Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Outline

Time Delays in Control and Estimation (038806)


e

Fiagbedzi-Pearson reduction

sh

Smith controller revised

lecture no. 6

Modified Smith predictor and dead-time compensation

Leonid Mirkin

Modified Smith predictor vs. observer-predictor

Faculty of Mechanical Engineering


TechnionIIT

Coprime factorization over H 1 and Youla parametrization

Nontrivial complication

Yet another transformation

So far, we studied problems in which h has finite spectrum. Let now

Consider

h W x.t/
P
D A0 x.t / C Ah x.t

h/ C Bu.t /;

which might have


I

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).

x.t
Q / Qx.t / C

eA.t

 h/

QAh x. /d

t h

Q n
Q
Q
for some AQ 2 Rn
and Q 2 Rnn
. Then,
Z t
Q
Q
Px.t/
Q
Q
D Qx.t
P /CA
eA.t  h/ QAh x./d C e Ah QAh x.t / QAh x.t h/
t h
Z t
Q
Q
D QA0 x.t / C QBu.t / C AQ
eA.t  h/ QAh x. /d C e Ah QAh x.t /
t h

D AQx.t/
Q C QBu.t /

Q
AQ

QA0

Q
Ah


QAh x.t /:

If we can choose AQ satisfying the left characteristic matrix equation


Q D QA0 C e
AQ
h reduces to

Q
Ah

QAh ;

PQ / D AQx.t
Q W x.t
Q / C QBu.t /:

(l.c.m.e.)

Q D QA0 C e
Properties of solutions of l.c.m.e. AQ

Q
Ah

QAh

Q and Q be the corresponding left eigenvector. Then


Let Q 2 spec.A/
Q D QA
Q Q
Q D Q AQ
Q
Qe
0C
Q
In other words, Q.
Q I
I

Q
Ah

Q
h

QAh D Q.A
Q
0Ce

Suppose we can solve (l.c.m.e.) in AQ and Q. Then



Z
u.t / D FQ x.t
Q / D FQ Qx.t / C

Ah /:

Q
h

Ah / D 0, so that
whenever  Q
Q 0, every eigenvalue of AQ is a pole of h

A0

(remember, those poles are the roots of h .s/ D det.sI

A0

Closed-loop spectrum

Ah e

Q
A.t
 h/

QAh x. /d

t h

for some FQ leads to the closed-loop system


sh

/).

FQ Q

sI
FQ e

A0 Ah e
Q Rh
Ah
.sI
0 e

sh
Q
A/

dQAh

B
I






X.s/
D I.C. :
U.s/

and the characteristic quasi-polynomial cl .s/ D det .s/, where


.s/

Closed-loop spectrum (contd)

.sI

Because
Q
Ah

Q
.sI A/

d QAh

0
Ah

sh

I /QAh D Q.sI

Ah e

A0

sh

/;

Q
Ah

Q
.sI A/

.s/ D

d QAh D .sI

FQ .sI

Q 1 Q.sI
A/

sI A0 Ah e sh
Q 1 Q.sI A0 Ah e
A/

Ah e

A0

sh

B
I

sh

/:

D det I


B
:
I

A0 Ah e sh / B FQ .sI

Q 1 Q det.sI
B FQ .sI A/

Q 1 Q.sI
A/
A0

Ah e

A0

cl .s/ D

In other words,

sh

/:

Ah e

sh

 

AQ Q
AQ C QB FQ Q
D
B FQ I
B FQ
I

FQ .sI

Q

Q 1 QB D det.sI A
A/
det.sI

QB FQ /
Q
A/

and thus

and
cl .s/ D det .sI

A0 Ah e sh

Q Rh
Q
Ah
.sI A/
dQAh
0 e

we have that

det I

QCe

Then

Q C .e
AQ

D sQ

so that

sI
FQ Q C e

Closed-loop spectrum (contd)

Next,

Q
A/ Q C e

det.sI A0
det.sI

Ah e
Q
A/

sh

det.sI

AQ


S
Q
spec.h;cl / D spec.h / n spec.A/
spec.Q cl /
Q  spec.h /).
(remember, spec.A/
I

QB FQ /:

Implications

Distributed state / input delays

Thus, if we can

Let

I
I

solve (l.c.m.e.) so that AQ contains all unstable modes of h ,


Q QB/),
find FQ so that AQ C QB FQ is Hurwitz (requires stabilizability of .A;

h W x.t
P /D

Then transformation

the control law


Z
u.t/ D FQ Qx.t / C

eA.t

 h/

QAh x. /d

t h

Q to the
stabilizes h by moving all its unstable modesthose in spec.A/
Q
Q
eigenvalues of A C QB F and keeping the other modes of h untouched.

x.t/
Q
Qx.t / C

t
t h

0
h


. /x.t C  / C . /u.t C  / d :

 t

eA.t
h

with l.c.m.e.
Q D
AQ

and
I

Q D
Not quite, solving AQ
I

(l.c.m.e.0 )

eA Q./d
h

where BQ

eA Q./d
h


S
Q
spec.h;cl / D spec.h / n spec.A/
spec.Q cl /.

Example
0

eA Q. /d is highly nontrivial. Specifically,

Let
h W x.t
P /D

we have to find all troublesome modes of h

x.t / C x.t

h/ C u.t /;

whole characteristic quasi-polynomial is (here s D  C j! )

(in most cases, have to rely on numerical approaches)


I


Q ./x. / C ./u./ d d

yields reduced system


PQ / D AQxQ C Bu.t
Q /;
Q W x.t

Is it that simple ?

C/

solve (l.c.m.e.) / (l.c.m.e.0 )

h .s/ D s C 1

(solution is non-unique and not especially elegant)

sh

D  C 1 C j!

h

j!h

Solutions of h .s/ D 0 must satisfy the magnitude condition


Only a handful of cases where the steps above can be solved analytically.
One example is
2


60
6
6
./ D 6 :::
6
40

 
 
:: : :
:
:
0 
0 0 

3
2

0
60
7
7
6
:: 7 . / C X 6 ::
7
6:
:7
6
i
40
 5
0 
0


::
:

 
0 
:: : :
:
:
0 
0 

3

7
7
:: 7 . C h /
i
:7
7
0 5
0 0



::
:

in which case spec.h / is finite and (l.c.m.e.0 ) is solvable with Q D I .

. C 1/2 C ! 2 D e

2h

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,
h .s/
1
D 1 C lim
s!0
s!0
s

lim

e
s

sh

which implies that s D 0 is a single root.

h
D 1 C h;
s!0 1

D 1 C lim

Example (contd)

Outline

Thus, we have only one unstable pole to shift and may pick nQ D 1, AQ D 0.
Eqn. (l.c.m.e.) then reads 0 D q C q , so we may pick q D 1. Then

Fiagbedzi-Pearson reduction

PQ / D u.t/;
Q W x.t

Smith controller revised

which is stabilized by u.t/ D k x.t


Q / for any k > 0. Thus
u.t/ D


Z
k x.t / C

x./d
t h

Modified Smith predictor and dead-time compensation

stabilizes h and renders its closed-loop characteristic polynomial


sCk
h;cl .s/ D
.s C 1
s

Modified Smith predictor vs. observer-predictor


sh

/:

Coprime factorization over H 1 and Youla parametrization

In fact, the controller above has the transfer function


C.s/ D


1
k 1C

e
s

sh 

2 H 1:

Smith controller: preliminary conclusions

More rigorous analysis

d
y

Pr .s/e

sh

wu
u

eQ

CQ .s/
Pr .s/.1

sh

Pr .s/e

sh

C.s/

wy

C.s/

Remember, Smith controller


I

works if Pr .s/ is stable


(stabilization problem reduces then to that for delay-free plant)

System is said to be internally stable if


w   
I transfer matrix from wy to y ,
u
u
Tcl

does not necessarily work if Pr .s/ is unstable

1
Pr C e

sh




1 Pr e sh
S Td

2 H 1:
C Pr C e sh
Tu T

(might lead to unstable loop)

Lets
I

analyze Smith controller from internal stability perspectives.

Loop shifting

Loop shifting: signal transformations


wu

Pr .s/

Pr .s/e

Pr .s/.1

Pr .s/.1

CQ .s/

sh

sh

Pr .s/

Pr .s/e
Pr .s/.1

sh

Pr .s/.1

sh

sh

CQ .s/

C.s/

wy

CQ .s/

/ we

Pr .s/.1

We end up with a new loop with the plant Pr and the controller

C
CQ
Q
C
so
that
C
D
1 C CPr .1 e sh /
1 CQ Pr .1 e

sh /

Pr .s/.1

CQ .s/

Pr .s/
CQ .s/

uQ

wy

yQ

wQ y

1. if Pr .1
2. if Pr .1

To complete1 the picture, we have to calculate them:


  
 
yQ
y C Pr .1 e sh /u
I Pr .1 e
D
D
uQ
u
0
I

sh

 
y
u

1 Well,

wQ y
wy
D
wQ u

sh

./wu

C.s/

wy

sh

C.s/

sh

wQ y

Pr .1

sh


I
Q
Tcl
0

Pr .1
I

sh

/ 2 H 1 , then TQcl 2 H 1 implies Tcl 2 H 1


e sh / 62 H 1 , then TQcl 2 H 1 not necessarily implies Tcl 2 H 1

sh

Key question:
I

Pr .1 e
wu

sh

/wu

I
0

Pr .1

when does Pr .1

sh

/ 2 H1 ?

Obviously, this is true if Pr 2 H 1 . Yet this also true if

and


Pr .s/.1

(in fact, never; this can be verified by making use of explicit form of TQcl )

different signals.

I
Tcl D
0

The new system is delay-free yet with


I

Pr .s/

yQ

Thus,

wQ u

sh

C.s/

sh

Loop shifting (contd)


,

sh

wu

Loop shifting (contd)


Pr .s/e

Pr .s/.1

wy

redistribute loop components w/o changing the whole system.

wu

sh

wu

Pr .s/e

Adding and subtracting block Pr .1

Pr .s/

Pr .s/e

sh

C.s/

wu

sh

sh

we also have to guarantee that internal CQ loop is well posed.



wy
:
wu

Pr .s/ proper and its only unstable poles are single poles at j 2
k , k 2 Z,
h

which are simple zeros of 1

sh

Outline

Loop shifting: idea


wu

Fiagbedzi-Pearson reduction

Pr .s/

Pr .s/e

sh

Pr .s/.1

sh

Pr .s/.1

sh

Smith controller revised

Modified Smith predictor and dead-time compensation

CQ .s/

Modified Smith predictor vs. observer-predictor

C.s/

wy

Driving idea here is to


I

cancel (compensate) delay via Pr .s/e

sh

Coprime factorization over H 1 and Youla parametrization

C Pr .s/.1

sh

/ D Pr .s/.

The question:
I

Dead-time compensation question


wu

Pr .s/e

Dead-time compensation: aspirations


If we obtained required .s/ 2 H 1 , we would transform

PQ .s/

sh

can we do it with stable compensation element ?

wu

Pr .s/e

wQ u

sh

.s/
u

.s/

CQ .s/

C.s/

Pr .s/e

for some proper and rational PQ .s/.

C .s/ D PQ .s/

C.s/

uQ

wy

CQ .s/

yQ

wQ y

for rational (delay-free) PQ .s/. Resulting C.s/,


wy

Technicaly speaking, we are looking for .s/ 2 H 1 such that


sh

PQ .s/

C.s/ D CQ .s/ I


.s/CQ .s/

CQ .s/

.s/

called dead-time compensator (DTC) and


Q .s/ internally stabilizes PQ .s/ iff C.s/ internally stabilizes Pr .s/e
I C
Tcl .s/ D

I
0

.s/ Q
I
Tcl .s/
I
0

(provided DTC loop well posed).

.s/
I

sh

as

DTC question: stable Pr .s/

DTC question: unstable Pr .s/ in Example of Lect 5

Choice of .s/ apparent and non-unique:


I .s/ D PQ .s/
Pr .s/e sh for any PQ .s/ 2 H 1 does the job.

With Pr .s/ D 1=s , the Smith predictor


sh

s

d 2 H 1

indeed (integrand analytic and bounded in C0 and integration path finite).


In this case
1
PQ .s/ D Pr .s/ D
s

Some standard choices:


I PQ .s/ D Pr .s/ results in Smith predictor
I

.s/ D

e
s

can be interpreted as unstable rational part of partial fraction expansion2 of

PQ .s/ D 0 results in internal model controller (IMC)

Pr .s/e

sh

Res

sh


I0

C .s/

for some entire .s/ (Pr .s/ has only one pole, that at the origin).

2 Here

DTC question: another unstable Pr .s/


Let Pr .s/ D 1=.s

As Pr .s/e

sh

1
s

sh

Likewise,
1

2 H 1:

has one singularity, pole at s D 1, we have that



1
sh
e
I
1
Res
1
s 1
e sh D
C .s/;
s 1
s 1

for some entire .s/. This suggest choice

e
1

1
,
s a


e sh I a
e ah
C .s/ D
s a
s a

1
s a

ah

the choice

e
PQ .s/ D
s

e ah
.s/ D
s a
again.

sh

De

Res

ah

gives us required

for which .s/ is distributed-delay system


.s/ D

sh

so that if Pr .s/ D


Res s 1 1 e sh I 1
e h
Q
P .s/ D
D
;
s 1
s 1

.s/ D

c/f .s/ stands for residue of f .s/ at c .

DTC question: first-order unstable Pr .s/

1/. Were looking for stable PQ .s/ such that


.s/ D PQ .s/

Res.f .s/I c/ lims!c .s

.s 1/

d 2 H 1 :

e sh
De
s a

ah

e
0

.s a/

d 2 H 1

e
a

sh

First-order unstable Pr .s/: time-domain interpretation


sh

Impulse responses of P .s/ Pr .s/e


p.t/ D

it t < h

ea.t

h/

and PQ .s/ D

p.t
Q /D

and

if t  h

sh

e
a

(
0

1
s

e
a

ah

Completion operator
are

Impulse response of .s/ D PQ .s/

ea.t

h/

.t/ D p.t/
Q

p.t / D

ea.t

sh

h/

A B
C 0

. Impulse response of Gh .s/ G.s/e


gh .t / D

if t  0

then
it t < 0 or t  h

 h G.s/e

sh

Q
G.s/

Completion operator (contd)

h .t / D

h/

if 0  t < h
otherwise

Hence, transfer function

 h G.s/e


sh

h .t /e
0

y.t/ D C

eA.t
t h

 h/

sh

sh

sh

if t  h

A
B

C e Ah 0


A B
e
C 0

st

dt D


u writes

Bu. /d D C

C eA.t

h/

Be

st

dt

eA.
0


A B
:
Pr .s/ D
C 0


We can always choose .s/ D  h Pr .s/e sh 2 H 1 , for which


h/

Bu.t

/d:

is indeed rational.

sh

sh

Q , so we call it completion of
to rational G.s/

Let

PQ .s/ D Pr .s/e

is bounded in C0 and, as it also entire, belongs to H 1 . Thus,


I  h G.s/e sh stable for every proper rational G.s/.

In time domain, y D  h G.s/e

Ce

DTC question: general Pr .s/

is FIR with (bounded) impulse response

(
C eA.t

it t < h
A.t h/

G.s/e

completes, in a sense, G.s/e


G.s/e sh .

sh

is

LTI system

if 0  t < h

i.e., .s/ is FIR (finite impulse response).

For any rational G.s/,  h G.s/e

(
0

sh

Q
We are looking for rational G.s/
such that g.t
Q /  gh .t /, 8t  h. Obviously,
(


0
it t < 0
A
B
Q
:
g.t/
Q
D
so that G.s/ D
C e Ah 0
C e Ah eAt B if t  0

8t  h:

Pr .s/e
(
0

it t < 0

respectively. Thus,
p.t/  p.t/;
Q

Let G.s/ D

C  h Pr .s/e

sh

A
B
D
Ah
Ce
0

Modified Smith predictor


u

Example 1
y

CQ .s/

 h Pr .s/e

sh

d
y

1
s 1

sh

eQ

2eh
e

Then,
I
I

 h Pr .s/e

sh

CQ .s/

if Pr .s/ is strictly proper, internal loop is well-posed for every proper CQ ,


C.s/ stabilizes Pr .s/e s iff CQ stabilizes PQ .s/

This CQ stabilizes PQ .s/ D


Tu .s/ D

and

s 1

sh

s 1

C.s/

Transfer function
C.s/ D CQ .s/ I

and then

2eh .s 1/
;
sC1

2eh
e
sC1

T .s/ D


1
2.1 e .s
Td .s/ D
1C
sC1
s 1

1/h

sh

sh

are all stable, as expected.

Outline

Observer-predictor revised
Let Pr .s/ D

Fiagbedzi-Pearson reduction

Smith controller revised

Modified Smith predictor and dead-time compensation

A B
C 0

and consider observer-predictor control law

xP .t / D .A C LC /x .t / C Bu.t
o

Coprime factorization over H 1 and Youla parametrization

eA.t
t h

Ly.t /

/
Bu. /d

where F and L are matrices making A C BF and A C LC Hurwitz. Denote


.t / eAh xo .t / C

Modified Smith predictor vs. observer-predictor

h/


Z
u.t / D F eAh xo .t / C
Z

eA.t

/

Bu./d

t h

(which is an h-prediction of xo for y  0). Then


.t/
P D eAh .A C LC /xo .t / C Bu.t
C Bu.t /
D A.t / C Bu.t /

eAh Bu.t
eAh L y.t /


h/ Ly.t /
Z t

h/
A
e
t h

C xo .t / :

A.t  /

Bu. /d

Observer-predictor revised (contd)

Observer-predictor revised (contd)

In other words, observer-predictor control law writes as


(

.t/
P D A.t / C Bu.t / eAh L y.t / C xo .t/

Thus, we end up with control law

P D .A C BF C e

u.t/ D F .t /

Substituting xo .t/ D e

Ah

.t /

.t/
P D .A C e

eA.t

h /

Bu./d , we get

Ah

LC e Ah /.t / C Bu.t /
Z t
Ah
e LC
eA.t h / Bu./d

LC e

e L yCC

A.t h /

t h

u.t/ D F .t/

for

CQ .s/

 h Pr .s/e

t h

sh


A C BF C eAh LC e
Q
C .s/ D
F

Ah


eAh L
:
0

Outline
u

CQ .s/

 h Pr .s/e

sh

Fiagbedzi-Pearson reduction

Smith controller revised

This is clearly MSP with primary controller,



A C BF C eAh LC e
Q
C .s/ D
F

Ah


eAh L
;
0

which is observer-based controller for




A
B
PQ .s/ D
C e Ah 0
(note that A C eAh LC e
I

/

Ah

Now, noting that the last term above is output of  h Pr .s/e


its input, control law above is actually

eAh Ly.t/

Connections

Ah

u D F

t
t h

Ah

Ah

D eAh .A C LC /e

Ah

is Hurwitz). Thus,
observer-predictor is MSP when primary controller CQ is observer-based
controller for PQ

Modified Smith predictor and dead-time compensation

Modified Smith predictor vs. observer-predictor


Coprime factorization over H 1 and Youla parametrization

sh

Bu. /d

when u is

Coprime factorization over H 1

Coprime factorization and stabilizability

We say that transfer function P .s/ has (strongly) coprime factorization over
H 1 if there are transfer functions

wu

P .s/

M.s/; N.s/; MQ .s/; NQ .s/; X.s/; Y.s/; XQ .s/; YQ .s/ 2 H 1

such that
P .s/ D N.s/M

and

X.s/ Y .s/
NQ .s/ MQ .s/



.s/ D MQ

.s/NQ .s/

 

M.s/ YQ .s/
I 0
D
:
Q
0 I
N.s/ X.s/

C.s/

wy

Theorem
There is controller C.s/ internally stabilizing this system iff 3 P .s/ has strong
coprime factorization over H1. In this case all stabilizing controllers can be
parametrized as (Youla parametrization)

 1
Q
YQ .s/ C M.s/Q.s/ X.s/
C N.s/Q.s/
 1

Y .s/ C Q.s/MQ .s/
D X.s/ C Q.s/NQ .s/

C.s/ D

for some Q.s/ 2 H 1 but otherwise arbitrary.


3 Most

Coprime factorization and stabilizability (contd)


wu

se

Plant can be (weakly) coprime factorized over H 1 , i.e.,


se

se s
D
sC1

A B
C D

with .A; B/ stabilizable and .C; A/ detectable. Then


2

A C LC B C LD
X.s/ Y .s/
4
D
F
I
NQ .s/ MQ .s/
C
D


3
L
0 5
I

and

cannot be internally stabilized.

wy

This is example of system, that


I

Coprime factorization for rational systems


Let P .s/ D

C.s/

nontrivial part here, only if, was proved by Malcolm C. Smith (1989).

1
sC1

yet there is no strongly coprime factorization.

2

A C BF B
Q
M.s/ Y .s/
4
D
F
I
Q
N.s/ X.s/
C C DF D

3
L
0 5;
I

where F and L are any matrices such that A C BF and A C LC Hurwitz.

Reduction to rational factorization

Reduction to rational factorization (contd)

Let P .s/ be (not necessarily rational) proper transfer function such that

Lemma

P .s/ D Pa .s/

P .s/ D Pa .s/

.s/

for some .s/ 2 H 1 and rational Pa .s/ with coprime factorization




Xa .s/ Ya .s/
NQ a .s/ MQ a .s/



 

Ma .s/ YQa .s/
I 0
D
:
0 I
Na .s/ XQ a .s/

 


X.s/ Y .s/
Xa .s/ Ya .s/
I 0
D
2 H1
NQ .s/ MQ .s/
NQ a .s/ MQ a .s/ .s/ I

and


Were looking for


I

.s/ has strongly coprime factorization

strongly coprime factorization of P .s/ in terms of that of Pa .s/.

 


M.s/ YQ .s/
I
0
Ma .s/ YQa .s/
D
2 H1
.s/ I
N.s/ XQ .s/
Na .s/ XQ a .s/

Proof.
By direct substitution.

Resulting stabilizing controllers

Dead-time systems

Youla parametrization in this case is

If

C D . YQ C MQ/.XQ C NQ/

Hence,

C.XQ a C Na Q/

P .s/ D Pr .s/e

D . YQa C Ma Q/ XQ a C Na Q

. YQa C Ma Q/

C. YQa C Ma Q/ D

sh


A B
e
D
C 0

sh

we already know how to present it in form


P .s/ D PQ .s/

YQa C Ma Q

or, equivalently,

h


P .s/ D

A
B
Ah
Ce
0

h




A B
e
C 0

sh

Thus, any stabilizing controller for P .s/ is of the form


C.XQ a C Na Q/ D .I C C/. YQa C Ma Q/:

Thus, denoting Ca . YQa C Ma Q/.XQ a C Na Q/ 1 , we end up with


u

.I C C/ 1 C D Ca

C D Ca .I

Ca /

Ca .s/
.s/

 h Pr .s/e

sh

where CQ .s/ is stabilizing controller for PQ .s/. Thus, we end up with


I

Hmm, looks familiar. . .

CQ .s/

modified Smith predictor yet again.

You might also like