Professional Documents
Culture Documents
Accommodation of External Disturbances in Linear Regulator and Servomechanism Problems
Accommodation of External Disturbances in Linear Regulator and Servomechanism Problems
w(t) = H(t)z; w = (zL‘1, . ., 2 U P ) (la) where t.he prime denot,es transpose, S , Q(t) are symmetric
nonnegat,ive-definite n X 12 matrices, ~ ( ut ,) is a positive-
i = D(t)2 + d(t); 2 = (21, * * *, x,), definite scalar function of (u, t ) , and t,he t.ermina1t>imeT
d = (cT~, e ‘ - , u,) (lb) may be eit,her fixed a priori or unspecified.
Since t.he availableplant. out.put. is y = C(t)x, rat.her
where H ( t ) , D ( t ) are presumed kno1t-n mat.rices and t.he than x ( t ) , we nil1 further specify that, t,hr control u”(t)
e1ement.s oi(t)of d(t) are complet.elyunknown sequences of n-hich minimizes (3) mustultimately be expressed in a
random-intensity random-occurring isolated’ delta func- form which can be physically realized by an output-feed-
tions. We will refer t.o the p-vector z as t.he “stat.e” of the back controller: u”(t) = ~ [ y ( s ) s],, t, < s 5 t , to _< t 5 T ,
disturbance w(t). All dist.urbances of the t.ype described in where x[. , . ] is a physically realizable functional.
the previous paragraph, and much more complex2 func-
tions as well, can be simula.ted with t.he model (1) by the A . Th,e Counteraction U o d e of Distwbance Accom.moclatwn
proper choice of ( H , 0). Some pract.ica1 techniques for In [I], [2]it was shown that a control u(t) which regu-
choosing t,he mat.rices ( H , D ) are given in Sect,ion VII. It. lates x ( t ) ---t 0 , and simultaneously count,eract.s a class of
is remarked that. disturbance models of the form (1) were disturbances w(t) could be derived by minimizing a per-
apparent.ly first used in [SI; see also [6], [’i]. formance index of t>hetype (3) where t,he term ?(. , - ) is
We nill now proceed to demonst>rate hon- one can use permit,ted t,o contain also a certain combinat.ion of higher
the model ( 1 ) to design physically realizable regulat.or and order t,ime derivatives of u(t). It n-as furt.her shown in [2]
servomechanism controllers which are remarkably effective t.hat t.he sameform of control could be derived by an
in accommodatingunknown unmea.surable external dis- alternative, and computat.ionally more attractive, method
turbances. based on linear stabilization t,heory. In thissect>ionwe mill
utilize the la.t,ter met.hod, together with the t,echnique of
1 It will be shown in the sequel that the smallest permissible dis-
“st.at.ereconstruction” [ S I , as a device to solve the present
tance 6 > 0 between adjacent impulses of the vector ( u l ( t ) , . . ‘, u p ( t ) ) optimalregulator problem. Our final result will be a
will depend on the physical limitations of the available cont.ro1hard-
ware.
2 The model (1) can sirnulate any function ~ ( twhich
) is piecewise
expressible as linear combinations of the fundamental solutions of I n this paper we ail1 reqilire only linear matrix operations on y
some finite-dimensional time-varyinglinear differential equation. of the form Kg. If rank C = ?n < m, we can always find a maximal
c
This class of funct.ions includes the vast. majority of disturbances rank TLX n matrix such that, for any given K,Ky = Ccx
for some
encountered in realistic control problems. suitable matrix E.Similar remarks apply t.0 H in (la).
JOHKSON: ACCOMXODATIOK O F EXTERNAL DISTURB-4NCES 637
F(t) = B ( t ) r ( t )
for some matrix r(t).
. . If (6) is satisfied and rank B(t) 'r,
r(t)is unique and is given by r(t)= (B'B)-lB'F. Other-
wise, r(t)is nonunique. Assuming (7) is sat.isfied, we nil1
choose the counteractioncomponent of u(t) as D = Tp2 [ ATi2 + FHTz - P12] + T22 [DTzz - f22] (18)
u,(t) = -r(t)$(t)
and the regulating component of u(t) as Examination of (17) reveals several important, facts.First,
note that) if ~ ( t ) 0 the behvior of x ( t ) would be t.otally
If&) = K(t)i (9) independent of the dist>urbanccw ( t ) . Of course, in pract,ice
t,hat. ideal condition is not. realized (e.g., &(to) # 0). HOW-
where $ ( t ) , i ( t ) are accurate estimates of w ( f ) , x ( t ) which ever, if ( t ) is chosen so that. ~ ( t+
) 0 rapidly for all initial
we will obt.ain from y ( t ) by an on-line real-time st.ate re- values of (x,, z, E,) t,he closed-loop plant, st>a.t,ex(t) will be
construct,or device, and K(t)x(t)is t.he cont,rol which would essent.ially insensitive to any and allexterna.1disturbances
minimize (3), (5) if disturbances were absent in (2). We that can be generat,ed by (1). Moreover, in that case x ( t )
will say more about the actualdesign of K(t) in t.hesequel. will be governed, to a close approximation, by the homo-
The details of building a st,at,ereconst,ruct>ordevice for geneous equation
generat.ing accurat,e estimat.esd(t),i ( t ) from on-line mea-
surements of y(s), t, < s <
t, t, t < <
T , are out,lined in +
i = [ A ( t ) B(t)K(t)]x, (E(!) = 0) (20)
Appendix I. The essential result4 is that (4)is finally ex- so that, theregula.t.ingmatrix K(t) can then bedesigned by
pressed in the form of a dynamic out,put-feedbackcont,rol- set,ting 7 = iiR'(t)R(t)iiR(t) in (3) and minimizing (3) for
ler t.he undisturbed system
u = GMy +
H(tE (10) f = A(t)x B(t)iiR(t).+ (21)
where t ( t ) is t,he ( n +p - m)-vect,or output^" of the
For t.his purpose, Q ( t ) in (3) maybe chosen positive
linear dynamic device
definite if one desires t,o regulate x ( t ) to zero, and may be
cZt/dt = D(t)t E(t)y + (11) chosen as Q(t) = C'(t)Q(t)C(t), > 0, if one desires only
t,o regula.te y ( t ) t.o zero. This latter minimization problem
and (herea.fter we occasionally dropt,heargument f on is a sta,ndardproblemin conTrent,ionallinear-qua&atic
matrices in order t.0 simplify t.he notation)
regulat.or
t,lleoV andcan
be
readily
carried
out,
using the
procedures discussed in [16]. The result. ist,hat,under
4 Owing to the uniform symbol convention imposed on all papers appropriate controlla,bility and observabilit,y assumptions,
in this special issue, the meaning of the symbols (y, C, B, F, etc.) in
[?] does not coincide R4t.h the mea,&@ used in this paper, In par- the opt'ima1' E o exist's, is unique, and is gvenby
t.lcular, (t,Z, F , B ) in t.his paper correspond to (- y, - K ~ IB,
, F ) in
PI. iiR,(t) = K(t)x (22)
638 IEEE TRANSACTIONS OK AUTOMATIC CONTROL, DECEMBER 1971
-
tained, if it is t,heoretically possible, by properly letting SI/
m [or Ils(l
-+ \\-here S = C ’ ( T ) S C ( T ) ]in (3). Other-
0)
described in [ l l ] , [12]. It is remarked Ohat if rank B(t) = C. The Maximum Utilization Mode
r , then B f has the explicit form of Disturbance Accommdution
B+(t) = [B’(t)B(t)]-’B’(t). (28) The counteraction-type modes of dishrbance a.ccom-
modat.ion are effective mea.ns for realizing set-point regu-
If one uses (27) as a replacement. for ( 8 ) , replacing w in lationinthe face of persistently acting fluctuating
(27) by the est.imate 6,the corresponding cont,rol (4)has dishrbances. However, those
schemes essentially ig-
t,he same form as (lo), (11) wit.h the exception that the nore any pot,ent,ia,llyuseful regulatory effects t,hat may be
definitions (12)-(15) a.re replaced by associat.ed with the action of t,he exterm1 disturbances. If
+ + + one could somehow exploit the pot.entially useful effect.sof
D = (Tu X C ) [ ( A BK)Tlz ( I - BBf)FHTzz - PI,] the disturbances, it might be possible t o reduce st.ill furt.her
+ T ~ D -TP2d~ ~(29) the tot.al control effort (say, energy) required to achieve
+ + +
E = -DE (T, X C ) [ ( A BK)C*’ - &’I + dX/& regulation.
I n order to makemaximumutilisat.ion of the “free
(30)
energy” associat,ed with the disturbances, we will choose 7
H = KTE - BsFHT2, (31) in (3) as the positive-definite quadratic form
G = -HX + KC#‘. (32) 7 = u‘(t)R(t)u(t),
R(t) >0 (36)
The corresponding expression for (17) then turns out to and choose u(t) to minimize (3) for the plant (2) with dis-
be turbances (1). By t,his means, we\vi11 obtain a controller
(f)
+( -
-
o(t)
Tzzd(t)
) u0(O which lets the ext.erna.1disturbance a.ssist in the job
of regulation, whenever such assistance is available, and
which otherwise minimizes (3) when w ( t ) cont,ainsno useful
(33) effects.
where The minimizat-ion of (3), (36) subject to (l), (2) can be
carried out by st,andard techniques and an outline of such
o = (Z - BB+)FHz. (34) a procedure is given in Appendix111. The essential results
are t,hat,the optimal control uo(t) can be expressed as t.he
The regulating nlatxix K(t) must, now be chosen so as to folloning t,ime-varying 1inea.r funct,ion of ( x ( t ) , z ( t ) , t ) :
take into account the forcing function o(t).This step in-
volves adjoining (lb) to the first of (33) and will be illus- u’(x,Z, t; to, T ) = - [R-’(t)B’(t)K(t)]x
txated in the next section; see remarks below (51). It may - [R-’(t)B’(t)Pl(t)]z (37)
be observed that (29)-(32), (33) reduce to (12)-(15), (17)
whenever (6) is satisfied (because then ( I - BB+)F= 0 ) . where K(t) is an 11 x n nonnega.t.ive-definite symmet.ric
Anot,her possible approach, which is valid for the case matrix satisfying the Riccatidifferent,ial equation
when (1)-(3) are all time invariant,, w ( t ) = const.ant, and K = -A’K - K A +
KBR-IB’K - Q ,
T --t m , is t.0 choose uc so as t.0 minimize some spec.ified
K ( T ) = S (38)
function of the 1imit.ingterminal state x , = limT,, x(?”).
The latteris found by setting u = Kx + u,, i = 0 in (2a) and Pl(t) is an n. X p mat.rix sa.tisfying the differential
to obt.ain equation
X, = -(A + BK)-‘(Bu, + F w ) (35) PI = -( A - BR-’B’R)’Pl - PID - KFH,
where it is assumed that thechosen K is constant andsuch P l ( T ) = 0. (39)
that all eigenvalues of A +BK have negative real parts. Equat.ions(38), (39) canbesolved once and forallby
It is now a st.raightforward algebraic minimization problemintegrating in backwardt.ime using the “initial conditions”
to choose uct.o minimize some selected functionof x,, such K ( T ) = S, P l ( T ) = 0. I\Iat.hematical and computational
as a quadratic form’ x,’Wx,, W 2 0. The resulting solu- aspects of t,he matric Riccati equat,ion discussed are in [la].
tion u, = u,(w) is t.henphysically realized in the form The physical realization of t.he control (37) is carried out,
u,(G) as before. as before, by replacing ( x , z) mit,h (2, 2) where (2, 2) are
It should be noted that one must exercise care when genera.ted by an on-line stat,e reconst,ructordevice (Ap-
state z ( t ) . I n other words, there is no scientific way of im- bances [compare (40), (41) with ( l ) ] . I nthis way, we will
proving the control policy (37), in the sense of (3), when be able t.0 design a servocontroller which can “follow-” a
information about d(t) is completely lacking.* broad class of complex tracking comma,nds yJt) which are
It is remarked that,one must use care in letting T +- co not completely known a priori.
in (3), (36) since then the performance indexwill quite The primary servomechanism problem considered here
likely be unbounded for some disturbances w(t) [in fact,, if may be sta,ted as follows. Our plant is governed by (2),
Q > 0, for any disturbance such t,hat 1imt4- F(t)w(t) f where the unltnou-n disturbances w ( f )are modeled by (1).
01. This is the price one pays for using this particular mode The servocommands yc(t) are not known beforehand, but
of distmbance accommodation. A practical compromise are 1mon-n t>obe accurately modeled by .some solut>ion(s)of
would be to choose a dual mode controller which uses the (40),(41). The problem is to choose a feedback controller
dist.urbance utilization mode for the initial phase of regu- u( .) such t,hat>the plant, output y(t) faithfully follom any
lationandthedisturbance counteraction mode for the a.nd every servocommand yc(t) t,hat, can be generated by
final phase of regulation. (40), (41)-in the face of any plantinit,ial condit,ionsx, and
The ultimatedegree of regulation assistance t,hatr can be every disturbance w(t) that can be generat,ed by (1). We
exploited from the externaldisturbanceisconveniently further require that u ( - )be finally expressed in a phgsi-
measured by setting Q(t) = 0, S > 0 [or S > 01 in (3). cally realizable output-feedback form.
Then thedifference between t.he minimumof (3), (36) with To achieve the condition t.hat, y(t) follows y,(t) we mill
disturbances absent and theminimum of (3), (36) wit.11dis- minimize a performance index of t,he form
turbances present is a measure of the potential usefulness
J [ u ;x,, to, TI = +e’(T)Se(T)
contained in the disturbances.
V. ACCONMODATION
OF DISTURBANCES + 4 ST[e’(t)Q(t)e(t)+ M O ,
to
t ) ~ (42)
PROBLEMS
IN LINEARSERVOMECHANISM
where e denot,es the instantmaneow tracking error,e(t) =
The servomechanism problem is concerned w3h choos-
y c ( t ) - y(t),S, Q are posit,ive-definit,esymmetric matrices,
ing a feedback controller u( .) such that the plant, output
and q(. , .) is some positive-definit.e scalar function of u , t.
y ( t ) faithfully “follows” a t*ime-varyingcommand function
To accommodat,e t.he disturbances in t>his servomecha-
yc(t). I n many practical applications the plant is simul-
nism problem, we can use any of the accommodation modes
taneously subjected t.0 persistently acting external dist,ur-
discussed inSectionIV. To begin, we n-illconsider the
bances w(t) which make it more difficult for y(t) t,o ac-
counteraction mode.
curately follow yc(t). In this section we will show how the
disturbance accommodation techniques described in See- A . Th.e Cou.ntey.crctimLlfo& of D&uu.,.banE Accornmodotion
tion IV for the regulator problem can be adapted to the
Let, the t.otal control u be divided into t.wo parts:
servomechanism problem.
For purposes,
design the class of expect,ed servocom- u = us + uc (43)
mands {yc(t)] is traditionally anticipated a.s having t,he
where, as before, ucis reponsible forcounteract.ing w(t) and ’
form of steps, ra.mps, and other polynomials int.ime. Here,
usis responsible for making y(t) follow yc(t). Thedesign of
y e will expand and generalize this class of expect.ed com-
uc is now ca.rried out. just as in Sect,ion IV, starting wit.11
mands b17 assuming that the yC(t) are generat’ed by the dy- the Same assumptions[especially t,lle essential assumption
na&c command process
[S), (7)]
. .~ a.nd choosing t.he struct,ure (S). The resulting ex-
governed by t.he upper triangular set of equations [com- B . The Minimization Mode of Disturbance Accom.modution.
pare with ( 1 7 ) ] . If the condition (6) is not. satisfied, one can proceedjust
as in Section IV-B and design u , t.0 minimize the effect of
w ( t ) in some specified sense. The result is that the com-
posit,e system (50) then has t,he form (33) with K = 0 and
where a, and x are defined by (lS), (19). We now proceed the term Bus added t.0 i .
to design X so that, E --t 0 [e.g., dist.urbance is counter- Assuming that &(t) = 0 has been achievedbyproper
act,ed) quickly, using the various procedures described in choice of B ( t ) , one can proceed to design the servomecha-
Sect,ion IV-A. Then, for a.11 pra.ct,ica.lpurposes, the plant nism component us just, as in Section V-A. A new term
stat.e x ( t ) in (50) obeys the undisturbed equation analogous to o(t) in (34) nil1 now appea.r in (52); so one
(52) a.nd w.rit.e X = (x1 clz)’,
i = A(t)x +
B(t)u,. (51)
must add the dynamics (lb) to
e = [- CI GlOIX. The expression for uso\+-ill then be ob-
I n order t.0 design t,he servomechanismcomponent. us of tained in the form u,O = K l ( t ) x Kz(t)c &(t)z and the + +
u we now choose 7 in (42) as 71 = us’(t)R(t)us(t), R ( t ) > 0, implementation ca.n be carried out, just as in Sect,ionV-A,
and proceed t.0 minimize (42) along solutions of (40), (41), where for z^ we use (see Appendix I) 2 = T2B(y - zy).
(51). For this purpose, we set X = (x1 c)’, write (40), (41),
C . The Maaimunl Utilization Mod8
(51) as t,he composite system
of Disturbance Accommodation
It is entirely possible that, some of the actionof ext.ernal
disturbances can beput touseful work in t.he servomecha-
nism problem if u ( t ) is chosen appropriately. To explore
and express e as
this possibility, we can proceed as in Sect.ionIV-C choosing
e = [--ClG]i. (53) 71 as (36), setting X = (xlclz)’, e = [-ClGIO]X, and
minimizing (42) subject to t.he composite syst,em (l), (a),
Then, (42) has the form
(40), (41). Thisleadsto a problem a.nalogous to that
J[uS;io, to, TI = + i ’ ( T ) S X ( T ) treated in Section IV-C and can be solved by the same
+ a ST to
+ methods used there (see Appendix 111). The result is t.hat,
[ i ’ ( t ) O ( t ) i ( O u,’(t)R(t)us(t)1 dt (54) uo(t)can be expressed as uo = z ~ ( t ) x-I- z2(t)ci- Z3(t)z
where t,he f i ( t ) are comput,ed by solving mat,ric differen-
where,g S = [ - C l G ] ’ S [ - C I G ] , Q = [ - C l G ] ’ Q [ - C I G ] . tia.1equations of the type(3S), (39). As in Sect.ionV-B, the
The standard linea,r-quadratic regulator t.heory can now implementation of u” is carried out,by use of (45)-(47) (re-
be applied t.0 (52), (54) [disregarding the impulse sequence placing us by u ) and the relations f = (C$’ - T12X)y +
~ ( t for ) reasons analogous to those given below ( 3 9 ) ]with Tlzy; i = Tzz(y - X C ~ ) ; = ( E NG)? - Ny,.
the result that., if ( 2 ) , (54) satisfy certain pa.l%ialcontrol-
+
lability a.nd observability properties,u;(t) exist*, is unique, VI. ACCOMMODATION OF A MOKZERO
and can be expressed as REGULATION POINT
SET
+
u,o = K1(t)x K&)c (55) So far, we have consistently assumed that the desired
regulation set, point, is x = 0 or y = 0. If the desired set
where Kl(t),K2(t)are independent. of x , c and are det,er- point, is nonzero, one can proceed as follows. I n t,he caseof
mined by certa.in auxiliary matric differential equations. st.ate regulation, define the error stat,e vectorX , = X - X “
The case T -+ a must, be approached wit.h the same care and replace (2a) by
as previously indicated for ( 3 ) , (21). I n part.icular, the re-
ma.rksbelow (22) are applicable _. here. Finally, - . to imple-
+
f , = A(t)x, + +
B(t)u(t) F(t)w(t) A(t)x* (56)
ment, (55) me replace ~ ( t )c(t) , in (55) by accurate estimat,es
Itrhere now the object,ive of control is t o achieve Xe(t) = 0.
f ( i ) , ?(t) obtained from appropriate on-line est.imator de-
The termA(t)x* plays the role of an additionalkn.0zL.ndis-
vices operating on the available data (y(s), yc(s)), to < s
5 t, to 5 t _< T . For this latter purpose, we get, f immedi- t.urbance-t,ype term in (56).
One way to acconmlodate the term A ( t ) x * in (56) is to
ately from the already exist,inghardware (45)-(49) by not-
ing t,hat (see Appendix I) f = (C,’ - Tlz8)y Tlzy. The + + +
choose u ( t ) = u, u u p where t.he pa.ir ue,u R counter-
acts w and regulates x, + 0 , as before, and U P counteracts
estimate 2 can be obtained from t,he additional hardware
i? = ( E + N G ) ; - N y , where N is chosen to make the
A ( t ) x * . To a.ccomplish the latt.er, it. is necessary and s a -
cient that B ( t ) u p - A ( t ) x * , which means that @ [ A ( t ) x * ]
error e, = c - 2. approa.ch zero quickly. Since e,(t) is
governed by I., = ( E + +
NG)e,
c
p ( t ) , the stabilizat,ion
, 5 t 5 T , or, in ot.her words
@ [ B ( t ) ]to
methods described below (23) are applicable here. A(t)x* G B(t)<(t;x * ) , for
some
vector [ ( t ; x * ) . (57)
given by t(t;x*) = (B‘(t)B(t))-lB’(t)A(t)x*.Suppose u p holds, the model (1) can accommodate any and every un-
is chosen as indicated anduc,u R are designed as in Section known distmbanc.e function which can be piecewise repre-
IV-A. Then, the result,ing controller will consistently reg- sented as some 1inea.rcombination of the individual modes
ulate x ( t ) to any specified set-point x* in t.he face of any previously incorporated into (1). The “const.ant” weight-
disturbance that can be generated by (1). ing coefficients for t,he individual modes are, of course, as-
If (57) is not. sat,isfied, one can proceed in the spirit of sumed completely unknownand t,hey may even jump from
Sect.ion IV-B and choose u p to minimize, say, IIB(t)up + time to t>imein a random-like manner [this is t,he effect
A(t)x*ll. The minimal norm vector minimizing the latter modeled by t,he unknown impulse sequence d(t)].
quantity is u p = - B t ( t ) A ( t ) x * where B+ hasthe same The use of (1) as a model for disturbances implies that
meaning as in (27). This leaves the knoxm residual forcing the actual dist.urbances are to be imagined, over short, in-
term ( I - BB+)A(t)x*in (56), which must be taken into tervals of time, as the impulse response of some linear
account when choosing uR. homogeneous differential equation. Thus, the problem of
To exploit the possible useful regulatory effects of the choosing a specificnumerica.1 st.ructurefor (1) coincides
set.-point “forcing term” A(t)x*, one can write (3) as with the classic problem of “ident,ificat,ionof a linear dy-
namic system from observation of its impulse response.”
J [ u ;%eo, t o , TI = 3xe’(T)Sxe(T) The lat,ter problem has received considerable attention in
+ ST
4
to
[X;(t)Q(t)xe(t) + u ‘ ( t ) R ( t ) u ( t ) l dt (5s) t.he control literature where a variety of on-line experi-
mental identification schemes have been proposed. It is
and proceed to minimize (5s)subject t o (l),(56). The opti- remarked t.hat since modes of z(t) which do not appear in
mal cont.rol uofor this case can be determined by a minor (la) a.re of no int,erest,,one loses no generality by assuming
alt,eration of t.he procedure used t.0 obt.ain (37) [see Appen- the pair (D(t), H ( t ) ) in (I) is completelyobservable. This
dix IV]. The result. is means that (1) can always be assumed t o be in any one of
t,he canonical forms for complet.ely observable systems.
uqx, 2, t; x*, to, T ) = - [R-’(t)B‘(t)K(t)]x If the disturbance ca.n be mathematically described in
- [ R - l ( t ) B ’ ( t ) p l ( t ) ] z + R - ’ ( t ) ~ ’ ( t ) ~ ( t ) x *(59) piecewise analyticfashion,atime-invariant model (1) can
be obtained by use of the Laplace transform. For instance,
Where K ( t ) , arc given (381,(391, as before, and suppose lc(t) isa scalar and assume that u.(t) can be mathe-
L(t) is given by matically described as
L = [-A‘ + KBR-’B‘]L - Q, L(T) = S . (60) +
~ u ( O= cfi(t) + +
czfi(t) * * ~,ltfN(t) (61)
The case of nonzero output regulation set, point, y x # where t.he ci, i = 1,2, . . .,M , are unknown constants(over
0, can be accommodated inasimilar way. K;amely, we short intervals of time) and the fi(t) are known functions
simply set, v = m , G(t) I , E ( i ) = 0 in (40), (41), and (which are Laplace t,ransformable). Let. S [ f i ( t ) ]= ft(s) de-
writ,e c ( t ) =ys = a known out,put,set.-point.vect,or. Then, note t.he Lapla.cetransform of f f ( t ) and assume that ft(~) is,
proceeding as in Section IV, we set u = uc u R where uc + or can be closely approximated by, a rational algebraic
is designed as before and u R is now chosen as us in (50)- function of s. That is
+
(55). This gives u R = Kl(t)x &(t)y* where y * is com-
pletely known.
VII. CHOOSING
THE DISTURBANCEMODEL(1) where p i , qr a.re finite-degree real polynomials in s and qt
MODEL( M ) , (41)
AKD SERVOCOAIMAKD is monic.
The effectiveness of t,he disturbance controllers described Then, taking theLaplace t,ransform of (61), and collecting
in t.his paper is dependent upon how well the actual distur- terms, gives
bance functions w ( t ) are modeled by (1).Thus, inchoosing
a model (1) to “fit” experimental recordings of dist.urbance
data, the designer should make sure that all significant
waveform modes observed inthedisturbancedataare where O(s) is t.he monic LCD of the product (q1.q2. +
model of the form (1). For instance, one could choose the the present paper can be extended to a more general class
canonicalform [2] w(t) = (1, 0, 0, ., 0)z; i = D,z + of optimal control problems. This topic is explored in [17].
d ( t ) , where Do is
I
APPENDIX
Let Tu@), T&) be, respectively, any pair of once-dif-
ferentia.ble n X (n + p - m) and p X (n + p - m) ma-
Do = . . tricessuch that. [C(t)lO][TlZ’(t)lT22’(1)]’ = 0 andrank
. . . [TI2‘(t)lT,’(t)] s n + p - m for all to 5 t I 5”. Suitable
0 1 choices for these mat.rices always exist and are nonunique -
in general. Then, set Tu = (TI2’Tl2f T22’T22)-lT12‘;TB
= (T12’T12 + T22’T22)-1T22‘. Motivation for t,he structure
Another possible canonical form one might choose is de- of (11) comes from the following. Introduce tThe auxiliary
scribed in [2, eq. (54b)l. variable J! = z y + T12x + T22zwhere z ( t ) is an arbitrary
+
In theevent that experimental dat,aabout actual distur- once-differentiable (n p - m) X m matxix. The deriva-
bances is lacking, one can invoke the argument that over tive of J!,using (l), (2), is comput,ed t o be
sufficiently short int,ervals of time each disturbance com-
ponent wi(t) should be represent.able as some low-order,
&/dt = (Df ZX)a + [(TI2 + XC)(AC” - 6:’)
+ + +
(say, 7nith degree) polpomia.1 int,ime with unknown coeffi- - (D Z X ) Z d X / d i ] y (T12 X C ) h TZd (66) + +
cients (e.g., the classical power series expa.nsionidea). This where a, X are defined by (18), (19).
leads to a very simplemodel (1) in which each independent Now, guided by the structareof (66), let a(t) be the out-
component wf(t) is modeled by dmi+’wi(t)/dtmi+‘= 0. I n put, of a dynamic device built according to the equation
many practical applications, the disturbances vary suffi-
ciently slowly that onecanassume them t.0 be essen- da/dt = (D + ZX)t; + +
[(Tu XC)(AC+’ - e:’)
tially piecewise constant (like randomsteps). For such + + + +
- (a SX)X d X / d t ] y (TI2 XC)BU (67)
cases, an accurate model (1) is obtained by simply set,t,ing
p = p, H = I, D = 0. The cont.rollers described in t.his where u(t),y ( t ) are the aclua2 cont.ro1input and plant out-
paper will t.hen have the form of cIa.ssica1 “proportional- put of ( 2 ) . Define ~ ( t )= -i(t) + +
Xy(t) TI&) +
plus-integral” (PI) or “proportional-plus-integal-plus- Tz2z(t), and let, i , 1 be defined as
derivative” (PID) cont.rollers [l], [3], [17]. The design of i ( t ) = (Ce’ - T12z)~(t) TaE(t),
such controllers for multi-input multi-output time-varying
+
1(t) = 7’22(T;(t) - Z y ( l ) ) (68)
plant,s has t,raditionally been a. difficult problem.
All of the disturbance modeling techniques described in where x ( t ) , z(t), y ( t ) are a d d m y solutions of (l), (2).
t,his section apply equally well to the problem of mat,he- Then, a simple calculat,ionl0shows that.
matically modeling servocommands y,(t). I n t,his way, the
(x(t.1 - i ( t ) ) = T12(04t) (694
elements of ( G ( t ) ,E ( t ) ) in the servocommand model (40),
(41) can be chosen t o fit experimental recordings of rea,l- (z(t) - i ( t ) ) = Tzz(t)e(t) (69b)
time commands y,(l) as measured under t.ypical field con-
ditions. and, from (11, (21, (671,
k = [D(t) + +
z(t>X(t)]E 1122(t)d(t). (70)
VIII. CONCLUSIONS Thus, t,he outputs (a, G ) , G = Hi,of the state reconstruc-
I n many realistic cont.ro1problems, t,he plant t.0 be con- tor (67), (68) are accurat.e est,imatesof (x, w ) provided Z ( t )
trolled is subjected t o persistently acting externa.1 distur- is chosen so t,hat ~ ( t+
) 0 rapidly between impulsesof d(t).
bances which a.re not known beforehand and are not ac- Substituting (68) into (4),(8), (9) gives (lo), and substi-
cessible for measurement, but. which do have a (more-or- tuting (10) int,o (67) gives (11)-(15). Expressions (45)-
less) known set. of possible waveforms. I n t.his paper it has (49) follow fromsimilarconsiderations[within(67),
been shown that if such disturbances can be modeled by (68) replaced by y]. An altermtive form of st,at,erecon-
it is pos-
solutions of some linear different,ial equation, t,hen st.ruct,or, which uses more integrators than (67), (68) but
sible to const.ruct> a dynamical feedback controller which, has a somewhat simpler algebraicst,ructure,is describedin
bymeasuringonly the availableplant. output y ( t ) , can ~71.
maint>ain accurate set-point, regulation(or accurate servo-
txacking) in the face of any such dist,urbances. Explicit, APPEKDIXI1
proceduresfordesigningt,hese“disturbance counterad- The st,abilitv properties of (23), (24) can be est,ablished
ing” controllers have been given. I n addition, it. has been by Lpapunov’s second method. For t,his purpose, choose
shown horn to design regulator and servomechanism con-
t,rollerswhich make optimal utilizationof any useful effects
that, may be present in the action of ext,ernal dist.urba.nces.
‘O-Herewe haJe made m e of the four_algekraic identities CI’C +
~ ~0 ; T % TI : -
T12TI? = 2; T I ? T = -= -
- 0: TZTB 2, which are a natural
The dist,urbance accommodat,ion t,echniques described in ~ ,T’?, T ~ ~ -
consequence of t,he definit,ions of T ~ T?~,
644 IEEE TR4NS9CTIOh’S OK -4UTOMATIC CONTROL, DECEMBER 19’71
V ( t ) = e’@-l(t)& and note that, under t.he condition ( 2 9 , Pl(t) in (74) are given by (38), (39) just as before, and the
the derivative of V ( t )between impulses of d(t) is new t,erm L(t) is given by (60). Expression (59) then foI-
lon-s from the fact t,hat x, = x - x*.
3 = -e’@-l(t)o(t)@-i(t)e, g ( t ) > 0. (71)
Now, under the conditions stated below (23), t,he solu- REFEREKCES
C. D. Johnson, “Optimal control of the linear regulator wit.h
tions @(t) of (25), corresponding to any @ > 0,
, are always const.a.nt.disturbances,” I E E E Trans.Automat.Contr. (Short
positive definite and uniformly boundedawa.y from t,he null Papers), vol. AC13, pp. 41+21, Aug. 1968.
matrix [15]. It, follows t.l& one can always choose Q ( t ) > -, ‘.Further studyof the h e a r regulator with dkturbances-
The case of vector disturbances satisfying a linear differential
0 so t,hat : 1) is negat,ive definit,e and uniformly bounded equat,ion,” IEEE Tra.ns. Automat. Confr. (Short Papers), vol.
AC-15, pp. 222-2228, Apr. 1970.
away from the null matrix, and 2) for any E # 0, 17 can be _ _ , “Further commentson‘Optimalcontrol of the linear
made as small as desired. regulator with constant disturbanw-’,” I E E E Trans. Automai.
Contr. (Corresp.), vol. ACl.5, pp. 516518, Aug. 1970.
APPEKDIXI11 C. D. Johnson and R. E. Skelton,“Optimal desaturation of
momentum exchange cont,rol systems,” A I A A J., vol. 9, no. 1,
The minimization of (3), (36) subject to (l),(2), and the pp. 12-22, 1971.
R . E. Kalman and R.W . Koepcke, “Opt.imal synt.hesis of linear
condit,ion t,hat d ( t ) = 0 can be ca.rried out by mca.ns of the samplingcontrolsystems using generalized performance in-
Hamilton-Jacobitheory. First, let, (I), ( 2 ) be combined dexes,” Trans. AS:WE, vol. 80, pp. 1820-1826, Nov. 1958.
A. T. Fuller, “0pt.imization of non-linear control systems with
into onecomposite system x = Ax + Bu where x = transient input*,” J . Electron. Confr.,vol. 8, pp. 46;i-479, 1960.
-, “Optimization of non-linear control systems n-it,h random
(xlz)’, and n-rit,e (3), (36) a.s J [ u ; To, t,, TI = (1/2)x‘(T) inpnts,” J . Electron. Contr., vol. 9, pp. 65-80, 1960.
%(T) + JE [X’GX+
(1/2) u’Ru] dt. Then, set v(?,f) = -4.E. Bryson, Jr., and 11. G. Lnenberger, “The synthesis of
regulator logic using state-variable concepts,” Proc. IEEE, 1701.
J[u’; x , t , TI, X, = X , to = t. It. can beshown that thefunc- 58, pp. 1803-1811, Kov. 19’70.
tion u(X,t ) sat.isfiesthe Hamilton-Jacobi partial differen- R. E. Kalman, P. L. Falb, and AT. A . Arbib, Topics i n Mathe-
tial equation [16] matical SystemTheory. Nerr York:McGrax-Hill, 1969.
H. D’Angelo, Linear Tinae-Va.rying Systems. Boston: Allpn
and Bacon, 19’70.
C. E. Langenhop, “On generalized inverses of matrices,” SIAM
J . A p p l . Math., vol. 15, pp. 1239-1246, Sept,. 1967.
R. E. Kalman and T. s. Englar, “A user’s manual for the auto-
mat.ic synthesisprogram,” NASA, Washington, D.C., Con-
tractor Rep. CR-472, June 1966.
W . 3.1.Wonham, Random different.ia1 equations in cont.ro1
theory,” ProbabilisticMethodsin. Applied Mathematics, A. T.
and t.he optimalcontrol uo is given by uo(X, t) = Bharucha-Reid, Ed. Xew York: Academic Press, 19’70.
“Optimal stochastic control,” An~.fomalica,vol. 3, pp. 113-
- R-lB’V;W. m e now assume a nonnegative-definit.e spm-
~
118, ’1969.
metric solution t.0 (72) of the form W(X, t ) = (1,/2))x’P(t)X R. E. Kalman, “Contribut.ions to the theory of opt,imal con-
trol,” Rol. SOC.Math. :Vex., vol. 5, pp. 102-119, 1960.
where 11.Athans and P. Falb, Optind Control; A n Introduction to the
Theory and Its Application. New York: AIcGraw-Hill, 1966.
C. D. Johnson, “Accommodation of disturbancesin optimal
control problems,” Int. J. Contr., to be published.