for certain Input Constraints
WolfgangReinelt
DepartmentofElectricalEngineering
Linkoping University,S-58183Linkoping,Sweden
WWW:
http://www.control.isy.liu.se/~wolle/
Email:wolle@isy.liu.se
June 1999
REG
LERTEKNIK
AUTO
MATIC CONTR
OL
LINKÖPING
Reportno.: LiTH-ISY-R-2165
Technical reports from the Automatic Control group in Linkoping are available by
anony-mous ftp at the address
ftp.control.isy.liu.se
. This report is contained inthe portablefor certain Input Constraints
WolfgangReinelt
June 1999
Abstract
Wedeterminethemaximumoutputamplitudeofasystem,whenitsinputfullscertain
constraints. Inparticular,theamplitudeandtherateofchange(i.e.therstderivative)
havetobebounded. Weshowpropertiesoftheworstcaseinputandpresentanalgorithm
thatallowstoconstructthisinputandcalculatesthemaximumamplitudeoftheoutput.
Thisproblemhasapplicationswithinthenon-conservativedesignofcontrollersforplants
withboundedinputs. Nevertheless,itisinterestingasasystemtheoretictaskitselfand
thereforestatedseperately.
Keywords: LinearSystemTheory,Maximum OutputAmplitude,HardConstraints.
1 Introduction and Motivation
Linearsystemswithboundedinputsproduceboundedoutputs,stabilityprovided. Giventhe
maximumamplitudeoftheinputandcalculatingthemaximumamplitudeoftheoutputisthe
problemofdeterminingthe
1
-normofthesystem.Amorechallengingtaskistocalculatethemaximumoutputamplitudeforasystemwithinputsthatfullcertainadditionalconstraints.
Formotivationwelookatprocessengineeringforexample,andinparticularatthe
liquid-levelina tank,interpretedas atimesignal. Notonly the liquid-level inatankis bounded
(bythetank'sheight),additionallytheliquidcannotchangeitslevelarbitrarilyfast. Sothis
signalisboundedinamplitude,aswellasitsrstderivativedue tothetime. Regardingthe
liquid-levelasareferencesignalforacontrolsystemwithcontrolsignallet'ssaytheopening
angleofsomevalve(feedingthetank)itisquitereasonabletodesignacontrollerthattakes
careforthepossiblereferencesignalsfortheliquidlevelaswellasfortheamplitudebound
forthecontrolsignal. Theboundforthecontrolsignalis,looslyspeaking,therangebetween
fullyopenandfullyclosed.
Interpretingthisinsystem-theoreticterms,thetransferfunctionfromreferencesignalto
controlsignalisasystemwithaninputthatfullscertainconstraints(bounds foramplitude
and rst derivative) and a bounded output signal. Task of the controller design is now to
designthecontrollersothatthemaximumamplitudeofthecontrolsignalisnothigherthan
"fullyopen" inorder toprevent wind-up and stability problems. Therefore itis obviously
helpfultocalculatethemaximumoutputamplitudeforagivensystemwithprescribedinput
constraints onamplitudeandrstderivative.
At a rst glance it might be suÆcient to neglecet the constraint on the reference
resultingcontroller willbetooconservative. Duringrun-time,theamplitudeofthecontrol
signalwillbemuchsmallerthantheallowedone;thiseectwillobviouslyreducethecontrol
systems'sspeedandperformance.
However, the solution of the technicalproblem has many applications in the design of
controllers thatguaranteeboundedcontrolsignals[1]. In particular,optimalcontrol
prob-lems can be solved [4, 5], as well as robust control problems [6, 7]. The term optimal
refers to the fact, that the constraint
|
r(t)
_| ≤
_R
allows a minimization of the error signal|e(t)| = |r(t) − y(t)|
. An extension for the discrete time case is possible [3, 2, 8]. Thesecontrolapplicationswillbepresentedseperately.
Thisworkisorganizedasfollows: section2denestheproblem,whichwillthenbesolved
in section 3. Section 4 sketches shortly the numerical realisation. The ideasare extended
to the multivariable case in section 5, suitedfor multivariable control systems. Section 6
illustrates the theory with anexample. Conclusions aredrawnin section 7, related works
as well as thedirectionof future research are also given. Problem statementand solution
aredue to[1,4], anumericalimplementationwasoutlinedin[5]and theextensiontothe
multivariablecasewasgivenin[6, 7].
2 Problem Statement
We examinea linear and time invariantstable system, which is represented by its transfer
function
G(s)
resp.itsimpulseresponseg(t)
. Wepostponetheextensiontothemultivariablecase tosection 5and concentrate on the SISO case. The input is denoted by
r
, the outputby
x
. The followingconstraints hold forthe continuousand piecewise 1 dierentiable input signalr
:|r(t)| ≤ R
(1)|
r(t)
_| ≤
_R
(2) fort > 0
,whereR,
_R > 0
aregivenconstantvaluesandr(t) = 0,
t
≤ 0.
(3)Wecallthosereferencesignals,whichfulleqns.(1-3)
(R,
_R)
-admissible,orshortr
∈ A(R,
_R)
. WearelookingforthemaximumamplitudeX
m
(t)
oftheoutputx
(uptoacertaintime)with(R,
R)
_-admissibleinputs,i.e.
X
m
(t) :=
supr
∈A(R,
R)
_ sup0<τ
≤t
|x(τ)| =
supr
∈A(R,
R)
_ sup0<τ
≤t
|g(τ) ? r(τ)|.
(4)where"
?
"is the convolution:g(t) ? r(t) =
R
t
0
g(τ)r(t − τ)dτ
. Quiteclear fromlinear systemtheoryis,thatforsystemswiththeonlyinputconstraint(1),themaximumoutputamplitude
isgivenbymax
t
≥0
|x(t)| = R
R
∞
0
|g(τ)|dτ
,producedbythebang-banginput:r(t − τ) = R
·
sign(g(τ)).
(5)Thustheproblemistrivialunlesstheadditionalconstraint(2)isimposed.
1
We now turn to the construction of the maximum output amplitude as stated in eqn.(4)
according to [4]. Basic idea is rst to show some properties of the input signal
r
, whichproduces the output with the maximum output amplitude. In the following, we call this
inputsignaltheworstcaseinput. Thisstrategyismotivatedbytheexistenceofaworstcase
inputinthesimplecaseineqn.(5). Westartwiththecalculationoftheoutputbyconvolution
foracertaintimestamp
t
:x(t) =
Z
t
0
g(τ)r(t − τ)dτ =:
Z
t
0
g(τ)r
t
(τ)dτ
(6)andabbreviatethe"timeinverted"inputsignalby
r
t
(τ) := r(t − τ)
withthefollowing conse-quencesfortheconstraints(1-3):|r
t
(τ)
| ≤ R, τ < t
(7)|
r
_t
(τ)
| ≤
_R,
τ < t
(8)r
t
(τ)
=
0,
τ
≥ t
(9)3.1LemmaThefunction
X
m
(t)
ismonotoneincreasing(i.e.notdecreasing)intimet
.There-fore, the maximum amplitude as dened in eqn.(4) appears for sure for
t
→ ∞
, soX
m
=
limt→∞
X
m
(t)
isthemaximumoutputamplitude.Proof. Let
t
0
> 0
andr
t
0
∈ A(R,
_
R)
theinput 2thatproducesthemaximumamplitude
X
m
(t
0
)
.For
t
1
> t
0
dener
t
1
(τ) = r
t
0
(τ), 0
≤ τ ≤ t
0
resp.r
t
1
(τ) = 0, t
0
< τ
≤ t
1
. Clearlyr
t
1
∈ A(R,
_R)
andfromeqn.(6) followssup
0<τ
≤t1
|x(τ)| =
sup0<τ
≤t0
|x(τ)|
andthusX
m
(t
1
)
≥ X
m
(t
0
)
.2
3.2DenitionSupposethereexists aninputr
∞,o
=: r
o
withmaximumoutputamplitudeX
m
. Then−r
o
producedobviouslythemaximumoutputamplitudeX
m
,too. Foroneofthem,sayr
o
,holdsX
m
=
Z
∞
0
g(τ)r
o
(τ)dτ
≥ 0,
(10)i.e., ineqn.(4) the absolute valueis super uous. In thefollowing, we constructthisworst
caseinput 3
thatproduces thismaximumoutputamplitudeaccordingtoeqn.(10).
3.3Algorithm (Construction ofanauxiliaryinput)Let
r
∈ A(R,
_R)
beanarbitraryadmissible input. Weconstructan auxiliaryinputr
H
forr
uniquelyby the steps givenbelow (gure 1 illustratestheconstruction). Thesetofallpossibleauxiliaryinputs (i.e.allsignalswiththesameproperties)isdenotedby
A
H
(R,
_R)
.1. Let
t
i
, i = 1, . . . , N
thezerosofg
. Dener
H
(t
i
) = r(t
i
)
.2. If
g(t) > 0
in(t
i
, t
i+1
)
, letr
_H
(t) = +
_
R
in the neighbourhood oft
i
andr
_H
(t) = −
_R
inthe neighbourhood of
t
i+1
. In the case that this denition leads to the non-unique situationthat the two'slopes' intersect insomet
∗
∈ (t
i
, t
i+1
)
, letr
_H
(t) = +
_
R
in[t
i
, t
∗
]
resp.r
_H
(t) = −
_
R
in[t
∗
, t
i+1
]
. Denenallyr
H
=
min{r
H
, +R
}
. 2uniquenessisnotnecessaryforthisargumentation.
3
0
0
0
g(t)
_r
H
(t)
+R
−R
+
R
_−
R
_r
H
(t)
r(t)
t
t
t
t =
∞
t =
∞
t =
∞
t
1
t
2
t
3
Figure1: Construction ofauxiliaryinput
r
H
forgiveninputr
.3. If
g(t) < 0
in(t
i
, t
i+1
)
, do as in step 2. but with changed signs forr
_H
and resultingobviousmodications.
4. Finally,choose
|
r
_H
(t)
| =
_R
forlarge timest
inordertoensurelimt
→∞
r
H
(t) = 0
, arising fromconstraint(9).3.4Corollary thefollowingpropertiesoftheauxiliaryinputareclearbyconstruction:
1.
A
H
(R,
_R)
⊂ A(R,
R)
_, i.e.anauxiliaryinputisalsoadmissible.
2. Twodierentinputs
r
1
, r
2
∈ A(R,
_R)
withr
1
(t
i
) = r
2
(t
i
)
havethesameauxiliaryinput.3.
r
H
(t)
≥ r(t)
forg(t)
≥ 0
andr
H
(t)
≤ r(t)
forg(t)
≤ 0
.4. Fixanadmissibleinput
r
andsupposeanarbitraryadmissiblesignalr
∗
6= r
H
with prop-erty3.,thenR
∞
0
g(τ)r
H
(τ)dτ >
R
∞
0
g(τ)r
∗
(τ)dτ
.5. Themaximumwidthofthepulsesof
r
_H
inAlgorithm3.3is givenbyT = 2
R
_
R
.3.5TheoremThe worstcaseinputisauxiliaryinput:
r
o
∈ A
H
(R,
_Proof. Forall
r
∈ A(R,
_R)
thefollowingholds byconstructionofr
H
,seeCorollary3.4(3.):Z
∞
0
g(τ)r(τ)dτ
≤
Z
∞
0
g(τ)r
H
(τ)dτ
(11)and"
=
"holdsonlyforr
≡ r
H
(exceptthecaseg
≡ 0
). Assumer
o
∈ A(R,
_R)\A
H
(R,
R)
_,thenthe
constructionofanauxiliaryinput
r
oH
is possible(becauser
o
is admissibleinput). Applying eqn.(11)tor = r
o
:X
m
=
Z
∞
0
g(τ)r
o
(τ)dτ <
Z
∞
0
g(τ)r
oH
(τ)dτ
(12)which contradicts the denition of
X
m
as the maximum output amplitude. Consequentlyr
o
∈ A
H
(R,
R)
_.
2
Until now, we did not construct a unique worst case input thatleads to the maximum
outputamplitude,butweshowedsomenecessarypropertieswhicharelistedinthefollowing
3.6LemmaThefollowingnecessarypropertiesoftheworstcaseinput
r
o
hold:1. The(derivativeofthe)worstcaseinputhasapulse-shape:
r
_o
(t)
∈ {±
_R, 0}
,where_r
o
(t) =
0
implies|r
o
(t)
| = R
.2. Thewidthofthesinglepulsesisconstrainedby
T = 2
R
_
R
. 3. Twoadjacentpulseshavedierentsigns.4. lim
t
→∞
r
o
(t) = 0
and|
limt
→∞
r
_o
(t)
| =
_
R
.The previous Lemmastates mostly properties of
r
_o
. Partial integration of eqn.(10)to-getherwithLemma3.6(4.) givesanexpressionforthemaximumoutputamplitudeinthese
terms,where
s
isthestepresponseofthesystem(i.e.s = g
_ ):X
m
= −
Z
∞
0
s(τ)
_r
o
(τ)dτ
(13)Despitetheminusineqn.(13),
X
m
≥ 0
holdsbyDenition3.2. Itappearsonlyduetothe partialintegration.Lookingontoeqn.(13)andknowingtheshapeoftheworstcaseinputasstatedinLemma3.6,
thesolutionisquiteclear: inordertomaketheintegralmaximal,putsomepulses(of
maxi-mumwidth,seeCorollary3.4(4.)) inthenearofextremaofthestepreponse: positiveones
inthenearoftheminimaandnegativeonesinthenearofthemaxima. Inthefollowing,we
willprove this. Inorder nottooverloadthediscussion withtechnicaldetails,wemakethe
1. Lettheimpulseresponse
g
haveonlyanitenumberN
ofzeros,i.e.thestepresponseonlyanitenumberofextrema.
2. Lettherstextremumof
s
tobea(local)maximum.Undertheseassumptions,themaximumamplitude
X
m
is givenby:X
m
=
R
_N
X
i=1
(−1)
i+1+k
Z
t
i
00
t
i
0
s(t)dt + R(−1)
N+k
limt
→∞
s(t).
(14)The lastpart of the sum exists becausethe system is stable. The pairs
(t
0
i
, t
i
00
)
refer to the unknownpositionsofthepulses. Additionally,thesignofthepulsesisstillunknown,there-fore we added
k
∈ {0, 1}
in eqn.(14). Obviously the problem is solved, when weknow the exactlocationofthepulsesandtheirsign. Wemakeonemore3.8TemporaryAssumptionLettheextremaofthestepresponsehaveadistance
> 2T
,whichensuresthatallpulseshavemaximumwidth
T
,i.e.t
00
i
= t
i
0
+ T
.3.9TheoremNecessaryconditionforamaximum
s(t
0
i
) = s(t
i
0
+ T )
,suÆcientconditionisk = 0
ineqn.(14).
Proof. Necessaryforamaximumis
∂X
m
∂t
i
0
= 0
foralli
,whichleadsimmediateleytos(t
0
i
) = s(t
i
0
+
T )
. SuÆcientis thattheHessianis negativedenite,whichleadsdirectlyto_
R(−1)
i+k
(g(t
0
i
+
T ) − g(t
i
0
)) < 0
thusk = 0
, becauseofassumingarstlocalmaximumofs
inAssumption3.7.2
ApplyingTheorem3.9toeqn.(14),wegetthemaximumoutputamplitudeby
X
m
=
R
_N
X
i=1
(−1)
i+1
Z
t
i
0
+T
t
i
0
s(t)dt + R(−1)
N
limt
→∞
s(t).
(15)whereweplacethepulses
(t
0
i
, t
i
0
+ T )
,sothats(t
0
i
) = s(t
i
0
+ T )
holdswithanegativesignunderthemaximaof
s
andwithapositivesignundertheminimaofs
.3.10Remarkontemporaryassumptions3.7and3.8:
1. Thestabilityofthesystemensures
g(t)
→ 0
forlarget
,i.e.s(t) =
const. Inthecase,thatg
has a innite numberof zeros the maximum output amplitude is givenby eqn.(15) withanarbitraryprecisionwhen usingarbitrarymanyzeros(i.e.N
→ ∞
)Foranexactproof,wereferto[4 ,AppendixA2.1].
2. Supposearstlocalminimumfor
s
. Inthiscase,thesuÆcientconditionwouldleadtok = 1
, i.e.toachangeofsignsforallpulses. We cansimplytreatthesedierentcasesbytaking theabsolutevalueofthesumineqn.(15).
3. The"well-distinctness"oftheextrema,assumedinAssumption3.8ensuresthatpulses
withwidth
T
posed aroundtheextremaofs
willnotintersecteachother. Technically,this assumption simplies the proof of Theorem 3.9. In general, we can prove that
R
t
i
00
t
i
0
g(t)dt = 0
is necessary and suÆcient for a maximum. The proof is of technicalT
,isviewedasonemaximumresp.minimum. Fortheentirediscussionofallcases,we referagainto[4, AppendixA2.2].3.11 Remark For _
R
→ ∞
(i.e. no restriction on the input's speed), the worst case input convergestothebang-banginput,asgivenineqn.(5).4 Numerical Algorithm
Theorem 3.9 and Remark 3.10 give us necessary and suÆcient conditions to construct the
worstcaseinput. Themaximumoutputamplitude canbecalculatedwithhelp oftheworst
caseinputandeqn.(15). Werestrictourselvesonanotveryformal descriptionofthe
algo-rithmthatconstructsthe worstcaseinputnumerically. Numericaldetailsand thehandling
ofspecialcasesareoutlinedin[5].
1. Determineextremaofthestepresponse.
2. Putinitialpulseswithcorrectandalternatingsignsundereachextremumof
s
.3. Makethemwider,ensuring
s(t
0
i
) = s(t
i
00
)
,aslongastheydonotintersect.4. Inthecaseofintersections,handle"not welldistinct"pulsesseparately.
5. Calculatethemaximumoutputamplitudefromeqn.(15).
The problem canalso be viewed as a nonlinear optimisation problem in the pulse
po-sitions: determine
t
0
i
, t
i
00
, so that rhs of eqn.(15) becomesmaximal. Indeed, a solution by nonlinearoptimisationispossible,havingagoodinitialvaluegivenbystep(2.) ofthepro-cedureabove;thissolution isoutlinedin[6].
5 Multivariable Case
Weextendthepreviousresulttothecaseofmultivariablesystems,i.e.
r
andx
arenowvec-tor signals[6 ]. What wehaveinmind withthisextensionis thetreatmentof multivariable
control systems with constraint control signals, i.e.weregard
x = u
andr
is the (external)referencesignalof thecontrol system. Therefore, itis usefultorestrict theinput
r
compo-nentwise:
|r(t)| R, t > 0
(16)|
r(t)
_|
_R, t > 0
(17)r(t)
=
0, t
≤ 0.
(18)inanalogy toeqns.(1-3). Read
asacomponentwise
≤
and evaluate| · |
therefore compo-nentwisely. Consequently, wecall the set of allsignalsr
fullling these constraints (R,
_
R
)-admissible, with
R,
_R
arenow beingvectors with positive entries. Furthermore, we dene themaximumoutputamplitudecomponentwisebywherethe
X
i,m
aredenedasinLemma3.1.Letusrstlookontoasystemwithone output
x
andk
inputsr = (r
1
, . . . , r
k
)
T
∈ A(R,
R)
_ .Then
x(s)
isgivenbyx(s) = G
1
(s)
· r
1
(s) +
· · · + G
k
(s)
· r
k
(s)
(20)Weabbreviate
x
~i
(s) = G
i
(s)
· r
i
(s)
. Now wearelookingforthemaximumoutputamplitudeX
m
. Usingeqn.(20),themaximumoutputamplitudeisgivenbyX
m
=
k
X
i=1
~X
i,m
.
(21)Itfollowsdirectly,that
X
m
isachievedforacertainvectorr = (r
1
, . . . , r
k
)
T
∈ A(R,
R)
_ .Inthemultivariablecasewith
n
outputs,wesimplyapplytherststep: accordingtothedenition,thecomponents
X
i,m
ofX
m
canbecalculatedasinequation(21).6 Illustrative Example
Weexaminethesystemrepresentedbythetransferfunction
G(s) =
s
2
+ 0.4
s
2
+ 1.4s + 1.0
,
with input constraints
R = 1.0
resp. _R = 0.8
. According to Lemma 3.6 (2.), the maximumpulsewidthis
T = 2.5
. Numericalsolution byconstructionoftheworstcaseinputyieldsX
m
= 0.76.
(22)Fig. 2 shows the construction of the worst case input with pulses of
r
o
located at[0, 1.20]
,[1.20, 2.39]
,[5.24, 7.74]
and at of widthT/2
at innity (in reversed time) with signs "- +-+" following max-min-max sequence of the step response. For simulation of this worst
case input we need to reverse this reversed time, therefore we choose the "innite" time
to
t
∞
= 120
. With this transformation, we receive the depicted worst case output witha maximum amplitude of
x(t
m
) = 0.73
fort
m
= 108.80
as a good approximation for themaximumamplitudeascalculatedineqn.(22).
7 Conclusions and Related Works
Weshowedconditionsoftheworstcaseinputwithboundedamplitudeandspeed,that
pro-duces thatmaximumoutputamplitudeforagiven(stable)multivariablesystem. A
numeri-calalgorithmwassketchedtoconstructthisworstcaseinputandtocalculatethemaximum
outputamplitude. Thisis,aswebelieve,animportantstepwithinthenon-conservative
con-trollerdesignforsystemswithhardboundsonthecontrolsignal. Minimisationoftheerror
signalinacontrolsystemispossiblewiththeseassumptionsonthereferencesignal.
More-over itenables us to check the maximumamplitude anarbitrary signal within the control
0
1
2
3
4
5
6
7
8
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t
s(t)
(a)Step response
s
with calculated posi-tionofthepulsesfortheworstcaseinput.0
20
40
60
80
100
120
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
t
r_o(t)
(b)Worstcaseinput.
0
20
40
60
80
100
120
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
t
y(t)
(c)Worstcaseoutput.
[1] N. Dourdoumas. Prinzipien zum Entwurf linearer Regelkreise mit Beschrankungen {
eineEinfuhrung. Automatisierungstechnik,35(8):301{309,Aug.1987.
[2] P.Muller. EntwurfoptimalerAbtastregelkreisebeibeschrankten Systemgrossen.
Mas-ter'sthesis,DeptofEE,UniversityofPaderborn,33095Paderborn, Germany,1994.
[3] X. Peng. Rechnerunterstutzte Synthese von Abtastregelkreisen mit Beschrankungen.
PhDthesis,DeptofEE,UniversityofPaderborn, 33095Paderborn,Germany,1992.
[4] R. W. Reichel. Synthese von Regelsystemen mit Beschrankungen bei stochastischen
Eingangsgrossen. PhD thesis, Dept of EE, University of Paderborn, 33095 Paderborn,
Germany,1984.
[5] W.Reinelt.EntwurfoptimalerReglerbeiamplitudenbeschrankten Systemgrossen.
Mas-ter'sthesis,DeptofMath,Univ ofPaderborn,33095Paderborn,Germany,Dec.1993.
[6] W. Reinelt. RobustControl ofSystems subjecttoHardConstraints. PhD thesis, Dept
ofEE,UnivofPaderborn, 33095Paderborn,Germany,Apr.1998.
[7] W.Reinelt.
H
∞
loopshapingforsystemswithhardbounds. InProc.oftheIntSympon QuantitativeFeedbackTheoryandRobustFrequencyDomainMethods,Durban,SouthAfrica,pages89{103,Aug.1999.
[8] I. Zacharias. Robuste Regelung von Abtastsystemen unter Berucksichtigung von
Stell-grossenbeschr ankungen. Master'sthesis,DeptofMath,UniversityofPaderborn,33095