Lecture 07 EAILC

Lecture 7: Deterministic Self-Tuning Regulators
Feedback Control Design for Nominal Plant Model
via Pole Placement

Indirect Self-Tuning Regulators Direct Self-Tuning Regulators
c Leonid Freidovich.
April 22, 2010.
Elements of Iterative Learning and Adaptive Control:
Lecture 7
p. 1/20
Indirect Self-Tuning Regulators

Consider a single input single output (SISO) system
y (t) + a1 y (t 1) + + an y (t n) = b0 u(t d0 ) + + + bm u(t d0 m)
April 22, 2010.
Lecture 7
p. 2/20

It can be re written in the regression form

y (t) = a1 y (t 1) an y (t n) + b0 u(t d0 ) + + b1 u(t d0 1) + + bm u(t d0 m) = (t 1)T 0
April 22, 2010.
Lecture 7
p. 2/20

It can be re written in the regression form

y (t) = a1 y (t 1) an y (t n) + b0 u(t d0 ) + + b1 u(t d0 1) + + bm u(t d0 m) = (t 1)T 0
where
= a1 , . . . , an , b0 , . . . , bm
(t 1) =
y (t 1), . . . , y (t n), u(t d0 ), . . . , u(t d0 m)

T
April 22, 2010.
Lecture 7
p. 2/20
RLS Algorithm (discrete time)

Given the data y (t), (t 1)
y (t) = (t 1)T 0 ,
N t0
dened by the model
t = t 0 , t 0 + 1, . . . , N
April 22, 2010.
Lecture 7
p. 3/20
RLS Algorithm (discrete time)

Given the data y (t), (t 1)
y (t) = (t 1)T 0 ,
N t0
dened by the model
t = t 0 , t 0 + 1, . . . , N
(t0 ) and a measure of our trust in this With the initial guess guess: P (t0 ) > 0, we can use the RLS algorithm: (t) = (t 1) + K (t) y (t) (t 1)T (t 1) K (t) = P (t)(t 1) = 1 P (t 1) (t 1) + (t 1)T P (t 1)(t 1) P (t 1) (t 1) (t 1)T P (t 1) + (t 1)T P (t 1) (t 1)
Lecture 7
p. 3/20
P (t ) =
P (t 1)
April 22, 2010.
RLS Algorithm (continuous time)

Given the data y ( ), ( )
t t= t0
dened by the model

[ t0 , t]
y ( ) = ( )T 0 ,
obtained from A(p) y ( ) = B (p) u( ) using ltered signals

yf ( ) = Hf (p) y ( ) and uf ( ) = Hf (p) u( ) as follows y ( ) = pn yf ( ), ( )T = pn1 yf ( ), . . . , pm1 uf ( )
April 22, 2010.
Lecture 7
p. 4/20
RLS Algorithm (continuous time)

Given the data y ( ), ( )
t t= t0
dened by the model

[ t0 , t]
y ( ) = ( )T 0 ,
obtained from A(p) y ( ) = B (p) u( ) using ltered signals

yf ( ) = Hf (p) y ( ) and uf ( ) = Hf (p) u( ) as follows y ( ) = pn yf ( ), ( )T = pn1 yf ( ), . . . , pm1 uf ( )
(t0 ) and a measure of our trust in this With the initial guess guess: P (t0 ) > 0, we can use the RLS algorithm: (t) (t ) = P (t ) (t ) y (t) (t)T (t ) = P (t ) P (t ) (t ) (t )T P (t ) P
c Leonid Freidovich. April 22, 2010. Elements of Iterative Learning and Adaptive Control: Lecture 7
p. 4/20
Algorithm using RLS and MD Pole Placement

Off-line Parameters: Given polynomials Bm , Am , Ao
April 22, 2010.
Lecture 7
p. 5/20

Off-line Parameters: Given polynomials Bm , Am , Ao Step 1: Estimate the coefcients of A and B = B + B , i.e. 0 , using the Recursive Least Squares algorithm.
April 22, 2010.
Lecture 7
p. 5/20

Off-line Parameters: Given polynomials Bm , Am , Ao Step 1: Estimate the coefcients of A and B = B + B , i.e. 0 , using the Recursive Least Squares algorithm. Step 2: Apply the Minimum Degree Pole Placement algorithm
(deg A = deg Am , deg B = deg Bm , deg Ao = deg Adeg B +1, Bm = B Bpm )
R = Rp B + ,
T = Ao Bpm ,
S, Rp : A Rp +B S = Ao Am
with estimates for A and B taken from the previous Step.
April 22, 2010.
Lecture 7
p. 5/20

R = Rp B + ,
T = Ao Bpm ,
with estimates for A and B taken from the previous Step. Step 3: Compute the feedback control law
R u(t) = T uc (t) S y (t)
April 22, 2010.
Lecture 7
p. 5/20

R = Rp B + ,
T = Ao Bpm ,
with estimates for A and B taken from the previous Step. Step 3: Compute the feedback control law
R u(t) = T uc (t) S y (t)
Repeat Steps 1, 2, 3 (until the performance is satisfactory).

p. 5/20
Example 3.4 (Recursive Modication of Example 3.1)

Given a continuous time system
y +y = u,
April 22, 2010.
Lecture 7
p. 6/20

y +y = u,
its pulse transfer function with sampling period h = 0.5 sec is

G (q ) = b1 q + b2 q2 + a1 q + a2 = 0.1065q + 0.0902 q 2 1.6065q + 0.6065
April 22, 2010.
Lecture 7
p. 6/20

y +y = u,
its pulse transfer function with sampling period h = 0.5 sec is

G (q ) = b1 q + b2 q2 + a1 q + a2 = 0.1065q + 0.0902 q 2 1.6065q + 0.6065
The task is to synthesize a 2-degree-of-freedom controller such that the complementary sensitivity transfer function is
B m (q ) A m (q ) = 0.1761q q 2 1.3205q 0.4966
We will take
uc (t) = 1
uc ( t ) = 1
for
t0
except for for

50 t 70
Lecture 7
p. 6/20
for
15 t < 30,
uc (t) = 0
April 22, 2010.

We have found the controller via MD pole placement
R(q )u(t) = T (q )uc (t) S (q )y (t)
with
R(q ) = q + 0.8467 S (q ) = 2.6852 q 1.0321 T (q ) = 1.6531 q
that stabilizes the discrete plant. Will it stabilize the continuous-time system as well?
April 22, 2010.
Lecture 7
p. 7/20

We have found the controller via MD pole placement
R(q )u(t) = T (q )uc (t) S (q )y (t)
with
R(q ) = q + 0.8467 S (q ) = 2.6852 q 1.0321 T (q ) = 1.6531 q
that stabilizes the discrete plant. Will it stabilize the continuous-time system as well? Let us design the adaptive controller which would recover the performance for the case when parameters of the plant - the polynomials A(q ) and B (q ) are not known!
p. 7/20
1.5
0.5 output
0.5
1.5
10
20
30
40 time
50
60
70
80
90
The response of the closed-loop system with adaptive controller

p. 8/20
1 control
10
20
30
40 time
50
60
70
80
90
The control signal

p. 9/20
1 0 a
1
1 2 1 0.5 a
2
10
15
20
25
30
35
40
0 0.5 0.25 0.2 b1 0.15 0.1 0.4

2
10
15
20
25
30
35
40
10
15
20
25
30
35
40
0.2 0
10
15
20 time (sec)
25
30
35
40
Adaptation of parameters dening polynomials A(q ) and B (q )

p. 10/20
Lecture 6: Deterministic Self-Tuning Regulators
Feedback Control Design for Nominal Plant Model
via Pole Placement

Indirect Self-Tuning Regulators Direct Self-Tuning Regulators
April 22, 2010.
Lecture 7
p. 11/20
Direct Self-Tuning Regulators

The previous algorithm consists of two steps
Step 1:
Computing estimates for the polynomials A(q ), B (q )
April 22, 2010.
Lecture 7
p. 12/20

The previous algorithm consists of
Step 1: Step 2:
(q ), B (q ) for A(q ), B (q ) Computing estimates A
Computing polynomials dening the controller

T (q ), S (q )
R (q ), (q ), B (q ). based on A
April 22, 2010.
Lecture 7
p. 12/20

Step 1: Step 2:

T (q ), S (q )
Such procedure has potential drawbacks like

Redundant step in design: why to know A(q ) and B (q )?
April 22, 2010.
Lecture 7
p. 12/20

Step 1: Step 2:

T (q ), S (q )
Such procedure has potential drawbacks like

Redundant step in design: why to know A(q ) and B (q )? Control design methods (pole placement, LQR, . . . ) are
(q ), B (q ). Is there a way to avoid sensitive to mistakes in A such poorly conditioned numerical computations?
April 22, 2010.
Lecture 7
p. 12/20

Step 1: Step 2:

T (q ), S (q )
The procedure has potential drawbacks like

Redundant step in design: why to know A(q ) and B (q )? Control design methods (pole placement, LQR, . . . ) are
(q ), B (q ). Is there a way to avoid sensitive to mistakes in A such poorly conditioned numerical computations? E.g.: (q ) = q (q 1), A (q ) = q + B
April 22, 2010.
Lecture 7
p. 12/20

The idea of another approach comes from the observation that for the plant
A(q ) y (t) = B + (q ) B (q ) u(t)
and for the target system (the model to follow)

Am (q ) y (t) = Bmp (q ) B (q ) uc (t)
the Diophantine equation

A o (q ) A m (q ) = A (q ) R p (q ) + B (q ) S (q )
can be seen as a regression to estimate coefcients of R and S
April 22, 2010.
Lecture 7
p. 13/20

A(q ) y (t) = B + (q ) B (q ) u(t)

Am (q ) y (t) = Bmp (q ) B (q ) uc (t)


A o (q )A m (q )y (t ) = A (q )R p (q ) + B (q )S (q ) y (t )
April 22, 2010.
Lecture 7
p. 13/20

A(q ) y (t) = B + (q ) B (q ) u(t)

Am (q ) y (t) = Bmp (q ) B (q ) uc (t)


A o (q )A m (q ) y (t ) = A (q )R p (q )y (t ) + B (q )S (q ) y (t )
April 22, 2010.
Lecture 7
p. 13/20

A(q ) y (t) = B + (q ) B (q ) u(t)

Am (q ) y (t) = Bmp (q ) B (q ) uc (t)


A o (q )A m (q ) y (t ) = R p (q ) A (q )y (t ) + B (q )S (q ) y (t )
April 22, 2010.
Lecture 7
p. 13/20

A(q ) y (t) = B + (q ) B (q ) u(t)

Am (q ) y (t) = Bmp (q ) B (q ) uc (t)


Ao (q )Am (q )y (t) = Rp (q )B (q )u(t) + B (q )S (q )y (t)
April 22, 2010.
Lecture 7
p. 13/20

A(q ) y (t) = B + (q ) B (q ) u(t)

Am (q ) y (t) = Bmp (q ) B (q ) uc (t)


Ao (q )Am (q )y (t) = Rp (q )B + (q )B (q )u(t)+B (q )S (q )y (t)
April 22, 2010.
Lecture 7
p. 13/20

A(q ) y (t) = B + (q ) B (q ) u(t)

Am (q ) y (t) = Bmp (q ) B (q ) uc (t)


Ao (q )Am (q )y (t) = B (q ) R(q ) u(t) + S (q )y (t)
April 22, 2010.
Lecture 7
p. 13/20
Direct Self-Tuning for Minimum-Phase Systems

Suppose that B (q ) is stable and can be canceled, i.e.
A(q )y (t) = B (q )u(t) = b0 u(t d0 ) + b1 u(t d0 1) + + bm u(t d0 m) = B + (q )B (q )u(t) = B + (q )b0 u(t)
April 22, 2010.
Lecture 7
p. 14/20

then Diophantine equation

A o (q )A m (q ) = A (q )R p (q ) + B (q )S (q )
April 22, 2010.
Lecture 7
p. 14/20


leads to the regression

Ao (q )Am (q )y (t) = b0 Rp (q )u(t) + S (q )y (t)
April 22, 2010.
Lecture 7
p. 14/20


leads to the regression

Ao (q )Am (q )y (t) = b0 Rp (q )u(t) + S (q )y (t)
which is in the form
(t ) = (t )T
with
A o (q )A m (q ) y (t ) = (t ),
c Leonid Freidovich. April 22, 2010.
(t)T = b0 R(q )u(t) + S (q )y (t)

Lecture 7
p. 14/20

In the formula
(t) = b0 R(q )u(t) + b0 S (q )y (t) = (t)T
the vector of regressors is

(t) = [u(t), . . . , u(t l), y (t), . . . , y (t l)]T
and the vector of parameters is

= [b 0 r 0 , . . . , b 0 r l , b 0 s0 , . . . , b 0 sl ]T
April 22, 2010.
Lecture 7
p. 15/20

In the formula
(t) = b0 R(q )u(t) + b0 S (q )y (t) = (t)T
the vector of regressors is

(t) = [u(t), . . . , u(t l), y (t), . . . , y (t l)]T
and the vector of parameters is

= [b 0 r 0 , . . . , b 0 r l , b 0 s0 , . . . , b 0 sl ]T , then Suppose is estimated by 2 q l 1 + . . . , R (q ) = q + 1
l
c Leonid Freidovich. April 22, 2010.
l+2 q l 1 + . . . S (q ) = q + 1
l
Lecture 7
p. 15/20

Some modication to the regression model
Ao (q )Am (q )y (t) = (t) = b0 R(q )u(t)+b0 S (q )y (t) = (t)T
can be made if we introduce the signals

uf (t) = 1 A o (q )A m (q ) u(t), y f (t ) = 1 A o (q )A m (q ) y (t )
April 22, 2010.
Lecture 7
p. 16/20

Some modication to the regression model
Ao (q )Am (q )y (t) = (t) = b0 R(q )u(t)+b0 S (q )y (t) = (t)T
can be made if we introduce the signals

uf (t) = 1 A o (q )A m (q ) u(t), y f (t ) = 1 A o (q )A m (q ) y (t )
Then the regression model is

y ( t ) = b 0 R ( q ) uf ( t ) + b 0 S ( q ) y f ( t ) = R(q )uf (t) + S (q )yf (t) = (t )T
April 22, 2010.
Lecture 7
p. 16/20
Direct Self-Tuning Regulators: Summary

Given
polynomials Am (q ), Bm (q ), Ao (q ) the relative degree d0 of the plant (deg Ao = d0 1)
April 22, 2010.
Lecture 7
p. 17/20

Given
Step 1: Estimate coefcients of polynomials R(q ) and S (q ) from the regression model
y ( t ) = b 0 R ( q ) uf ( t ) + b 0 S ( q ) y f ( t ) = ( t ) T
April 22, 2010.
Lecture 7
p. 17/20

Given
Step 2: Compute the control signal by

R(q )u(t) = T (q )uc (t) S (q )y (t)
where for Bm (q ) = q d0 Am (1) (minimal delay and unit static gain):

T (q ) = Ao (q )Am (1)
April 22, 2010.
Lecture 7
p. 17/20

Given
Step 2: Compute the control signal by

R(q )u(t) = T (q )uc (t) S (q )y (t)
where for Bm (q ) = q d0 Am (1) (minimal delay and unit static gain):

T (q ) = Ao (q )Am (1)
Repeat Steps 1 and 2

p. 17/20
Direct ST Reg. for Non-Minimum-Phase Systems

If B (q ) is not stable or cannot be canceled, i.e.
A(q ) y (t) = B (q ) u(t) = b0 u(t d0 ) + b1 u(t d0 1) + + bm u(t d0 m) = B + (q )B (q )u(t), B (q ) const
April 22, 2010.
Lecture 7
p. 18/20


A o (q )A m (q ) = A (q )R (q ) + B (q )S (q )
April 22, 2010.
Lecture 7
p. 18/20


A o (q )A m (q ) = A (q )R (q ) + B (q )S (q ) leads to the regression Ao (q )Am (q )y (t) = B (q ) R(q )u(t) + S (q )y (t)
April 22, 2010.
Lecture 7
p. 18/20


A o (q )A m (q ) = A (q )R (q ) + B (q )S (q ) leads to the regression Ao (q )Am (q )y (t) = B (q ) R(q )u(t) + S (q )y (t)
which is nonlinear in parameters and has unstable factor

(t) = B (q ) R(q )u(t) + S (q )y (t) = (t 1)T
p. 18/20

Given
polynomials Am (q ), Bm (q ), Ao (q ) the relative degree d0 of the plant
April 22, 2010.
Lecture 7
p. 19/20

Given
Step 1: Estimate coefcients of R(q ) and S (q ) from

y (t) = R(q )uf (t) + S (q )yf (t) = (t)T
Cancel possible common factors of R(q ) and S (q ) to compute R(q ) and S (q )
April 22, 2010.
Lecture 7
p. 19/20

Given

y (t) = R(q )uf (t) + S (q )yf (t) = (t)T
Cancel possible common factors of R(q ) and S (q ) to compute R(q ) and S (q ) Step 2: Compute the control signal by R(q )u(t) = T (q )uc (t) S (q )y (t) where
T (q ) = Ao (q )Bmp (q )
April 22, 2010.
Lecture 7
p. 19/20

Given

y (t) = R(q )uf (t) + S (q )yf (t) = (t)T
Cancel possible common factors of R(q ) and S (q ) to compute R(q ) and S (q ) Step 2: Compute the control signal by R(q )u(t) = T (q )uc (t) S (q )y (t) where
T (q ) = Ao (q )Bmp (q )
Repeat Steps 1 and 2

p. 19/20
Next Lecture / Assignments:

Next meeting (April 26, 13:00-15:00, in A205Tekn).
Homework problems: Consider the process: 1 , where a is an unknown parameter. Assume G (s ) = s (s + a ) that the desired closed-loop system is 2 G m (s ) = 2 . Construct 2 s + 2s +
continuous-time indirect algorithm, discrete-time indirect algorithm, continuous-time direct algorithm, discrete-time direct algorithm.
April 22, 2010.
Lecture 7
p. 20/20

Lecture 07 EAILC

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Lecture 07 EAILC

Transféré par

Droits d'auteur :

Formats disponibles

Lecture 7: Deterministic Self-Tuning Regulators

Feedback Control Design for Nominal Plant Model

via Pole Placement

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Indirect Self-Tuning Regulators

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Indirect Self-Tuning Regulators

It can be re written in the regression form

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Indirect Self-Tuning Regulators

It can be re written in the regression form

y (t 1), . . . , y (t n), u(t d0 ), . . . , u(t d0 m)

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

RLS Algorithm (discrete time)

dened by the model

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

RLS Algorithm (discrete time)

dened by the model

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

RLS Algorithm (continuous time)

dened by the model

obtained from A(p) y ( ) = B (p) u( ) using ltered signals

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

RLS Algorithm (continuous time)

dened by the model

obtained from A(p) y ( ) = B (p) u( ) using ltered signals

Algorithm using RLS and MD Pole Placement

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Algorithm using RLS and MD Pole Placement

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Algorithm using RLS and MD Pole Placement

with estimates for A and B taken from the previous Step.

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Algorithm using RLS and MD Pole Placement

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Algorithm using RLS and MD Pole Placement

Repeat Steps 1, 2, 3 (until the performance is satisfactory).

Example 3.4 (Recursive Modication of Example 3.1)

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Example 3.4 (Recursive Modication of Example 3.1)

its pulse transfer function with sampling period h = 0.5 sec is

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Example 3.4 (Recursive Modication of Example 3.1)

its pulse transfer function with sampling period h = 0.5 sec is

except for for

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Example 3.4 (Recursive Modication of Example 3.1)

April 22, 2010.

Elements of Iterative Learning and Adaptive Control:

Example 3.4 (Recursive Modication of Example 3.1)