Sei sulla pagina 1di 14

Control Eng. Practice, Vol. 4, No. 5 pp.

685-698, 1996
Copyright © 1996 Elsevier Science Ltd
Pergamon Printed in Great Britain. All fights reserved
0967-0661/96 $15.00 + 0.00
PII:S0967-0661(96)00052-4

COMBINED IDENTIFICATION AND CONTROL: ANOTHER WAY

L. Keviczky
Computer and Automation Institute, Hungarian Academy of Sciences, H-1111 Budapest, Kende u. 13-17, Hungary
(hl Okev@huella.bitnet)

(ReceivedJuly 1995; infinal form February 1996)

Abstract. A new generic optimal controller structure and regulator design method are
introduced, avoiding the solution of polynomial equations. The model sensitivity properties
of some combined identification and control schemes are investigated. It is shown that the
new structure is superior to the others. An applicable strategy for iterative control
refinement based on the generic scheme is presented and illustrated by simulation
examples. The adaptive version of the control refinement strategy is also demonstrated.

Key words: Pole-placement controller, internal model principle, optimal control, computer
controlled system, model sensitivity, combined identification and control, adaptive control

1. INTRODUCTION This is mostly the case if the control is discrete-time


and based on sampled linear dynamic systems. In a
The need to design high-performance control general case the system S = S÷S is factorable for
systems has not lessened in importance, in spite of inverse stable S÷ and inverse unstable S_
the thousands of methods and algorithms published components. Because S_ cannot be eliminated by a
in recent decades. The huge number of papers
simple cancellation mechanism it is sometimes called
indicates that no unique or best method has been
the invariant system component. A heuristic but
found. The solution depends on the model, criterion,
uncertainties, disturbances, constraints, etc. widely applicable solution is to choose X = S~.I ,
(sometimes even on the designer's taste). when the inverse stable part is canceled; however the
invariant inverse unstable factor is untouched.
Several controller design schemes are based on this
method, even though there is no optimality
connected. Unfortunately, the remaining invariant
factor can sometimes cause intolerable dynamics, so
its effect must be attenuated. This can be done, if a
proper criterion is used for this purpose, e.g.,

Fig. 1. The simple problem of control


(1)
Many experts believe that the essence of all control = arg m~n{E{[(1- XS+S_)n(k)]2}}
problems can be led back to the simple problem
shown in Fig. 1, i.e., how to choose a serial
compensator transfer function X to S ensuring a which is the formulation of the classical Wiener
design of optimal stochastic systems, providing the
unity dynamic transfer. The trivial solution X = S -1
is not always applicable because S is not invertible. minimum mean-square error (H2) problem by

This paper is a shortened version of the author's invited x = g_-~s~~ (2)


plenary paper presented in the ACASP'95 IFAC
Conference (see the References).

685
686 L. Keviczky

where ~ is obtained by reflecting the zeros (they are They can also be interpreted as predictors for the
reference and output disturbance signals. In this case,
unstable!) of S_ through the unit circle and
e.g., ff is the estimated (or predicted) disturbance.
providing (Here an ideal case was assumed when the true
process S is known to ease the understanding of the
S_-' (1) S_ (1) = 1. (3) basic schemes. This assumption is good to explain
the operation of the system; however, only a process
Note that the solution depends on the applied input model M is available in most practical cases, as will
excitation n(k). Here n(k) is assumed as a white be discussed later.)
noise sequence.
pARTIALLy I I PROCESS Iw
INVERSEMODELI I /

' I-'11 II Ill /


~ - - IL . . . . . .

" I ~ INTERNAL
MODEL
CONTROLLER
Fig. 2. A classical pole-placement controller

A characteristic approach to optimal controller Fig. 3. Modified internal model principle for inverse
schemes is called the pole-placement technique unstable processes
(/~strtm and Wittenmark, 1984; Landau, 1990)
Note that the whole system should be taken into
targeted at providing prescribed transient properties
consideration in the internal model, because a
for the servo and disturbance-rejection paradigm of
realizability problem arises when using only the
closed-loop controller design. The standard
inverted model. If the invariant system component is
technique representing a two-degrees-of-freedom,
so-called ~9~- Sa - J controller assumes basically to be attenuated, then ~VS~_l must be used instead of
the structure shown in Fig. 2, where y,, u, y and w S~_l in the partially inverse model, according to the
are the reference, input, output and disturbance optimality of Eq. (1). The advantage of this optimal
signals, respectively. Here, discrete-time control scheme is that the principal operation of the
representations are considered for computer regulator is very easy to follow, and the
controlled systems. The argument k of variables computations of the regulator polynomials are also
means the integer value discrete time (integer easy and obvious. The disadvantage of this scheme is
multiple of the sampling period) and z -n means the that the identification (parameter estimation) method
backward shift operator (z -ly(k) = y(k - 1)). indicated by the internal model in the closed loop
gives rise to several difficulties.
The . g L S a a n d J - are polynomials, therefore the Both the above approaches have the general
advantage of this scheme is that the implementation problems of the known optimal controller structures,
of a so-called "direct adaptive regulator method" is that the identification and control errors are different.
very easy, because it is easy to construct a predictor Therefore these schemes are not the best ones for
equation which is linear in the parameters of these developing and analyzing simultaneous identification
polynomials. The disadvantage of this scheme is that and (adaptive) control algorithms. In this paper a
it hides the internal operation of an optimal system, new generic structure (Keviczky and B~inyfisz, 1994)
and special considerations are necessary in a is discussed, which allows a very simple procedure
recursive parameter estimation algorithm because the to design optimal control systems when the
~ , Saand J - are redundant, having more parameters identification and control errors are identical. Their
than are minimally necessary. Furthermore, the relationships to the previously discussed schemes are
solution of a Diophantine equation is necessary to also presented.
obtain the regulator polynomials.
r i
Another well-known classical scheme of optimal / ] i Iw
control systems is called the internal model
principle, whose less well-known form modified for
inverse unstable factors is shown in Fig. 3, if only the
-t,,s-'pLLp l odls
lyrr-'-71 r--7 ~ .~, lu [ - - - . - ~ . ~
T y

inverse stable factors are canceled. The name


originates from the system model applied in the
controller. Here P~ and Pw are the desired overall
tracking and disturbance rejection transfer functions Fig. 4. The new control system structure
(or reference models) for the design requirements.
Combined Identification and Control 687

2. A NEW CONTROLLER STRUCTURE Sa=ArB(Aw -Bw) (12)


Another, new structure (shown in Fig. 4) is
introduced in this section for the design of optimal
~=AA~Bw. (13)
controllers.
(Observe the common factors that make the
parametrization for a direct self-tuning technique
Here S is the system, R is the regulator, fir is a redundant and the recursive estimation difficult.)
precompensator transfer functions. The system
output If the above precompensator and regulator are used,
the controller shown in Fig. 4 can be further
1 simplified according to Fig. 5. The advantage of this
y(k) = firSYr(k ) + 1 + R S w(k) = Yt(k) + ya(k) (4) approach is that the three transfer functions of the
controller can be computed directly, without solving
is very special, because in spite of the closed loop the the usual polynomial equation. A further and very
important advantage is that the identification and
tracking behavior (yt(k) = firSYr(k))is independent control errors are identical, so this scheme is
of the regulator R. This structure in practice opens especially good for combined identification and
the loop for the command signal and the selected control problems. (The very limited usability for
feedforward compensator (or observer) inverse stable systems will be eliminated in the next
section.)

fir = Pr S -~ (5) F I
i F j '~ J w
provides a desired iyf-Y"l V'E"I .~ I--"-1 l u l - ' - ' l "i y

Yt(k) = Pr Yr(k) (6) J N I


i I
i
i
tracking (servo) transfer function by Pr" Observe that I CONTROLLER I
by selecting a regulator
Fig. 5. Equivalent form of the new controller
R = Pw S -1 = C S -1 (7)
1-Pw
3. A GENERIC SCHEME FOR OPTIMAL POLE-
PLACEMENT CONTROLLERS
a desired regulating (or disturbance rejection)
behavior
Because the above method is based on full pole and
zero cancellation the extended applicability can be
yd(k)=(1-Pw)W(k) (8) reached by partial cancellation of only those system
(model) components, which are inverse stable.
can be reached by Pw. The Pr and Pw transfer Assume that
functions contain the desired poles to be placed, so
they can be called as reference models. Note that S=S+S_=S+'ff_z -a, (14)
arbitrary zeros can also be placed; however, the
calculation of the precompensator fir and regulator is stable and S+ means the inverse stable factor. Here
R requires the inverse of the process S. (Here, also, S_ = ~ z -a is the non-invertible part where the
the ideal case was assumed, when the true process S
is known. If Pw is the best LS predictor of w, then discrete time delay z -a is also a factor whose inverse
the regulator obtained is the minimum variance (MV) z a is not realizable. In case of a partial cancellation
regulator.) It is easy to prove that this scheme can be the precompensator
rearranged to the scheme in Fig. 2 by straightforward
block manipulations. Assuming that fir =P~S~ ] (15)

B should be used instead of (5). This results in


S = -- (9)
A
Yt(k) = PrS+IS+Lz -d Yr(k) prLZ -d Yr(k).(16)
=

Pr =Br and Pw =Bw. (10)


Ar Aw Selecting the regulator as

the classical pole-placement regulator polynomials


can also be easily computed according to the e = Pw S-~ I = ca-+ 1 (17)
previous formulas, and are given by 1-ewLJ
~z'= AAwBr (11) one can obtain the regulatory transfer as
688 L. Keviczky

Yd(k)=( 1- PwS'- z-a)w(k) • (18) The final optimal controller is already very general,
because it covers the most critical processes where
the design is not trivial. In practice, the model M
Eqs (16) and (18) show that in the case of partial should be used rather than the system S in the
cancellation, which is the general case, when there internal model, so the design procedure should apply
are inverse unstable system factors (e.g.,
nonminimum phase systems) or time delay, the ideal
servo (Pr) and disturbance rejection (1-Pw)
transfer functions cannot be reached, but only their (21)
modified (Pr~Z -d) and (1-Pw'S_z -d) forms. Note -- = - ~ Z

that the modifications do not depend on us, instead


they depend on the system itself only. Therefore S_ instead of (14). Here 1t4+,M_ and z -d" are the
is sometimes called the "invariant process factor"
inverse stable, inverse unstable factors and the delay
(mostly zeros). (Because reasonably Pr(1) = 1 and time of the model, respectively. So the optimal
Pw(1) = 1 are selected, it is also reasonable to choose controller will change according to Fig. 8. This
S_(1)=I in the factorization Eq. (14).) Figure 6 controller is very easy to implement in a computer
shows the transfer functions in the new controller controlled system and it retains the advantage of the
original idea of the new controller structure, i.e., it
scheme, which can even be called a generic optimal
does not require the solution of a polynomial
pole-placement controller in this case. This scheme
(Diophantine) equation to obtain the controller
can be further simplified according to Fig. 7. It is
transfer functions or polynomials. Instead, (11), (13)
easy to show that Fig. 6 can be rearranged to the
schemes in either Fig. 2 or Fig. 3 by straightforward and (20) should be applied now, where B+B_z-din
block manipulations. So the classical oole-olacement and ,~ are the numerator and denominator
reeulator schenl¢ and the modified internal model polynomials of the model M instead of the system
orinciole scheme are identical to each other and t0 S, as Fig. 9 shows.
the new generic scheme if the appropriate transfer r I
functions are selected. I [ ] I Iw
[- -I
j jw

[ CONTROLLER I
Fig. 8. The generic-model-based pole-placement
I CONTROLLER I controller
[ q
Fig. 6. The generic pole-placement controller i [ | t Iw
[ q
I I I ~ Iw I iNv~lr
~--" I P~rl~u_v
_.l s /
'yrD~.9 ~ lu **y LCONTROLLER j
I .v~.r "1 P~YI ~ /
Fig. 9. The polynomial design of the generic pole-
placement controller
I CONTROLLER I
It has already been mentioned that the uncancelable
Fig. 7. The equivalent form of the generic controller factors of the system will be factors of the modified
reference models, so they are invariant to any control
Assuming that instead of (9) the system transfer strategy. However, their influence on the original
function corresponding to (14) is given by reference models can be minimized, if a criterion is
used for this purpose. Using the Wiener design
S = B+B z_ d (19) principle shown in the introduction the precompen-
A sator

only the polynomial Sa will be changed /St = Pr ~--1S+l (22)

-" ) (20) and the regulator

and the computation of ~" and .~ remains R= Pw S--~SZ ] = CS~.1 (23)


according to (11) and (13).
1 - Pw ~-] L z -~
Combined Identification and Control 689
must be used and the overall system equation the a priori information on the model and ignores the
fact that the characteristics of the plant could be
learnt even when it is being controlled. Therefore,
y(k) = PrL-lS_z-dyr(k)+(1- Pw~_-lS_z-d]w(k) classical robust control approaches mostly result in a
(24) conservative design in terms of performance.
is obtained. Figure 10 shows the practical realization
of this strategy if the model of the system is used. 4.1 Open-loop identification and closed-loop control
The simplest strategy to combine modeling and
Note that the application of the compensator ~-l on control, if the identification to obtain an optimal
S_ is optimal for a given excitation only, in this case model M*, selected from a model class ~¢/, is
for white noise (or approximately for wide performed in an open-loop experiment by
bandwidth) disturbance. Therefore the optimal pole- minimizing an identification criterion Qio(eo)
placement controller shown in Fig. 9 is suggested for function of the open-loop identification error e o , i.e.
most practical applications.

A~ED I M* = arg min Qio(eo) = arg min


M~..R M~,.~
Qio(M,x, A) =
' :" 1 = Mo(./[,x, S) (25)

A~TF..t~ ] e J S ! where x = {x(k);k= 1..... N} is the applied input


I FN~fOR l l~'~'k4t ~" I j
excitation series and
r CONTROLLER I
Fig. lO. Attenuating the invariant factors by Wiener A = S- M (26)
design
is the additive model uncertainty between the system
S and model M (see Fig. 11). The optimal regulator
4. COMBINED IDENTIFICATION AND R*, selected from a regulator class ~n~, is obtained
CONTROL SCHEMES
by minimizing a control criterion Qc(ec) function of
Since the start, the key paradigm in designing control the closed-loop control error ec (see Fig. 11), i.e.,
systems is how to handle uncertainties associated
with the plant. One of the main techniques is R = arg min Qe(ec) = arg min Qc(R, S) = R*(~, S).
adaptive control, intended to learn parameter and RE,R R¢,~
disturbance uncertainties in varying circumstances. (27)
Another important approach is to implement ~"OPEN-LOOPIDENTIFICATION '°
robustness features in simultaneous identification I Yo
and control procedures.
'
4x 1
(~)
In the traditional approach to the analysis and design J
of an adaptive control system the unknown plant is
represented by a model, which is almost known
except for a few parameters, which are assumed to be
time-varying. Having the estimated parameters, the REGULATOR
DESIGN
I
controller is updated according to the certainty
equivalence principle. The unstructured uncertainties
are mostly ignored in these cases; therefore, these
adaptive regulators are not robust. To tell the truth it
is not at all easy to consider these uncertainties in the
classical adaptive systems, and to guarantee proper
transients during the learning adaptation phase or Fig. 11. Combination of open-loop identification and
abrupt parameter changes. closed-loop control

In a classical robust control approach, the regulator is (Note that the open-loop input excitation x is
designed on the basis of a nominal model of the different from the closed-loop input u acting on the
plant, with the associated parametric and process.) Because only the model of the system is
unstructured model uncertainties explicitly taken into known in a practical case, the most widely applicable
account. (Unfortunately mostly analytical forms are strategy is to substitute S by M in (27). This
required, which are very rarely available in practical strategy is called the certainty equivalenceprinciple
applications, except for a few special real cases or and is realized by
the examples of "god-given" a priori information.)
Here stability robustness is guaranteed and R = argmin ~:(~c) = argmin Q_.c(R,M)=/~*(,~g,M).
performance robustness is sometimes achieved. The R~JI~ R~.R
weakness of this approach is that it considers only (28)
In this case the optimal regulator does not reach the
690 L. Keviczky

theoretical optimum, because the model M is used It is possible to form different combined schemes,
instead of the system S. However, it is possible to depending on the structure of the optimal control and
form an iterative control refinement procedure, the combination of the sequential identification and
improving the model and regulator step by step: control steps. In this sequential procedure it is a
general observation that:
1. Identify the model using the modeling step 1. The human first learns to control over a limited
bandwidth, and learning extends the bandwidth
(29) over which an accurate model of the plant is
= Mo(* , S ),
known.
2. The human first implements a low-gain controller,
2. Calculate the optimal regulator and learning allows the loop to be tightened.

= k'(m, (30) On the basis of these observations, an adaptive


robust control philosophy, the windsurfer approach,
3. Determine an optimal input excitation for the was proposed by Anderson and Kosut (1991). They
open-loop identification use a parallel closed-loop optimal controller scheme,
which is very widely used in the analysis of control-
relevant identification and iterative control-
refinement procedures. Note that the classical so-
called '!direct adaptive control" algorithms generally
Here ~ is the (mostly amplitude or energy) use a somewhat different scheme, which is called a
constrained input signal domain. This step is parallel in-loop optimal controller scheme. These
sometimes called "optimal input design" and the schemes and the new generic optimal controller
operation is denoted by -~(...). scheme will both be analyzed below. For the sake of
simplicity, this comparison will be discussed here for
. Once M i and R i have been found one can inverse stable processes first.
continue to increase the closed-loop bandwidth by
repeating the procedure. The iterative process is The generic optimal controller scheme. This scheme
continued from step 1, while a stop condition is (Keviczky and B~iny~isz, 1994) was shown in
not fulfilled (until the ultimate control objective is Section 2.
achieved, or it is terminated because it has
reached some vital constraints). The structure of the optimal controller gives a special
insight to explain the operation of a feedback loop
For comparison, it is interesting to see that in the for the servo and disturbance rejection paradigm. It is
above case the identification and control errors are easy to see the role of the knowledge of the model of
the system and the role and appearance of the factors
H A of the system that are invariant to any control
(32) strategy. A slightly modified form of the generic
scheme of Fig. 4 will be used here, as Fig. 12 shows.
and
r I
1 1 I I ee )1 }~
ec = Yr + W ; Yr = PrYr (33)
1 + RS 1 + RS

where P~ is the reference model, Yr = PrYr is the


desired process output (model output or predicted Leo Nm°tL" . . . . . . . . ,
reference signal) and w is the output disturbance.
Fig. 12. The generic optimal controller scheme
4.2 Closed-loop identification and closed-loop
control This structure has further advantages in handling
system uncertainties; a new canonical sensitivity
The formal description of the procedure for scheme can also be obtained, and opens up a new
combining modeling and control if the identification way to analyze combined modeling and control
is performed in a closed-loop experiment is very issues. Figure 12 is the long-sought ideal scheme for
similar to the previous section; however, the the combined identification and control problem,
identification criterion Qic(ec) is now a function of because in this case the control and modelinQ errors
the closed-loop identification error e c , i.e. are identical:

M* = arg min Qic(Ee) = m W --


Me..~t
(34) el = el 1 + RS Yr 1 + RS
=arg min Qic(M,R, Yr, A)=Mc(.~,Yr, S) (35)
M~,.~

and the regulator is again designed by equation (28).


Combined Identification and Control 691

So the new scheme also provides an excellent The parallel in-loop optimal controller scheme. This
possibility for studying robust identification for scheme is a special version of the parallel closed-
control. The uncertainty scheme of the generic loop optimal controller scheme, and can be seen in
optimal controller structure is shown in Fig. 13. Fig. 16. While the previous scheme concentrates in
practice on how to provide the equivalence of the

lw two (real and model-based) closed loops and the


modeling error is an open-loop error, this scheme
conversely requires identification in the real closed
loop to minimize the modeling error in the closed
loop as well. This scheme corresponds to the
classical self-tuning regulator scheme.

w
Fig. 13. Uncertainty scheme of the generic optimal
controller structure
.t t....a I t....a ]£ ~
The parallel closed-loop optimal controller scheme.
L_ i
The most widely used optimal controller structure
discussing control-relevant identification or
simultaneous identification and control issues is a
parallel closed-loop optimal controller scheme. This Fig. 16. The parallel in-loop optimal controller
scheme is shown in Fig. 14. scheme

In this case the identification and control errors are


also different. The identification error

RS ; a ( Ra ]w
¢3 = ~ ' - ~ ) y r + 1- I+RSJ =
(39)
= H3 "~" Yr + E3 (A)w

Fig. 14. The parallel closed-loop optimal controller


scheme and the control error

In this case the identification and control errors are 1 1 1 1


different. The identification error is e3 = I + RS PrYr + I + RS w= I + R s Y r + I+ RS w
(40)
RM ( A ~. 1 The uncertainty scheme of the parallel in-loop
e2 = (1 + R ~ I + RM) ~--MJ yr + 1+ RS
W=
optimal controller structure is shown in Fig. 17.
(36)
= H2(--~)Yr + E2W

and the different control errors are


Yr Yr ~ ~ 2
1 1
e2 = ~ Yr + 1 + RS w (37)

Fig. 17. Uncertainty scheme of the parallel in-loop


1 1
&2 P~Yr 1+ RM yr" (38) optimal controller structure
1 + RM
It is interesting to note that in this structure the input
The uncertainty scheme of the parallel closed-loop signal u, used in the identification, depends on the
optimal controller structure is shown in Fig. 15. output noise w, too, which makes the input
correlated (caused by the so-called "circulating
noise") and the parameter estimation difficult.
Observe that E(A)la~ 0 =1 in (39), which is a
special classical explanation of the possible
convergence of the identification: if the algorithm
converges (i.e., A - > 0 ) , then the additive
disturbance becomes independent.

The combination of the generic and the parallel


Fig. 15. Uncertainty scheme of the parallel closed- closed-loop optimal controller scheme. It is possible
loop optimal controller structure to form several combinations of the above controller
692 L. Keviczky

schemes. Here only the special combination of the The greater this filtering weight, the smaller the
generic and the parallel closed-loop optimal modeling error is. Observe that IH2( 0)I has a
controller scheme will be shown because of its
maximum at o)c , i.e., when II1 = [R ~ = 1, if M --->S.
interesting properties. This combined structure is
shown in Fig. 18. This scheme is easy to understand;
however, it is possible to simplify it further, as Table ~, 1 Identification error weiuhtinu functions
Fig. 19 shows.
H0 Hi ~ /43
1 RS RS
I"--1 e ' = e ~ l " " l + ] , r"-'~ ,lw,J S
1+ RS (1 + RS) 2 1+ RS

-2'00
........ .-H°
.. I 7°0i ..........
...... o
........... t
. z .................. 1o
10-/ lO-i 10o 10 i 10-2 lO-i 10o 10 i

, . Ix~l . . . . I~1
Fig. 18. The combination of the generic and the
parallel closed-loop optimal controller
scheme -l°
-20 i . t .//'-"]--'"-
....... .I.... . . . . . . ¢o .............. . .... ,,>

10-2 10-1 o~c 100 101 10-2 10-1 10o 101

Fig. 20. The frequency shapes of error weighting


functions (see Ex. 7.3)

Y :C2 The weighting function H l emphasizes the high


frequencies, and H 0 and H 3 the low ones, while H 2
concentrates on the medium frequencies around (oc .
This last feature implies the widely used analysis
techniques based on this optimal control structure.
Fig. 19. The simplified form of the combination of Note that H 0, H 1 and H a correspond to the
the generic and the parallel closed-loop frequency characteristics of the process, the
optimal controller scheme sensitivity function ( SF), and the complementary
sensitivity function (CSF), respectively. H 2
It is easy to see that in this case the parallel closed- corresponds to the product of the CSF and SF, and
loop optimal controller scheme becomes equivalent can be calculated as
to the generic scheme with the minor difference that
e 2 = - e l, i.e., ~2 = - e l . This means that by the Y RS
above combination the parallel closed-loop optimal H2IM=S = "t(r) = ~ = (1
- +~ RS) -"

controller scheme will also have the very important (42)


behavior that the identification and control errors are =~I[RS=I/Y--~O , ~--*0
identical [ RS = Y --->O , ¢o--->oo

Observe that this product represents a specially


e2 = HlI-~)~r + Elw =-el = - e 1. (41)
c o m p u t e d m e d i u m - f r e q u e n c y r a n g e (MFR)
weighting. It is easy to see that ~/(Y) has a maximum
at II1 = 1, i.e., at the crossover frequency ¢oc .
5. COMPARISON OF THE DIFFERENT One should know that there is a heuristic - not
SCHEMES exactly proved - inverse proportional relationship
between the prediction error weighting of the
The simplest way to compare the different parameter-estimation procedure and the obtained
uncertainty schemes is to plot the absolute values of model error frequency dependence. So the weighting
the identification (prediction) error weighting shapes mean that H o and H a are not good for
functions IHe(jo~)i (e=0,1,2,3) - s u m m a r i z e d in iterative control refinement to increase the
Table 5.1 - as a function of o~, if M---> S. It is bandwidth, because they provide a larger model error
interesting to see how different they are (see Fig. 20), in the desired higher-frequency domain. H 2 seems to
where, plotting H o and H 2, M = S is assumed. Here be good for iterative control refinement; however, its
¢o denotes the natural frequency. The external special shape definitely needs a cautious, really
excitation Yr acts through the relative model error windsurfer, approach, because it gives the smallest
(given by A / M or A / S ) and these weighting filters. error around o~c . The weighting function H 1 is the
Combined Identification and Control 693

best, because by amplifying the high-frequency use a regular identification algorithm based on the
uncertainties in the identification error, a more auxiliary variable fi and the measured controlled
accurate model is expected in the targeted high- variable y, as Fig. 21 shows. Note that t~ and y must
frequency domain (where the bandwidth should be
be obtained from the closed loop operated by the
increased). Perhaps the best strategy could be to start generic optimal controller structure.
with method 1 and finish with 2 in the vicinity of an
optimum.
Because ~ depends on the model M, only an
It is also interesting to investigate the relationships iterative control refinement procedure can be
between the identification and control errors in performed. Its simplest - the so-called "relaxation"
d i s t u r b a n c e - f r e e cases (i.e., w = 0 ) . Simple type - iteration can be formed in the following way
calculations give the following very surprising for an off-line case using N samples (the i-th
relationships iteration is shown):

e I = eI (43) 1. Calculate the auxiliary variable /~i based on the


available model Mi_ l , the reference model P~
R and the applied reference signal series
E2 - - - A e2 (44)
l+ RM y ~ = { y [ ( k ) ; k = l ..... N}
I~3 = R A e 3 (45)
ui(k) = M[llPr Yr(k) ; k = 1..... N (47)
Equations (44) and (45) show uncertain and not
deeply analyzed behavior of these combined Here i denotes the index of the iteration; note
identification and control (or adaptive) schemes. In a
d i s t u r b a n c e - f r e e (or n o i s e - f r e e ) case the that y~ does not necessarily change by iteration.
identification errors e 2 and e 3 can be zero, if the
control errors e 2 and e 3 are zero, in spite of a 2. Identify a model between /~i and Yi-! using the
nonzero modeling error A. This bad feature can halt modeling step
the identification at an unpredictable wrong model,
and prevent any iterative refinement. The generic M i = a r g M~..#¢
min Qic(d[,a,t~i, gi_l, Pw)=
optimal controller scheme behaves well because of (48)
(43), i.e., the identification and control errors are the = Me(d[, Mi_ I, Ri_,,y r, S, Pr, Pw)
same, and they can vanish only at the same time.
a,--{ai(k);k--1 ..... N}
Comparison of the forms of ~2 and e 3 explains the
Yi-i = {Yi-i(k); k = 1..... N }
special well-known method of prefiltering A (the
data used for the identification) by 1/(1 + RM), i.e., 3. Calculate the optimal regulator based on (7)
by the reciprocal of the model-based characteristic
equation to make I~3 similar to e 2 . In this way one
Ri : CM[I _ Pw MFI (49)
can introduce the special closed-loop relevant weight 1-P w
H 2 even if the identification is performed in a closed
loop, but using the parallel in-loop optimal then compute the process input u i as
controller structure.
ui(k)= Pr (Ri + M]-I)y~(k) - Ri Yi(k)
(50)
6. ON THE GENERIC OPTIMAL CONTROLLER k = l ..... N
SCHEME
and apply to the process.
Inverse stable processes. It was demonstrated above
that the generic optimal controller scheme has . (There is a possibility of optimizing the applied
certain advantages compared to others shown. It is reference signal series in this step by a proper
also interesting to show that the common modeling input design procedure, see Section 8.)
and control error (e 1 = e 1) given by (35) can also be
expressed as . Once M i and Ri have been found, one can
continue to increase the closed-loop bandwidth
by repeating the procedure. The iterative process
el=el=- H1 Yr + Elw = is continued from step 1, while no stop condition
(46) is fulfilled (until the ultimate control objective is
=-[(Su+ w)- M;,]=-(y- ;,). achieved or it is terminated because of reaching
some vital constraints).
This form and Fig. 12 show an obvious way of
performing the identification step in a combined The robust stability condition (Morari and Zafiriou,
identification and control scheme, i.e., one should 1989) for the generic scheme has the general form
694 L. Keviczky

corresponds to Fig. 6. The identification and control


2~g..=sup R - ~ M g < l • 2~= RM (51) errors are identical in this scheme, too, i.e.,
o i+KM ' I+RM

where, assuming the norm-bounded multiplicative


error el = (1 + RS)M+M_ Yr - 1 + R-----~w = 1 + RS ~'-~+)Yr"

lel = t ~ _<g and


Pw M-I,
R = CM -I = 1- Pw
(52) - I+IRs w=-,,1 [ (M+) +ElW 3 (55)

finally, the condition (51) can easily be further The uncertainty scheme of the extended generic
simplified as optimal controller structure is shown in Fig. 23. It is
easy to see that the only difference from Fig. 13 is
1 that the relative multiplicative uncertainty A/M+ is
sup.-~C g
co I + C
_< suplPwl_<l
~
or suplPwl_<7.
o
(53) now computed on the basis of the invertible M÷.

[ ]
!
r"~ e l L e l ~ ÷It [lU~ ~Wy yr Iw

I
i !
'
I CONTROLLER
I
M,_q
Fig. 23. Uncertainty scheme of the extended generic
IIDENTIFICATIOHN REDGEU
SLIGNOR]
AT
optimal controller structure

It can be shown that the robust stability condition


Fig. 21. Identification and regulator design at the (51) has the form
generic optimal controller scheme
1
Observe that the left-hand side of this robust stability suplP
w ~ M_I<_ or suplv
w~ I-< ~-+ (56)
condition is independent of the model M or the real
process S, and depends only on the desired
disturbance model Pw. where

It is also easy to derive that the characteristic -}-+ = [e÷[-< ~+ (57)


equation of the closed-loop for small modeling error
in the generic scheme is
Observe that this robust stability condition is
K = (M + PwA) a~o = M = 0, (54) independent of the model M or the real process S,
and depends on the desired disturbance model Pw.
which also shows that this basic scheme is not
The characteristic equation of the closed loop for
applicable to inverse unstable models.
small modeling errors in the generic scheme is now
r
I e-e ]1 I~ K = [M+(1 + PwA_) + P~SA+](A_.A+)~0 = M+ = 0
(58)
where

ICONTROLLER [ A+=S+-M+ and A_=S_-M_, (59)

Fig. 22. The extended generic optimal controller i.e., the closed-loop system is stable if M÷ is the
scheme for inverse unstable processes inverse stable factor of the model.

The off-line iterative control refinement procedure


Inverse unstable orocesses. Note that both t~ and R described by equations (47), (48), (49) and (50) and
need the inverse of M. Because this method is based steps 1 - 5 can also be applied here, if instead of
on full pole and zero cancellation, the extended these equations the corresponding
applicability can be reached by partial cancellation of
only those system (model) components which are
inverse stable. This extended generic optimal
^ i 1 -1
ui(k)=(M+- ) PrYr(k) ; k = l ..... N (60)
controller scheme is shown in Fig. 22, which
Combined Identification and Control 695

noisy case the proper parameter-estimation method


(61)
R,- l_Pw (ELS, ML, etc.) should be applied, corresponding to
the given process and noise model structure.) The
outputs of the tracking reference model (continuous)
ui(k)=Pr( RiMi - +(M+) -I )Y~(k)- RiYi(k) (62)
and the controlled process (dashed) are shown in
Fig. 25, before and after the iterative control
k = l ..... N refinement.

formulas are used. Reference nladeland coanolledim~.s$ out, as

1.5
1
7. EXAMPLES FOR OFF-LINE ITERATIVE 0.5
REGULATOR REFINEMENT ( II /
-0.5

Example 7.1 Let the process be given by -1


-1.5
-2
S= 0.007869 z_.4 (63) 20 40 60 80 1~
1 - O.60653 lz -1 2 ~ f ~ . ~ m~.l ~1 o m ~ l t ~ gocess outlmrs

1.5
which is a sampled-time (sampling time is 1
h = 0.05sec and d = 4) first-order approximation of 0.5
Jl /
a helicopter "stick input/roll rate output" model. C ._1 I
-0.5 .- II

0.3 Control loss function -I


0.2

0.1
\ -t.5
-2
20 40 60 80 100
0
0 2 4 6 8 10 12 14 16 18 20
Fig. 25. Outputs of the reference model and the
Identification loss function
0.1 controlled process before and after the
0.05
i
xk iteration (LS)

0 Example 7.2 Let the process (Gevers, 1991) be


0 2 4 6 8 i0 12 14 16 18 20
Number of iterations
given by

Fig. 24. Loss functions in an iterative control 0.0364 z-l(1 + 1.2z -l )


refinement procedure (first-order example) S= (66)
1 - 1.6z -1 + 0.68z -2
Apply the unity gain tracking and disturbance where the same reference input and output noise was
rejection reference models used as in the previous example with the initial
model
0.5z -l 0.2z -l
Pr = -1 and Pw- -1 (64)
1-0.5z 1-0.8z O04z '(l+lOz-')
and start the iterative control refinement by the M° = 1 - 1.4z -1 + 0.5z -2 ' (67)
model
Control loss function
0.3

0.01 z_4 (65) 0.2


M° = 1 - 0.4z -l 0.1

0
0 2 4 6 8 10 12 14 16 18 20
i.e., d m = d = 4. Figure 24 shows the control and
Identification loss function
identification loss functions (variances) by iteration. 0.024
It can clearly be seen that the iteration is quite fast, 0.022
reaching the optimal values after 4 steps. 0.02
0.018
0 2 4 6 8 10 12 14 16 18 20
A unity-amplitude square-wave reference input
Number of iterations
signal with a periodic time of 40 samples was
applied, and the off-line procedure used N = 100 Fig. 26. Loss functions in an iterative control
samples. In the simulation runs an additive white refinement procedure (second-order example)
noise was used as output disturbance, with a standard
deviation ~, = 0.01. A simple off-line LS method was Figure 26 shows the control and identification loss
used for parameter estimation, only to demonstrate functions (variances) by iteration. The outputs of the
the operation of the iterative algorithm. (Experience reference model (continuous) and the controlled
showed that this method works very well in this process (dashed) are shown in Fig. 27, before and
scheme while the noise level is low. However, in a after the iterative refinement performed by off-line
696 L. Keviczky

LS parameter estimation. The outputs of the where the same reference input was used; however,
controlled process (continuous) and identified model the standard deviation of the output noise was
(dashed), before and after the iteration, are shown in ~, = 0.1, causing a much worse signal-to-noise ratio.
Fig. 28. This figure is a nice example to explain the Therefore an ELS
identification procedure was
necessity for iterative control refinement. One can applied to the initial model
see that the identified model output fits very closely
to the process output before and after the iteration, 0 . 1 z - l ( 1 - 0 . 1 z -1)
too, so the model error is small. However, the control (69)
error is very bad before the iteration shown in Fig.27. M°:(l-O.2z-')(1-O.9z-')
2
Reference model and c o n t r o l ~ proems output~
The outputs of the reference model (continuous) and
the controlled process (dashed) are shown in Fig. 29,

%
1.5 rx
1 r -~, i" I i +. ~'. before and after the iterative refinement.
0.5

-0.5
0 , -,f ii :"
Re ferela~ model ~

I
co~Wolled ~ o c e s s OUtpllts

t I

-1

-1.5 0.5

-2
20 40 60 80 100
I +*
2

t.5

1
11 t I
,I j
t i f
i/
I I '

t i ,I

0.5
2o ~o 6o g 100
Reference n'ax~l and com~ll~ process outputs
0

-0.5 ,i
-1

-1.5
o.i
-2
20 40 60 80 100
-o.i
Fig. 27. Outputs of the reference model and the _j ' '/ 'Cr

controlled process before and after the


20 40 60 80 100
iteration (LS)
C o m m l l e d process and iaentifled model o u t ~ t signals Fig. 29. Outputs of the reference model and the
2
controlled process before and after the
12
1
iteration (ELS)
02
Re fecea~ model ~ 1 conll'olled W o c e ~ oetpets
C

-0.5 1.5
1 I J'... ~, I I - ,'x I I ht,
-1

-1.5 0.5
I I r"
¢
-2
20 40 60 80 100
-0.5
"l ix x
-1
Collttolled p f o c e ~ ~ klexttili~ model Output signals I t t I I '
2
-1.5
t I
1.5 -2 i I I llt I
zo ~o 6o m mo
1
Refelence model ~ d contlolled p r o c e ~ otltpllts

S+i J
2
0.5
1.5
0
l
-0.5
0.5
-1
tl
-1.5
4),5
-2
20 40 60 80 I00 -1

-1.5

Fig. 28. Outputs of the controlled process and +2


20 40 60 80 100
identified model before and after the iteration
(LS) Fig. 30. Outputs of the reference model and the
controlled process before and after the
iteration (LS)
Example7.3 Let the process be given by
The same signals are shown in Fig. 30 when the LS
0.125z-1 (1 +0.6z -1 ) method was used. It is easy to see that the ELS
S = (68) method is much better in the case of significant
output noise.
Combined Identification and Control 697

8, ADAPTIVE SOLUTION FOR ON-LINE h = 0.05sec and d = 2) second-order approximation


ITERATIVE REGULATOR REFINEMENT of a helicopter "stick-input/roll-rate-output" model.
Here the same Pr,Pw, ~ and square wave Yr
The previous examples have demonstrated the nice
excitation was used as in Example 7.1. The initial
operation of the off-line iterative control refinement
model was
procedure based on the generic scheme. In spite of
the good convergence properties, the measurements
necessary are remarkable. In many applications it is a 0.001(20 + 0.5z -1 )
(77)
costly and long procedure to design the optimal M0 = (l_l.5z-~ +0.8z -2) z -2
regulator. However, it is easy to construct the
adaptive control refinement procedure based on the
iterative scheme. Following the same steps and i.e., d m = d = 2. The outputs of the reference model
properly changing the iteration i to sampling time k (continuous) and the controlled process (dashed), as
the following formulas are obtained (not discussing well as the control (continuous) and modeling
the details here): (dashed) error signals are shown in Fig. 31 for
N = 2 0 0 samples using the adaptive control
refinement strategy (70)-(75) with TO = 100l and
:-l(k)=(Mk+-l)-lprYr(k ) (7o)
p2 = 0.95.

f(k - d m ) = [/~(k -dm), u(k-d m - 1).... Ref { ~ a ~ model and cOtllrolkxl prlx;ea~ oeltputs
(71)
.... -y(k-1),-y(k- 2).... ]T ii
ol ' '
I I ; I X

+ Kk-_lf(k-dm)fT( k-dm)Kk-1
gk -5 Kk-i + ZT(k- am) J
(72) 50 100 150 200
ConltOl aad modelling ~rot signals
2

Pk =/~k-1 + Kk f(k-dm)[Y(k)-fT(k-dm)Pk-l] 1.5

1
0,5
= .. (73)
0
4).5

gkml_Pw(M+k)-1
.w =
(74)
-I
-1.5
.21
50 100 150 200

k -l
u(k)=Pr[RkMk+(M+) lyr(k)-Rky(k ) (75)
Fig. 31. Operation of the adaptive control refinement

where p is an exponential forgetting factor and


equations (71)-(73) represent a recursive LS method 10. CONCLUSIONS
(R/S) in the simplest so-called "naive" programming
form. Several other recursive parameter estimations This paper has introduced a new structure for the
can be applied instead of the above RLS algorithm. design of optimal pole-placement controllers. This
Note that after finding the estimated parameter new scheme allows one to avoid the explicit solution
vector /~k corresponding to (21), the M÷ and M_ of a polynomial equation obtaining the transfer
function elements of the optimal controller directly.
obtained should be computed by factorization. The The new design principle is quite general, and is
above procedure is an on-line strategy, performing applicable to nonminimum phase (inverse unstable)
all the refinement steps (47)-(50) in one sampling and delay-time systems, as well. The controller is
instance. Here it was assumed that the signal Yr(k) easy to implement in computer-controlled systems.
itself, or its generation rule, is given a priori. The structure of the optimal controller gives a special
insight into the operation of a feedback loop for the
servo and disturbance rejection paradigm. It is easy
9. ADAPTIVE EXAMPLES to see the role of the knowledge of the model of the
system, and the role and appearance of the factors of
Example 9.1 Let the process be given by the system that are invariant to any control strategy.

0.001(1.1 + Iz -1 ) This structure has further advantages in handling


(76) system uncertainties, and new canonical sensitivity
S = (1_1.6693z_1 +0.7788z_2 ) z -2 schemes can also be obtained. The paper investigates
the model sensitivity properties of some combined
which is a sampled-time (sampling time is identification (ID) and control schemes. The
698 L. Keviczky

properties of the simple combination of the open- An applicable strategy for iterative control
loop identification and closed-loop control, the refinement based on the generic scheme was
combined ID and control in a parallel closed-loop presented and illustrated by simulation examples.
optimal controller scheme, in a parallel in-loop The simulation was performed by MATLAB TM.
optimal controller scheme and in a new generic
optimal controller scheme are investigated. It is This is an engineering paper containing no exact
shown that the new generic optimal controller mathematical proofs. However, it tries to explain the
structure is superior to the others. The simple operation and the relationships of the schemes and
combination of the open-loop identification and solutions presented, and tries to show another -
hopefully, new and effective - way for the combined
closed-loop control does not provide good properties
identification and control paradigm. There are still,
for iterative control refinement. The widely used and of course, plenty of open questions that should be
analyzed parallel closed-loop optimal controller analyzed later with proper mathematical theories,
scheme is good for iterative control refinement; tools and methods. It is the author's hope that these
however, it really needs a cautious windsurfer ideas will inspire many other works and applications.
approach, because its robustness is highest around
the crossover frequency coc , which is not known a
11. REFERENCES
priori and can be approached only in an iterative
procedure. The parallel in-loop optimal controller
Anderson, B.D.O. and R.L. Kosut (1991). Adaptive
scheme, mostly applied in adaptive regulator robust control: on-line learning, 30th IEEE Conf.
schemes, is definitely not good for iterative control
Decision and Control, Brighton, England.
refinement, having improper frequency-domain
robustness shaping. It is, however, possible to use a /kstr6m, K.J. and B. Wittenmark (1984). Computer
special prefilter based on the inverse of the model- Controlled Systems-Theory and Design. Prentice-
based characteristic equation for the identification Hall, Englewood Cliffs, N.J.
data, to make this method similar to the previous Gevers, M. (1991). Connecting identification and
one. It is also interesting to know that in a noise-free robust control: a new challenge, 9th IFAC/IFORS
case the last two methods can easily fail because the Symposium on Identification and System
identification errors can be zero, if the control errors Parameter Estimation, Budapest, Hungary, 1-10.
are zero, in spite of a nonzero modeling error. This Keviczky, L. and Cs. B~tny~isz (1994). A new
bad feature can stop the identification at an structure to design optimal control systems. IFAC
unpredictable wrong model, and prevent any further Workshop on New trends in Design of Control
iterative refinement. The new generic optimal Systems, Smolenice, Slovakia, 102-105.
controller scheme seems to be the best among the Landau, I.D. (1990). System Identification and
investigated methods, because its frequency domain Control Design. Prentice-Hall, N.J.
robustness shaping amplifies the high-frequency Morari, M. and E. Zafiriou (1989). Robust Process
uncertainties in the identification error, and a more Control. Prentice-Hall International, Inc.,
accurate model is expected in the targeted high- London.
frequency domain (where the bandwidth needs to be
increased). It also behaves well because the
identification and control errors are the same, and This work was supported in part by US ARO Contract
they can vanish only at the same time. This scheme #N68171-94-0-9064 and by the Hungarian NSF (OTKA)
is easy to implement because general identification
techniques can be applied between a calculated
auxiliary input and the measured closed-loop output
variables.

Potrebbero piacerti anche