Вы находитесь на странице: 1из 6

Proceedings of the 2004 IEEE Intemational Conference on Control Applications Taipei, Taiwan, September 24,2004

? .

Generalized Minimum Variance with P c

Assignment Controller Modified

r Practical Applications

Abstract Among self-toning eontrollcrs, the explicit typt is the most commonly osed in industrial applications due to its simple strnctore; on the other hand the implicit shllctores are less attractive because of their largest shpchlr~lcomplexity. This paper presents a modification to the generalized minimum variance with pole assignment controller. The modiPr*tion pretends to simplify the origi0.1 algorithm and to make easier its implementation io practical projects, conserving most o f the

original attributes. PRINCIPAL SYMBOLS Polynomials of order n , , mb and n,corresponding to, respectively, system output, control input, and
disturbanceinput; e+=co=l.

presented and compared with the self-hming controller using pole a s s i p l e n t (STCPA) [5,6]. From this comparison a modification of the Allidina controller is also presented, incorporating the best characteristics of a STC and the GMVC. This modification simplifies the control algorithm and facilitates its implementation in a practical way when the process parameters are unknown at the controller s t . Another advantage of this modification is that the parameter selection variability is reduced, producing a more reliable controller.

A . M

Ed) Wt,
p.e,A

E
F

G H T

U(,),

CO)

k
I

Expectation operation, Generalized output function. Cos polynomials of order n , , nQ and na acting on, respectively, system output, input and reference. General polynomial acting on the reference. General polynomial acting on the perturbation General polynomial acting on the output. General polynomial acting on the input. Pole assigmnent polynomial r(l) System output, input and set point, respectively. Unconelated zero-mm random sequence. System time delay. Time measured in sample instants.

2. AUTOTUNING CONTROL WITH POLE ASSIGNMENT a i n objective of a pole assignment control is to force The m the system to achieve a set of previously-defined closed loop poles. The self-tuning cnnh-oller with pole assignment (STCPA) is briefly reviewed in the following paragraphs.
Consider the system described by the CARMA model
A y ( / )= z-'Bu(/)

+ c&r)

(1)

I. INTRODUCTION The self-tuning controllers have been studied during several decades. Some developments have been successfully adopted in industrial controllers [1,2]; being the most popular among instruments manufacturers, the explicit self-tuning controller type [3,4] which synthesize the control law using the process estimated parameters. On the other hand, the implicit selftuning controllers have not achieved the impact of the explicit controllers due to the complexity of their practical r y to achieve an optimal implementation. However the former t performance of the overall system: this is a characteristic that could be used in industrial applications in a broader way

A , R and C are polynomials in z, where the polynomial A is monic and coprime in relation to K, k represents the system delay in samples of time, while &I) is an unconelated zero mean random sequence. To achieve that the plant output y(0, follow to certain reference rp), inside a pole assignment outline (Fig.l), it is necessary to design the contml polynomials H,G and E. In Fig. 1 we can be observe that the controller structure is given by
H N ( I )= E r ( / ) G,"t)
~

(2 )

The closed loop poles (represented by the rmts of T) are assigned in the desired locations hy the appropriate selection ofHand G according to the diophantine equation [lo].
HA + : -IBG = TC

In this document, the generalized minimum variance controller (GMVC) [7,81 and the generalized m i n i " variance controller with pole assignment (GWCPA) [9] are

(3)

The self-tuning control with pole assignment (STCPA) c a n he synthesized in the following algorithm.

0-7803-8633-7/04/$20.00 02004 IEEE

1347

H U ( ~ ) + G Y ( ~~ )+ r(r)=o

(6)

and G is calculated from


PC= F A + ~ - ~ G
I

(7)

-___-___-_-__'________I

Lpg
Co*oller

where:
H=BF+QC y
E=-CR

(8)

Fig. 1 Closed Imp control s m m e

The generalized minimum variance control can he synthesized in the FolloWiog algorithm.

2.1. Algorithm: Self-tuning control with pole arsignment

Given T polynomial
A A A

Step 1 Use recursive least squares to estimate A , B , C from (1). Step 2 Synthesize the controller polynomials from
A A A A A

AH+ :-I B G = CT
A 76)" E=-C

Step 3 Generate the control output u t ) from


A A A

H u ( t ) = Rr(l)-Gy(Ij

Step 4 Wait for sample time interval to elapse and then return
to Step 1.

*
Fig. 2 Generalized output

This controller has several advantages; the most evident is the simplicity. This controller also offers robustness [6] and the ability to control n o n - m i n i " phase systems. Because of its capacities. the STCPA surpasses the elementag minimum variance controllers.

3.1 Algorithm: GMVC


Given P. Q and R, Step 1 Form the generalized output
((I)

3. GENERALIZED MINIMUMVARLANCE CONTROL In this section, the GMVC [7,8] is briefly reviewed as
follows. Let us defme a generalized output or pseudo output as ((t+@ =Py(t+k)+Qu(?)-Rr(t)
(4)

((1, =Py(1, +Qu(t-k)-Rr(t-k)


step 2 Estimate

,+,E

and

i from
-k ) + i ( t )

where P, Q, R (denominated cost polynomials) are used to modify the control performance (Fig. 2). The cost function'to be minimized is the mathematical expectation of& + k ) . J=E,
(&(tCv)

Kri-

&u(t-k)+&y(t-k)+kr(t

Step 3 Apply control up) using


ku(t)= -&(Ij-

&(l)

(5)

The control polynomials are computed according to the control law

Step 4 Wait for the sample time interval to elapse and retum to Step I.

1348

Convergence analysis related to this type of controllers can he found in [4, 141. The GMV algorithm has some disadvantages: 1. The algorithm is steady state one. the system performance depends directly on the parameters convergence [4].
2. A situation that complicates the convergence is the controller predictive character, which depends on a subdimensioned estimation of the control parameters from a timedisplaced #).
4. GENERALIZED MlNlMUM VARIANCE CONTROL

3. If it was necessruy to estimate the process when the plant is unknown, then it is not possible that P,Q and R converge, because the generalized output function cannot be formed appropriately on-line.

It can be deduced that the AUidina algorithm is only feasible when the process is previously h o w . In recent years numerous papers have been published about GMV controllers, some examples are [11-14]; However most of these developments consider that the process is known before beginning the control, this situation avoids implementing GMCVPA as a commercial controller.
5 . GENERALIZED MIMlMUM VARIANCE CONTROL WITH POLE ASSIGNMENT MODIFIED FOR PRACTICAL APPLICATIONS. The first practical application possibility for the GMVCPA is to introduce between the step 1 and the step 2 a process estimation step (see the section 4.1). But this modification is inefficient because requires double estimate. However a bigger problem exists, if the plant parameters are unknown, an appropriate generalized output function canuot be designed, and the convergence (in the case of existing) only would he a casual event. Such problems in the practical implementation motivate the modification of the existent algorithm; therefore we propose modifications that " i z e the caused inconveniences by insufficient previous information. To make n the modification (Generalized possible the convergence i minimum variance control with pole assignment modified for practical applications) is assumed that C(t) it is equivalent to C(t-1) (canceling the predictive mode on the controller). and we also assume that the coefficients estimation of the general polynomials tend to the m e ones (with this supposition the estimate of the general polynomials can be omitted). Instead of proposing P and Q arbitrarily, these are calculated on line using the equation
A A

WITH POLE ASSIGNMENT. Allidina, and Hughes [9] presented a modification, which consisted on "self-tune" P and Q in such a way that a desired closed loop poles is achieved, represented by a Tpolynomial.
4. I Algorithm: GMVCPA

Given P, Q, R initials and T, Step 1 Form the generalized output

&t)=Pv(i)+Qu(t-k)-Rr(t-k)
Step 2 Estimate H, G and E from

Step 3 Apply control U ( ! , using


.ri.(1)=

-&(1)-kr(1)

Step 4 Recompute P and Q from:


PH - q - k ~ = bm ~

BP+AQ =T

(9)

Step 5 Wait for the sample time interval to elapse and return to Step 1. The term h 8 acts as a polynomial observer where ho is the fust element of H . This structure pretends to eliminate the off line adjustment of P and Q; however this algorithm presents the following disadvantages:
1. P and Q convergence is subordinated to the process parameters knowledge, because this iuformation is used to design the generalized output function. A well designed function forces the parameter convergence.

maintaining the relationship between the optimization and the pole assignment. Next the modified algorithm is listed
5. I Algorithm: Modified GMVCPA
A A A

Step 1 Using recursive least squares estimate Step 2 Calcnlate P and Q from
f i n

A,B,C

RP+AQ=T Step 3 With P and Q calculate directly the control parameters

.G,6 and B from


PC = F A + ?G
H = BF+QC

2. There is not convergence guarantee with arbitrary initialization of P and Q [4].

E = -CR

1349

Step 4 Apply control u(t) using

whose discrete equivalent. using a sampling period of 0.1 seconds and considering an perturbation, is
y ( : ) = 1,81?y(:)-

f;U(/)=

-1.y ( / )- 2r(1)

0.819;~'y(.-)+ O.O0468;~*y(;)

+0.00438~-'+0.05e(.-)

Step 5 Wait for the sample time interval to elapse and return to Step 1.

In Table I, we present a comparison of the algorithms reviewed.

For this example. we select the following pole assignment polynomial: T(z)=:'-1.879~+0.8884 The initial conditions used for the simulation are: forgetting factor h=0.98, initial estimates t3=(0.5,0.5,0.5,0.5,0.5) and the initial covariance matrix was selected as PIOOI, where I is the identity matrix whose size is Q x Q. (The initial conditions described in this section will be effective in all the tests presented in this document).

6. CONTROLLER RESULTS The GMVCPA modified was tested with a set of linear processes defined in the literature [2]. One of these processes is

TABLE 1. ALGORITHMS COMPARlSON STCPA


.TOinitializethe recursive

esIimati0". .The process parameters are


Imkno\w.

GMVC - To initialue the recursive eninmion. . Select values P.Q and R to fom the
generalized o q d #(,)= P.d,)+Qri(t - k ) - R r [ t - k )

GMVCPA - To initialire the recursive eaimation. - Select initial valws P.Q and R to fom

modified GMVCPA
-To initialize the recursive

ealmation The prOEeSS parmeters are the generalized o q u t ~ [ ~ ) = P y [ l ) + Q ~ [ l - k ) - ~ ( r - k ) &om.

Using reclnnive least squares


A A A

loealmate

A,B,C

Enlmate H , G , E from # ( f ) herethe generalized o q l n is

Estimate H , G , E from ((f) generalized o q u is


# ( I ) = .&,(I

wherethe

using recursive least squares to


n ~ l

estimate A,B,C

#(I)=;U({

k)+
~

- k)+
For a pole assignment T,solve the

6yii - k ) + i r ( t

k ) + ;(I)

;."(I

- k ) + &(I- k ) + ;(I)

equation
Calculate H and G solvlng
A H + Z-' B G and E: hom
=

cT

BP+AQ=T With the estimated plynomials A, B, C.P,Q. and R to solve

PC = FA + :-'G
H =m+QC E = -CR
to obtain H, G and E

T(1)li=-C

&)
TO apply U ( l ) from
A h

'

To apply U ( ! ) h o m

Solve P H - ; ? CQ = h,FT to obtan P and Q effective for the ned s m p h


penod

1350

In Fig. 3, the controller and the evolution of the estimated parameters are presented.

certain pole assignment (9), the success of the suboptimal controller will depend on the perfect emulation of the dynamic behavior represented by T. Therefore if we choose T equal to the open loop process denominator, a successful pole assignment would cause a wutrol output with the same shape as the reference input. To illustrate this situation, we use the plant described in section 6 and T i s assigned to the pole placement of the open loop process denominator. In Fig. 4 we present the reference and the control input; note that U is equivalent to the r shape

1.Y

*o

80

IO

,m

1 m

1 4

?el

1 M

time (S)

4c

rm
(S)

120

$4

180

*lo

tune

Fig 3 Sundahoo of L h e mothtied GMVCPA

7. COMPUTATTONS EFFICIENCY When considering applying a controller in a practical application. computing efficiency becomes an essential feature of an algorithm. In this paper, we compare the above reviewed controllers regarding execution time and memory requirements. All tests were performed in a PC Pentium 4, Matlab was used as programming platform and the measurements were done using Matlab utilities. Table 2 summarizes these results

A Execution dme (seconds).

B Static memoly d g the execution (byies). C Memory occupied by the code @yies). D Propadonal increase 10 the d e s 1 execvfiontune. E: hoportiooal increase to the Smallest ststic memory Jlocatim.

From Table 2, we can state that STCPA is the fastest controller, followed by the modified GMVCPA; however, the latter performs much faster that GMVC and GMVCPA. Regarding static memory space, STCPA and modified GMVCPA need almost the same space: whereas GMVC and GMVCPA require a lot more. In terms of code length, STCPA generates the smallest code, fallowed closely by the modified GMVCPA.
8. SUBOPTDMALITY When we include a cost function in the design of adaptive controller. the control law is optimized when we also have a stochastic characteristic, such control systems must be considered suboptimal, If the cost function is defmed by a

Fig. 5 Generalized outputs and performance indicators.

To show the optimization process developed by the modified GMVCPA algorithm, the generalized output function (Fig. 5a) was reconstructed (because in the modified controller is implicit). The next equation

shows the difference that exists between the Ki,l and Theoretically, the cost function should be minimized to achieve 0. without considering the pmrbation (Fig. 5b). Findly, if the pole assignment was successful, for this particular case the relationship

eo).

1351

11. ACKNOWLEDGEMENTS

must have a tendency towards 0. In this test, the average m o r for a sampling period was 0.0181, which is equivalent to a percentual error of 0.181% with FIO, (Fig. Sc). The cost function in the modified GMVCPA algorithm is determined by (9). If an algebraic manipulation is carried out, the control law structure of the modified GMVPA cdli be equivalent to the control law used by the STCPA. However, the modified algorithm capacities are not restricted to the development of pole assignment, because there is the possibility of modifying the cost function. 9. PRACTICAL IMPLEMENTATION To cany out tests, a laboratory apparatus (CEl10 Servo Trainer, Tecquipment) was selected, which is a D.C. motor affected by inertial dynamics. The process is perturbed with a load change in the second 77; in the Fig. 6, we can see how the process response overcomes the load change. In this figure it can also be appreciated that the generalized output function amplitude. is notably small (standard deviation of 3 . 0 ~ 1 0 and ~ - 3 . 0 ~ 1 0 mean), .~ which shows the capacity of optimization of the algorithm in practice.

The authors wish to thank to CENIDET and Tecnologico de Monterrey Campus Cindad de M6xico for the support provided to develop this project.
1 1. REFERENCES

IO. CONCLUSIONS
We can say that the modification proposed for the GMVCPA gives verygood results With the cases studied so far, since it considerably reduces the number of parameters that need to he initialized, its implementation complexity is comparable with explicit control structures, and fmally, it allows the controller to cany out optimization tasks in a reliable way.

[I] Hagglund T., Astrom, K.J. Industrial adaptive controllers based on ffeauencv remonse techniques, Automatica, 2 7 , 1 9 9 1 , 599-609. ~ . 121 Lundh. M.. AstrBm. K.J. Automatic Initialization of a ._ robust self tuning controller, Automatica, 30, 1994, pp1640-1662. [3] Astrom, K.J., Wittenmark B. Adaptive control (2ded.). Addisan-Wesley 1995. [4] Wellstead, P.E., Zarrop M.B. Self-tuning sysfsms. John Wiley & Sous, England 1991. [SI Wellstead P.E., Edmunds J.M., Prager D., % d e r P. Self-tuning polelzero assignment regulators, Int. J. Control, 30, 1979, pp 1-26. [6] Wellstead P.E., Prager D., Zanker P. Pole assignment self-tuning regulator, IEE Proc., 126, 1979, pp 781-787. [7] Clarke, D.W., Gawthrop B.A. Self-tuning controller, IEE Proc., 122,1975, pp 929-934. [SI Clarke, D.W., Gawtluop B.A. Self-tuning control, IEE Proc., 126, 1979, pp 633-640. [9] Allidma, A.Y., Hughes F.M. Generalized self-tuning controller with pole assignment_IEE Proc., 127, 1980, pp 13-18. [IO] KnEera V. Diophantine equations in control-a survey Automatica. 29, 1993, pp1361-1375. [ l l ] lanaka H., Fujikawa, H.. Yamada S. A design method of generalized minimum variance for tracking reference with ramp input, Proceedings of the International Conference on Industrial Electronics, Coiitrol and Instrumentation, 1993, pp. 2212-2217. [I21 Yamamoto T., Fujii K., Kaneda M. Self-Tuning Temperature Control of a Polymerizing Reactor, Proceeding of the 1998 IEEE International Conference on Control Applications, Trieste, 1998, pp. 1110-1114. [I31 Yaacoh S., Mohamed F.A. On-line self-tuning controller for induction motor based on generalized minimum variance method, S C E _1998, pp. 875-878. [I41 Masayoshi D._ Yasuchika Mori, A study on Robust Asymptotic Tracking Property for Generalized Minimum Variance Control, Proceedings of the Americm Control Conference, Anchorage, 2002, pp. 1472-1477

1352

Вам также может понравиться