Upload
alex-vlad
View
220
Download
0
Embed Size (px)
DESCRIPTION
da
Citation preview
Time Delays in Control and Estimation (038806)
lecture no. 6
Leonid Mirkin
Faculty of Mechanical EngineeringTechnionIIT
esh
Outline
Fiagbedzi-Pearson reduction
Smith controller revised
Modified Smith predictor and dead-time compensation
Modified Smith predictor vs. observer-predictor
Coprime factorization over H1 and Youla parametrization
Nontrivial complication
So far, we studied problems in which h has finite spectrum. Let now
h W Px.t/ D A0x.t/C Ahx.t h/C Bu.t/;
which might haveI infinite number of open-loop poles complicates matters
Workaround:I move only part of these eigenvalues by feedback,
namely, unstable ones (we know that there is only a finite number of them).
Yet another transformation
Consider
Qx.t/ Qx.t/CZ tth
eQA.th/QAhx./d
for some QA 2 RQnQn and Q 2 RQnn. Then,
PQx.t/ D Q Px.t/C QAZ tth
eQA.th/QAhx./d C e QAhQAhx.t/ QAhx.t h/
D QA0x.t/CQBu.t/C QAZ tth
eQA.th/QAhx./d C e QAhQAhx.t/
D QA Qx.t/CQBu.t/ QAQ QA0 e QAhQAhx.t/:If we can choose QA satisfying the left characteristic matrix equation
QAQ D QA0 C e QAhQAh; (l.c.m.e.)
h reduces toQ W PQx.t/ D QA Qx.t/CQBu.t/:
Properties of solutions of l.c.m.e. QAQ D QA0 C e QAhQAhLet Q 2 spec. QA/ and Q be the corresponding left eigenvector. Then
Q QQ D Q QAQ D QQA0 C Q e QAhQAh D QQ.A0 C eQhAh/:
In other words, QQ. QI A0 eQhAh/ D 0, so thatI whenever QQ 0, every eigenvalue of QA is a pole of h
(remember, those poles are the roots of h.s/ D det.sI A0 Ahesh/).
Closed-loop spectrum
Suppose we can solve (l.c.m.e.) in QA and Q. Then
u.t/ D QF Qx.t/ D QFQx.t/C
Z tth
eQA.th/QAhx./d
for some QF leads to the closed-loop system
sI A0 Ahesh B QFQ QF e QAh R h
0e.sI QA/dQAh I
X.s/
U.s/
D I.C. :
and the characteristic quasi-polynomial cl.s/ D det.s/, where
.s/
sI A0 Ahesh B QF QC e QAh R h
0e.sI QA/dQAh
I
:
Closed-loop spectrum (contd)
Next,
.sI QA/QC e QAh
Z h0
e.sI QA/d QAh
D sQ QAQC .eAh eshI /QAh D Q.sI A0 Ahesh/;
so that
QC e QAhZ h0
e.sI QA/dQAh D .sI QA/1Q.sI A0 Ahesh/:
Then
.s/ D
sI A0 Ahesh B QF .sI QA/1Q.sI A0 Ahesh/ I
and
cl.s/ D det.sI A0 Ahesh/ B QF .sI QA/1Q.sI A0 Ahesh/
D detI B QF .sI QA/1Qdet.sI A0 Ahesh/:
Closed-loop spectrum (contd)
Because QA QB QF I
D QACQB QF Q
B QF I1
;
we have that
detI QF .sI QA/1QB D det.sI QA QB QF /
det.sI QA/and thus
cl.s/ D det.sI A0 Ahesh/
det.sI QA/ det.sI QA QB QF /:
In other words,I spec.h;cl/ D
spec.h/ n spec. QA/
Sspec. Qcl/
(remember, spec. QA/ spec.h/).
Implications
Thus, if we canI solve (l.c.m.e.) so that QA contains all unstable modes of h,I find QF so that QACQB QF is Hurwitz (requires stabilizability of . QA;QB/),
the control law
u.t/ D QFQx.t/C
Z tth
eQA.th/QAhx./d
stabilizes h by moving all its unstable modesthose in spec. QA/to theeigenvalues of QACQB QF and keeping the other modes of h untouched.
Distributed state / input delays
Leth W Px.t/ D
Z 0h
./x.t C /C ./u.t C /d:
Then transformation
Qx.t/ Qx.t/CZ tth
Z th
eQA.tC/Q
./x./C ./u./dd
with l.c.m.e.QAQ D
Z 0h
eQAQ./d (l.c.m.e.0)
yields reduced system
Q W PQx.t/ D QA Qx C QBu.t/; where QB Z 0h
eQAQ./d
andI spec.h;cl/ D
spec.h/ n spec. QA/
Sspec. Qcl/.
Is it that simple ?
Not quite, solving QAQ DZ 0h
eQAQ./d is highly nontrivial. Specifically,
I we have to find all troublesome modes of h(in most cases, have to rely on numerical approaches)
I solve (l.c.m.e.) / (l.c.m.e.0)(solution is non-unique and not especially elegant)
Only a handful of cases where the steps above can be solved analytically.One example is
./ D
2666664 0 ::::::: : :
::::::
0 0 0 0 0
3777775 ./CXi
26666640 0 0 ::::::: : :
::::::
0 0 0 0 0 0 0
3777775 . C hi /
in which case spec.h/ is finite and (l.c.m.e.0) is solvable with Q D I .
Example
Leth W Px.t/ D x.t/C x.t h/C u.t/;
whole characteristic quasi-polynomial is (here s D C j!)
h.s/ D s C 1 esh D C 1C j! ehej!h:
Solutions of h.s/ D 0 must satisfy the magnitude condition
. C 1/2 C !2 D e2h:
If > 0, this equation is unsolvable. If D 0, then ! D 0 is the only option.Indeed, s D 0 is a root. Then, by LHopitals rule,
lims!0
h.s/
sD 1C lim
s!01 esh
sD 1C lim
s!0h
1D 1C h;
which implies that s D 0 is a single root.
Example (contd)
Thus, we have only one unstable pole to shift and may pick Qn D 1, QA D 0.Eqn. (l.c.m.e.) then reads 0 D q C q, so we may pick q D 1. Then
Q W PQx.t/ D u.t/;
which is stabilized by u.t/ D k Qx.t/ for any k > 0. Thus
u.t/ D kx.t/C
Z tth
x./d
stabilizes h and renders its closed-loop characteristic polynomial
h;cl.s/ D s C ks
.s C 1 esh/:
In fact, the controller above has the transfer function
C.s/ D k1C 1 e
sh
s
2 H1:
Outline
Fiagbedzi-Pearson reduction
Smith controller revised
Modified Smith predictor and dead-time compensation
Modified Smith predictor vs. observer-predictor
Coprime factorization over H1 and Youla parametrization
Smith controller: preliminary conclusions
C.s/
reQeud
y- -Pr.s/e
sh QC.s/
Pr.s/.1 esh/
Remember, Smith controllerI works if Pr.s/ is stable
(stabilization problem reduces then to that for delay-free plant)
I does not necessarily work if Pr.s/ is unstable(might lead to unstable loop)
More rigorous analysis
wu
wyyu
Pr.s/esh
C.s/
System is said to be internally stable ifI transfer matrix from
wywu
toyu
,
Tcl 11 PrC esh
1 PreshC PrC esh
S TdTu T
2 H1:
LetsI analyze Smith controller from internal stability perspectives.
Loop shifting
QC.s/
Pr.s/wu
wyyu
Pr.s/esh
C.s/
Pr.s/.1 esh/
Pr.s/.1 esh/ -
Adding and subtracting block Pr.1 esh/ weI redistribute loop components w/o changing the whole system.
We end up with a new loop with the plant Pr and the controller
QC C1C CPr.1 esh/
so that C D
QC1 QCPr.1 esh/
Loop shifting: signal transformations
QC.s/
Pr.s/wu
wyyu
Pr.s/esh
C.s/
Pr.s/.1 esh/
Pr.s/.1 esh/ -
QC.s/
Pr.s/
wu
wy
./wu
yu
Pr.s/esh
C.s/
Pr.s/.1 esh/
Pr.s/.1 esh/ -
-
QC.s/
Pr.s/
wu QwyQy
yu
Pr.s/esh
C.s/
Pr.s/.1 esh/
Pr.s/.1 esh/ -
Loop shifting (contd)
wu
wyyu
Pr.s/esh
C.s/
,Qwu
QwyQyQu
Pr.s/
QC.s/
The new system is delay-free yet withI different signals.
To complete1 the picture, we have to calculate them: QyQuDy C Pr.1 esh/u
u
DI Pr.1 esh/0 I
y
u
and Qwy
QwuDwy Pr.1 esh/wu
wu
DI Pr.1 esh/0 I
wywu
:
1Well, we also have to guarantee that internal QC loop is well posed.
Loop shifting (contd)
Thus,
Tcl DI Pr.1 esh/0 I
QTclI Pr.1 esh/0 I
:
1. if Pr.1 esh/ 2 H1, then QTcl 2 H1 implies Tcl 2 H12. if Pr.1 esh/ 62 H1, then QTcl 2 H1 not necessarily implies Tcl 2 H1
(in fact, never; this can be verified by making use of explicit form of QTcl)
Key question:I when does Pr.1 esh/ 2 H1 ?
Obviously, this is true if Pr 2 H1. Yet this also true ifI Pr.s/ proper and its only unstable poles are single poles at j2h k, k 2 Z,
which are simple zeros of 1 esh.
Outline
Fiagbedzi-Pearson reduction
Smith controller revised
Modified Smith predictor and dead-time compensation
Modified Smith predictor vs. observer-predictor
Coprime factorization over H1 and Youla parametrization
Loop shifting: idea
QC.s/
Pr.s/wu
wyyu
Pr.s/esh
C.s/
Pr.s/.1 esh/
Pr.s/.1 esh/ -
Driving idea here is toI cancel (compensate) delay via Pr.s/esh C Pr.s/.1 esh/ D Pr.s/.
The question:I can we do it with stable compensation element ?
Dead-time compensation question
QC.s/
QP .s/wu
wyyu
Pr.s/esh
C.s/
.s/
.s/ -
Technicaly speaking, we are looking for .s/ 2 H1 such that
Pr.s/esh C.s/ D QP .s/
for some proper and rational QP .s/.
Dead-time compensation: aspirations
If we obtained required .s/ 2 H1, we would transformwu
wyyu
Pr.s/esh
C.s/
,Qwu
QwyQyQu
QP .s/
QC.s/
for rational (delay-free) QP .s/. Resulting C.s/,
C.s/ D QC.s/I .s/ QC.s/1 D yu QC.s/.s/
;
called dead-time compensator (DTC) andI QC.s/ internally stabilizes QP .s/ iff C.s/ internally stabilizes Pr.s/esh as
Tcl.s/ DI .s/0 I
QTcl.s/
I .s/0 I
(provided DTC loop well posed).
DTC question: stable Pr.s/
Choice of .s/ apparent and non-unique:I .s/ D QP .s/ Pr.s/esh for any QP .s/ 2 H1 does the job.
Some standard choices:I QP .s/ D Pr.s/ results in Smith predictorI QP .s/ D 0 results in internal model controller (IMC)
DTC question: unstable Pr.s/ in Example of Lect 5
With Pr.s/ D 1=s, the Smith predictor
.s/ D 1 esh
sDZ h0
esd 2 H1
indeed (integrand analytic and bounded in C0 and integration path finite).In this case
QP .s/ D Pr.s/ D 1s
can be interpreted as unstable rational part of partial fraction expansion2 of
Pr.s/esh D Res
eshsI 0
sC.s/
for some entire .s/ (Pr.s/ has only one pole, that at the origin).
2Here Res.f .s/I c/ lims!c.s c/f .s/ stands for residue of f .s/ at c.
DTC question: another unstable Pr.s/
Let Pr.s/ D 1=.s 1/. Were looking for stable QP .s/ such that
.s/ D QP .s/ 1s 1 e
sh 2 H1:
As Pr.s/esh has one singularity, pole at s D 1, we have that
1
s 1 esh D Res
1s1 e
shI 1s 1 C.s/;
for some entire .s/. This suggest choice
QP .s/ D Res1s1 e
shI 1s 1 D
eh
s 1;
for which .s/ is distributed-delay system
.s/ D .s/ D eh eshs 1 D e
hZ h0
e.s1/d 2 H1:
DTC question: first-order unstable Pr.s/
Likewise,
1
s a esh D Res
1sa e
shI as a C.s/ D
eah
s a eah eshs a ;
so that if Pr.s/ D 1sa , the choice
QP .s/ D eah
s agives us required
.s/ D eah
s a esh
s a D eah
Z h0
e.sa/d 2 H1
again.
First-order unstable Pr.s/: time-domain interpretation
Impulse responses of P.s/ Pr.s/esh D 1saesh and QP .s/ D 1saeah are
p.t/ D(0 it t < h
ea.th/ if t h and Qp.t/ D(0 it t < 0
ea.th/ if t 0
respectively. Thus,p.t/ Qp.t/; 8t h:
Impulse response of .s/ D QP .s/ Pr.s/esh then
.t/ D Qp.t/ p.t/ D(0 it t < 0 or t hea.th/ if 0 t < h
i.e., .s/ is FIR (finite impulse response).
Completion operator
Let G.s/ DA B
C 0
. Impulse response of Gh.s/ G.s/esh is
gh.t/ D(0 it t < h
C eA.th/B if t h
We are looking for rational QG.s/ such that Qg.t/ gh.t/, 8t h. Obviously,
Qg.t/ D(0 it t < 0
C eAheAtB if t 0 so thatQG.s/ D
A B
C eAh 0
:
LTI system
hG.s/esh
QG.s/ G.s/esh A BC eAh 0
A B
C 0
esh
completes, in a sense, G.s/esh to rational QG.s/, so we call it completion ofG.s/esh.
Completion operator (contd)
For any rational G.s/, hG.s/esh
is FIR with (bounded) impulse response
h.t/ D(C eA.th/B if 0 t < h0 otherwise
Hence, transfer function
hG.s/esh
D Z h0
h.t/estdt D
Z h0
C eA.th/Bestdt
is bounded in C0 and, as it also entire, belongs to H1. Thus,I h
G.s/esh
stable for every proper rational G.s/.
In time domain, y D hG.s/esh
u writes
y.t/ D CZ tth
eA.th/Bu./d D CZ h0
eA.h/Bu.t /d:
DTC question: general Pr.s/
Let
Pr.s/ DA B
C 0
:
We can always choose .s/ D hPr.s/esh
2 H1, for whichQP .s/ D Pr.s/esh C h
Pr.s/e
sh D A BC eAh 0
is indeed rational.
Modified Smith predictor
yu QC.s/
hPr.s/esh
Transfer function
C.s/ D QC.s/I hPr.s/esh QC.s/1Then,I if Pr.s/ is strictly proper, internal loop is well-posed for every proper QC ,I C.s/ stabilizes Pr.s/es iff QC stabilizes QP .s/
Example 1
C.s/
reQeud
y 1s1e
sh2eh
eh eshs1
- -
This QC stabilizes QP .s/ D ehs1 and then
Tu.s/ D 2eh.s 1/s C 1 ; T .s/ D
2eh
s C 1 esh;
and
Td .s/ D 1s C 1
1C 2.1 e
.s1/h/s 1
esh
are all stable, as expected.
Outline
Fiagbedzi-Pearson reduction
Smith controller revised
Modified Smith predictor and dead-time compensation
Modified Smith predictor vs. observer-predictor
Coprime factorization over H1 and Youla parametrization
Observer-predictor revised
Let Pr.s/ DA B
C 0
and consider observer-predictor control law
Pxo.t/ D .AC LC/xo.t/C Bu.t h/ Ly.t/
u.t/ D FeAhxo.t/C
Z tth
eA.t/Bu./d
where F and L are matrices making ACBF and ACLC Hurwitz. Denote
.t/ eAhxo.t/CZ tth
eA.t/Bu./d
(which is an h-prediction of xo for y 0). Then
P.t/ D eAh.AC LC/xo.t/C Bu.t h/ Ly.t/C Bu.t/ eAhBu.t h/ A Z t
theA.t/Bu./d
D A.t/C Bu.t/ eAhLy.t/ Cxo.t/:
Observer-predictor revised (contd)
In other words, observer-predictor control law writes as(P.t/ D A.t/C Bu.t/ eAhLy.t/ Cxo.t/u.t/ D F.t/
Substituting xo.t/ D eAh.t/ Z tth
eA.th/Bu./d , we get P.t/ D .AC eAhLC eAh/.t/C Bu.t/ eAhLC
Z tth
eA.th/Bu./d eAhLy.t/u.t/ D F.t/
Observer-predictor revised (contd)
Thus, we end up with control lawP D .AC BF C eAhLC eAh/ eAhL
y C C
Z tth
eA.th/Bu./d
u D F
Now, noting that the last term above is output of hPr.s/esh
when u is
its input, control law above is actually
yu QC.s/
hPr.s/esh
for
QC.s/ DAC BF C eAhLC eAh eAhL
F 0
:
Connections
yu QC.s/
hPr.s/esh
This is clearly MSP with primary controller,
QC.s/ DAC BF C eAhLC eAh eAhL
F 0
;
which is observer-based controller for
QP .s/ D
A B
C eAh 0
(note that AC eAhLC eAh D eAh.AC LC/eAh is Hurwitz). Thus,I observer-predictor is MSP when primary controller QC is observer-based
controller for QP
Outline
Fiagbedzi-Pearson reduction
Smith controller revised
Modified Smith predictor and dead-time compensation
Modified Smith predictor vs. observer-predictor
Coprime factorization over H1 and Youla parametrization
Coprime factorization over H1
We say that transfer function P.s/ has (strongly) coprime factorization overH1 if there are transfer functions
M.s/;N.s/; QM.s/; QN.s/; X.s/; Y.s/; QX.s/; QY .s/ 2 H1
such thatP.s/ D N.s/M1.s/ D QM1.s/ QN.s/
and X.s/ Y.s/
QN.s/ QM.s/ M.s/ QY .s/N.s/ QX.s/
DI 0
0 I
:
Coprime factorization and stabilizability
wu
wyyu
P.s/
C.s/
TheoremThere is controller C.s/ internally stabilizing this system iff 3P.s/ has strongcoprime factorization over H1. In this case all stabilizing controllers can beparametrized as (Youla parametrization)
C.s/ D QY .s/CM.s/Q.s/ QX.s/CN.s/Q.s/1D X.s/CQ.s/ QN.s/1Y.s/CQ.s/ QM.s/
for some Q.s/ 2 H1 but otherwise arbitrary.
3Most nontrivial part here, only if, was proved by Malcolm C. Smith (1989).
Coprime factorization and stabilizability (contd)
wu
wyyu
ses
C.s/
This is example of system, thatI cannot be internally stabilized.
Plant can be (weakly) coprime factorized over H1, i.e.,
ses D ses
s C 1
1
s C 11
yet there is no strongly coprime factorization.
Coprime factorization for rational systems
Let P.s/ DA B
C D
with .A;B/ stabilizable and .C;A/ detectable. Then
X.s/ Y.s/
QN.s/ QM.s/D24AC LC B C LD LF I 0C D I
35and
M.s/ QY .s/N.s/ QX.s/
D24 AC BF B LF I 0C CDF D I
35 ;where F and L are any matrices such that AC BF and AC LC Hurwitz.
Reduction to rational factorization
Let P.s/ be (not necessarily rational) proper transfer function such that
P.s/ D Pa.s/ .s/
for some .s/ 2 H1 and rational Pa.s/ with coprime factorizationXa.s/ Ya.s/
QNa.s/ QMa.s/ Ma.s/ QYa.s/Na.s/ QXa.s/
DI 0
0 I
:
Were looking forI strongly coprime factorization of P.s/ in terms of that of Pa.s/.
Reduction to rational factorization (contd)
LemmaP.s/ D Pa.s/ .s/ has strongly coprime factorization
X.s/ Y.s/
QN.s/ QM.s/DXa.s/ Ya.s/
QNa.s/ QMa.s/
I 0
.s/ I
2 H1
and M.s/ QY .s/N.s/ QX.s/
D
I 0
.s/ I Ma.s/ QYa.s/Na.s/ QXa.s/
2 H1
Proof.By direct substitution.
Resulting stabilizing controllers
Youla parametrization in this case is
C D . QY CMQ/. QX CNQ/1D . QYa CMaQ/
QXa CNaQ . QYa CMaQ/1Hence,
C. QXa CNaQ/ C. QYa CMaQ/ D QYa CMaQor, equivalently,
C. QXa CNaQ/ D .I C C/. QYa CMaQ/:
Thus, denoting Ca . QYa CMaQ/. QXa CNaQ/1, we end up with
.I C C/1C D Ca C D Ca.I Ca/1 Dyu
Ca.s/
.s/
Hmm, looks familiar. . .
Dead-time systems
If
P.s/ D Pr.s/esh DA B
C 0
esh
we already know how to present it in form
P.s/ D QP .s/ hP.s/
D A BC eAh 0
h
A B
C 0
esh
Thus, any stabilizing controller for P.s/ is of the form
yu QC.s/
hPr.s/esh
where QC.s/ is stabilizing controller for QP .s/. Thus, we end up withI modified Smith predictor yet again.
Fiagbedzi-Pearson reductionSmith controller revisedModified Smith predictor and dead-time compensationModified Smith predictor vs. observer-predictorCoprime factorization over H and Youla parametrization