intervention amount roll the seam cross section, the local cooling of especially roller declination placements, roll bending, roll displacement and/or roll, and roll degree of eccentricity adjusting under the situation of frame (Sendzimir-Ger ü sten) at Sen Shi.

10. method according to claim 8 is characterized in that: working point parameter (b ₁) comprise milling material width, roll-force and/or roller diameter.

11. a process control equipment, it has a regulating device (7), wherein by the measured value (y of process output quantity _j) calculate the regulated quantity (u of a plurality of governor motions that process is worked _i), it is characterized in that: be provided with one and be used for study as output quantity variation (dy _i) change (du with regulated quantity _i) relation governor motion performance (w _Ij) neural network (10); And governor motion performance (w in order to be learnt _Ij) send regulating device (7) to improve regulated quantity (u _i) calculating, neural network (10) is connected with regulating device (7).

12. equipment according to claim 11 is characterized in that: neural network (10) has an input layer, by it its input regulated quantity is changed (du _i), reach an output layer, be used for the output network response

, wherein input layer and output layer by with the corresponding network parameter (w of governor motion performance to be learnt _Ij) and according to relational expression ${d\tilde{y}}_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}\·{\mathrm{du}}_i$ Be associated with each other; And being provided with a device (12) with learning algorithm, it responds according to network And output quantity changes (dy _j) between deviation (e _j) make network parameter (w _Ij) according to reducing this deviation (e _j) principle change.

13. equipment according to claim 11 is characterized in that: neural network (10) has an input layer, by it its input regulated quantity is changed (du _i) and process working point parameter (b ₁), one is used for the output network response

Output layer and an overlayer that is positioned in the middle of it; This overlayer comprises the unit (17) of predetermined number, and wherein each unit (17) represent one respectively by working point parameter (b ₁) definition hyperspace in work fulcrum (P _a) and by the working point parameter (b that inputs to it ₁) calculate a weighting coefficient O _a, this weighting coefficient is at the actual value (b by the working point parameter ₁) working point (P) and this work fulcrum (P of definition _a) between the measuring of distance; Overlayer also comprises other unit (18), therein to regulating quantitative changeization (du _i) weighting coefficient (O that provides by unit (17) respectively is provided _a) come weighting; Overlayer and output layer are by network parameter (w _Aij) and according to relational expression ${d\tilde{y}}_j=\underset{a=1}{\overset{s}{\mathrm{\Σ}}}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{aij}}\·o_a\·{\mathrm{du}}_i$ Be associated with each other, in the formula $\underset{a=1}{\overset{s}{\mathrm{\Σ}}}{o}_a\·w_{\mathrm{aij}}=w_{\mathrm{ij}}$ Be governor motion performance to be learnt; And being provided with a device (12) with learning algorithm, it responds according to network And output quantity changes (dy _j) between deviation (e _j) make network parameter (w _Aij) according to reducing this deviation (e _j) principle change.

Images