Upload
walid153
View
238
Download
0
Embed Size (px)
Citation preview
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 1/170
Advanced Feedback Control
. . . Models, Identification, Control, Robustness.
Monterrey, June 2007
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 2/170
Various structures of feedback
We will study various structures of feedback laws.Those structures depend on whether the feedback isapplied
• from the all state vector (if it can be measured),
• of from the ouput vector.
Besides, the feedback itself can be either• static (the control vector entries are linearcombinations of the measurements),
• or dynamic (the control vector becomes the ouputof dynamic system - the controller - from whichthe measurement vector is the input).
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 3/170
Various structures of feedback
Therefore, three kinds of structures will be considered• static state feedback (usually simply referred to
as state feedback),
• static ouput feedback,
• dynamic ouput feedback.
Indeed, dynamic state feedback is rarely used.
The system model to be considered is simply a
realization
x = Ax + Bu
y = Cx + Du
(1)
(x ∈ IRn, u ∈ IRm and y ∈ IR p) Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 4/170
Static state feedback
This is a control law that might have been studied inthe case of SISO sytstems
A
C yu
D
B + +
+
+xxH yc +
+
K
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 5/170
Static state feedback
The corresponding mathematical description is
u(t) = Hyc(t) + Kx(t). (2)
with
• K ∈ IRm×n: state feedback matrix;
• H ∈ IR p×n: feedforward matrix;• yc ∈ IR p: reference vector.
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 6/170
Static state feedback
The induced feedback model is given by
x = (A + BK )x + BHyc
y = (C + DK )x + DHyc.(3)
• K is computed to ensure stability and either to
possibly reach transient performances (poleplacement) or to minimize some criterion (e.g.LQ control, optimal control).
• H is rather computed to reach staticperformances.
Advanced Feedback Control – p.
t t t t
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 7/170
tat c output ee ac
Assume that not all the entries of x are measured butonly the entries of y.
A
C yu
D
B + +
+
+xxH yc +
+
F
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 8/170
t t t t
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 9/170
tat c output ee ac
If D = O (no direct transmission to make simpler)then the closed-loop model is
x = (A + BF C )x + BHycy = Cx
(5)
If D = Othen the control vector u complies with
u = Hyc + F Cx + F Du
⇔ (II m − F D)u = Hyc + F Cx
Advanced Feedback Control – p.
tat c o tp t ee ac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 10/170
tat c output ee ac
⇔ u = (II m − F D)−1H
yc + (II m − F D)−1F
C
⇔ u =ˆ
H yc +ˆ
F C (6)This leads to the following closed-loop model:
x = (A + BF C )x + BHyc
y = (C + DF C )x + DHyc.(7)
One can compute F and H for design purpose anddeduce F and H which are implemented in practice.
Advanced Feedback Control – p. 1
tat c output ee ac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 11/170
tat c output ee ac
Another possibility is to modify the control law:
u = Hyc + F (y − Du) = Hyc + F y (8)
y = y − Du = Cx ∈ IR p is the new "measure" onehas to built (it’s part of the controller) so that one gets
A
Ê
C yu
D
B + +
+
+xxH yc +
+
F +
−y
Advanced Feedback Control – p. 1
tat c output ee ac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 12/170
tat c output ee ac
Then, the closed-loop model directly depends on F and H :
x = (A + BF C )x + BHycy = (C + DF C )x + DHyc
(9)
F and H are computed to get satisfactoryperformances. Note however that, with this kind of structure, it might be preferable to measure u too.
Advanced Feedback Control – p. 1
ynam c ouput ee ac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 13/170
ynam c ouput ee ac
A
Ê
C yu
D
B + +++xxH yc +
+
F 3
F 4
Ê
F 1
+
+ zF 2
+
+
z
Advanced Feedback Control – p. 1
ynam c ouput ee ac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 14/170
ynam c ouput ee ac
The control law is given by
z = F 1z + F 2y
u = F 3z + F 4y + Hyc,(10)
where z ∈ IRlis the state vector of the feedback
system.
The transfer matrix of this controller is
GF (s) = F 3(sII l − F 1)−1F 2 + F 4. (11)
Advanced Feedback Control – p. 1
ynam c ouput ee ac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 15/170
ynam c ouput ee ac
Linking controller and process realizations yields
(II m − F 4D)u = F 3z + F 4Cx + Hyc
u = (II m − F 4D)−1F 4 Cx+ (II m − F 4D)−1F 3 z+ (II m − F 4D)−1H yu = F 4 Cx+ F 3 z+ H y
Consider a concatenation of process and controllerstate vectors ξ′ = [x′ z′]′ to get:
ξ = A + Bˆ
F 4C Bˆ
F 3F 2C + F 2DF 4C F 1 + F 2DF 3
ξ + Bˆ
H F 2DH
yc
y =
C + DF 4C DF 3
ξ + DHyc.
(12)Advanced Feedback Control – p. 1
ynam c ouput ee ac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 16/170
ynam c ouput ee ac
Let the next augmented dynamic model be defined:
ξ = Aξ + Bu
y = Cξ + Du,(13)
where ξ ∈ IRn+l and
A = A O
O Ol ; B = B O
O II l ;
C = C O
O II l ; D = D O
O Ol .
Advanced Feedback Control – p. 1
ynam c ouput ee ac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 17/170
ynam c ouput ee ac
Also let some control law (static ouput feedback) beapplied on this model:
u = F y + Hyc, with (14)
F = F 4 F 3F 2 F 1
an H = H
Ol,m , (15)
After some few calculation, one gets the same
closed-loop model as the one obtained by applyingthe dynamic feedback on the original process model.
⇒ Applying a dynamic output feedback controller ona linear model is equivalent to applying a static feedback gain on an augmented s stem. Advanced Feedback Control – p. 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 18/170
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 19/170
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 20/170
Reference
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 21/170
Reference
The developments in this part are actually very easy toproduce whith quite simple calculation and matrixmanipulations. Only one reference might deserve to
be quoted, where the dynamic controller is formulatedas a static one applied on an augmented system:
P. Hippe et J. O’Reilly.Parametric compensator design.International Journal of Control, Vol 45(4), p.1455-1468, 1987.
Advanced Feedback Control – p.
oup ng etween c anne s
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 22/170
oup ng etween c anne s
The purpose in this part is to highlight the inherentdifficulty of controlling MIMO models due tocoupling bewteen the various inputs to outputs
channels.
For example, consider a process with two ouputs y1
and y2 and to control signals u1 and u2.It is interesting two control y1 that should track somereference yc1 as well as to control y2 that should track some reference yc1.
Unfortunately, in most cases, those control lawscannot be designed independently. An action on yc1,
an thus on u1 has an influence on y2 and the other wayaround.
Advanced Feedback Control – p.
n examp e: c em ca reactor
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 23/170
n examp e: c em ca reactor
Mixer
Products (output chemicals)
Outgoing fluid
Heating/ Cooling
Reacting chemicals (input chemicals)
Entering fluid
Advanced Feedback Control – p.
n examp e: c em ca reactor
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 24/170
n examp e: c em ca reactor
The temperature of the reactor is highly influent onthe quality of the reaction which is itself influent onthe temperature of the environment.
The process includes two inputs:
• the rate (concentration) of entering chemicals,
• the temperature of the heating/cooling fluid,
and two ouputs
• the rate (concentration) of outgoing chemicals,• the temperature inside the reactor.
Advanced Feedback Control – p.
n examp e: c em ca reactor
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 25/170
n examp e: c em ca reactor
So two channels input/ouput:• one for the chemical rates;
•the other one for the temperature.
Why is there a couplig between the two channels?
• If the temperature of the outside fluid changes (in
order to control that of the reactor), then thequality of the reaction is modified and the rates of the products are changed.
• If the rates of the entering chemicals are changed(in order to control the rates of products), then thereaction is of course more or less important
inducing a change of temperature because thereaction either provides or absorbs heat.Advanced Feedback Control – p.
not er numer ca examp e
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 26/170
not er numer ca examp e
A = −1 0
0 0, 5
; B =
1 1
1 1
C = −1 0
0 1
; D =
0 0
0 0
.
It is an unstable square system (see the poles). Theemphasized entries are those responsible for thecoupling.
The corresponding transfer matrix is G(s) =
1
s2 + 0, 5s − 0.5 s
−0, 5 s
−0,5
s + 1 s + 1 = G11(s) G12(s)
G21(s) G22(s) .
Advanced Feedback Control – p.
not er numer ca examp e
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 27/170
a a p
Assume that one ignores (!!!) the coupling transfersG12(s) and G22(s) and that one designs somecontrollers only for diagonal transfers.
From u1 to y1:
G11(s) =s−
0, 5
s2 + 0, 5s − 0, 5 =1
s + 1 ,
From u2 to y2:
G22(s) =s + 1
s2 + 0, 5s − 0, 5=
1
s − 0, 5.
Advanced Feedback Control – p.
not er numer ca examp e
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 28/170
p
For each first order channel, ones applies
F i Gii(s)H iyci + yi
−
ui
WithH 1 = −1
,F 1 = −0.5
,H 2 = 0.5
andF 2 = 1
,one gets the two following closed-loop models:
G11 = G22 =
1
1 + 2s.
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 29/170
not er numer ca examp e
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 30/170
p
With such a simple (and false) reasoning, one shouldget the next step response:
−1
−0.5
0
0.5
1
T o : O u t ( 1 )
From: In(1)
0 2 4 6 8 10 12−1
−0.5
0
0.5
1
T o : O u t ( 2
)
From: In(2)
0 2 4 6 8 10 12
Step Response
Time (sec)
A m p l i t u d e
Advanced Feedback Control – p.
not er numer ca examp e
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 31/170
p
...whereas one actually getsStep Response
Time (sec)
A m p l i t u d e
−0.5
0
0.5
1From: U(1)
T o : Y ( 1 )
0 2 4 6 8 10 12−2
−1.5
−1
−0.5
0
0.5
T o : Y ( 2 )
From: U(2)
0 2 4 6 8 10 12
Advanced Feedback Control – p.
Hence...
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 32/170
... from these examples, one can conclude that:
• The coupling cannot always be neglected;
• The responses can be drastically distorted.
Indeed, some models can even be unstable due tocouplings...
Advanced Feedback Control – p.
Thus...
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 33/170
One can formulate several problems:
• Static decoupling (only for steady-stateresponse),
• Tansient decoupling (also for the transientresponse).
Those problems can be handled
• either from a frequency point of view (frequency
decoupling),• or from a time point of view (state-space
approach).
Advanced Feedback Control – p.
req. app. stat c ecoup ng
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 34/170
Some possibility is to use feedforward control:yc u y
H (s) G(s)
Y (s) = G(s)U (s) = G(s)H (s)Y c(s)
⇒ y∞ = limt→∞ y(t) = lim
s→0(sY (s)) = lim
s→0(sG(s)H (s)Y c(s)).
If ones considers that all the reference entries yci aresteps of magnitude αi, one has to satisfy:
Advanced Feedback Control – p.
req. app. stat c ecoup ng
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 35/170
y∞ = lims→0
sG(s)H (s)
1
s
α1
...
α p
=
α1
...
α p
⇔ G(0)H (0)α1
.
..
α p
= α1
.
..
α p
⇔ G(0)H (0) = II p. (18)
So H (s) = H (0) = H (constant feedforward matrix)has to check (18).
Advanced Feedback Control – p.
req. app. stat c ecoup ng
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 36/170
• If m = p (square model) then H = G(0)−1;• If m > p then H can be a pseudo-inverse of G(0)
(for example, the Moore-Penrose one);
• If m < p then no generic solution: not enoughactuators compared with the number of outputs.
So the limits are:• m ≥ p;
• G(0) must be of full rank;• The process must be stable or be stabilized firstbecause this is only a feedforward control.
Advanced Feedback Control – p.
xamp e
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 37/170
G(s) =
20(s + 1)
(s2 + 3s + 12)(s + 2)
−130(s − 0, 3)
s2 + 2s + 80
−10(s − 3)
(s2 + 3s + 12)(s + 8
15(s − 1)(s2 + 4s + 12)(s + 2)
43(s + 1(s2 + 2s + 32)(s + 2)
30(s + 1)s2 + 2s + 122
−9(s
−4)
s2 + 2s + 52
30(0, 5s + 4)
s2 + 2s + 412
3, 2
s + 2
This is a square stable model ⇒
H = G(0)−1 =
0, 833 −0, 572 −0, 075
0, 972 0, 927 −0, 332
−0, 537 0, 079 0, 718
.
Advanced Feedback Control – p.
req. app. ynam c ecoup ng
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 38/170
Just some idea that can sometimes be used !
The idea is to compute H (s) such that
Q(s) = G(s)H (s) checks
qii(s) = 0 ∀i ∈ {1,...,p}
qij(s) = 0 ∀{i, j = i} ∈ {1,...,p}2
But it is illusory to solve such constraints so one cansimply try to reach
|qij(iω)
|<< 1
∀{i, j
= i
} ∈ {1,...,p
}2
Advanced Feedback Control – p.
req. app. ynam c ecoup ng
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 39/170
There are several techniques in the literature based onthat simple idea (whose efficiency has still to beproved (author’s note)). With those techniques, one
has to check that the useful transfers qii(s)• have no instable zeros;
• are strictly proper;
• should be preferably of weak order.
In any case, one has to keep in mind that a decoupling
procedure does not ensure other performances andshould be accompanied by other control laws toguarantee stability, transient behaviour, and so son.
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 40/170
ecoup ng w t t me approac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 41/170
These p independent linear 1st order differentialequations correspond, in Laplace’s domain, to:
Y i(s)Y ci(s) = −qiis − qii
∀i ∈ {1,...,p}.
that is to some transfers with unit static gain and one
pole qii.
Advanced Feedback Control – p.
ecoup ng w t t me approac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 42/170
If one looks for a state feedback such that thesetransfers are obtained, ones can write
x = Ax+Bu = Ax+B(Hyc+Kx) = (A+BK )x+BHyc
⇔ C x = y = (CA + CBK )x + CBHyc.
to be identified to
y = QCx − Qyc,
leading to (assuming that W = (CB)−1
exists)
K = W (QC − CA)
H = −W Q.(19)
Advanced Feedback Control – p.
ecoup ng w t t me approac
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 43/170
Thus a very simple technique that is unfortunatelyonly useful when m = p = n.
Indeed,• It cannot be extended to static output feedback;
•it cannot be used when D
= O.
... so very restrictive !
Advanced Feedback Control – p.
ecoup ng: ome conc us on
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 44/170
• Frequential approach and feedforward sometimesefficient (not alone) for static decoupling.
• Time approach rarely used (except under drastic
constraints) but see the next part for some attemptto transient decoupling.
• There exist other methods of decoupling such as
the "relative gain" method whose efficiency hasnot convinced the author of these slides.
• Other techniques consists in tracking a reference
model which is usually chosen with no coupling...but it is not a decoupling approach in itself.
As a conclusion, decoupling is fundamental but sodifficult !Advanced Feedback Control – p.
References
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 45/170
• P. T. Tham
Notes - An introduction to Decoupling control.Department of Chemical and Process Engineering, University of
Newcastle upon Tyne, England... for the example of chemical reactor
•Course Notes, Chapter 6 : Analysis and Design of Multivariable
Control Systems.
Electrical Engineering Department, State University of Binghamton,
New-York, USA... for decoupling by time approach and other insights.
• E. H. Bristol
On a new measure of interaction in multivariable process control.
IEEE Transactions on Automatic Control, Vol 11, p. 133-134... for the
reader interested in "relative gain approach".• J. P. Corriou.
Commande des procédés.
Lavoisier Editions, TEC&DOC Collection, 1996 (in French, sorry !),
for connected information.Advanced Feedback Control – p.
genstructure ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 46/170
Motivation: assigning the poles and possibly theassociated eigenvectors in order to try to shape thetransient response of the closed-loop system.
It can be way to obtain some transient input/ouputdecoupling.
Techniques based upon eigenstructure placement arealso called Modal Control.
Advanced Feedback Control – p.
atr x e genstructure
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 47/170
λ is an eigenvalue of A
∈lC n×n iff
P (λ) = det(λII n − A) = 0. (20)
A owns n eigenvalues λi
(which will be assumeddisctinct for the sake of conciseness). This set isreferred to as the spectrum λ(A).
A ∈ IRn×n ⇒ λ(A) is closed under conjugation.
There exists n non zeros vectors vi
∈lC n, called right
eigenvectors, such that
Avi = λivi ∀ i ∈ {1,...,n}. (21)
One should talk about eigendirections since they canbe multiplied by any non zero scalar. Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 48/170
ee ac mo e e genstructure
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 49/170
Closed-loop model eigenstructure=eigenstructure of its state matrix
Ac = A + BFC.
⇒
Acvi = (A + BF C )vi = λivi ∀i ∈ {1,...,n}u′
iA = u′i(A + BF C ) = λiu
′i ∀i ∈ {1,...,n}
U ′V = II n
Ac = A + BF C = V ΛU ′.Advanced Feedback Control – p.
ee ac mo e e genstructure
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 50/170
Remark: In practice, the matrices are real meaningthat not only λ(Ac) but also the sets of eigenvectorsare closed underconjugation.
• Input directions:
wi = F CV i ∀i ∈ {1,...n}.• Output directions:
l′i = u′iBF ∀i ∈ {1,...n}.
Advanced Feedback Control – p.
n uence o t e e genva ues
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 51/170
It can be easily proved that the free response of amodel to an intial condition is
x(t) =
ni=1
αieλitvi. (26)
• Re(λi) < 0 ∀i otherwise there are non vanishingterms (instability).
•
|Re(λi)
| ց⇒the term (mode) reduces faster.
• |Im(λi)| ր⇒ the term induces strongeroscillation (none if λi is real).
So λ(A) has an influence on stability, settling time,oscillations, characterizing the transient behaviour.Advanced Feedback Control – p.
n uence o t e e genvectors
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 52/170
Consider the pertubed closed-loop model
x = (A + BF C )x + BHyc + Bd
y = Cx.
(27)
With the basis change x = V ξ, V being the modalmatrix of Ac = A + BF C :
ξ = Λξ + UBHyc + U Bd
y = C V ξ .(28)
Also consider the identity matrix:
II n = e1 . . . en , (29)Advanced Feedback Control – p.
n uence o t e e genvectors
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 53/170
• yc[i] has no effect on λ j iff
u′ jBHei = 0.
⇒ left eigenvectors distribute the effects of thereferences on the eignvalues
• λi has no effect on x[ j] (resp. y[ j]) iff
e′ jvi = 0.
⇒ right eigenvectors vi (resp. Cvi) distribute theeffects of the eigenvalues on the state entries(resp. outputs).
Advanced Feedback Control – p.
n uence o t e e genvectors
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 54/170
• d[i] has no effect on λ j iff
u′ jBei = 0.
⇒ left eigenvectors distribute the effects of somedisturbances on the eigenvalues
• λ j has no effect on u[ j] iff (less obvious)
e′ jwi = 0.
⇒ input directions distribute the effects of theeigenvalues on the control entries.
Advanced Feedback Control – p.
n uence o t e e genstructure
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 55/170
The effect of the environment on the system dynamicsis mainly described by the left eigenstructure whereasthe effect of these dynamics on the system outputs ismainly described by the right eigenstructure (IbrahimChouaib).
Remark: It can also be proved that eigenvectors have
an influence on the local sensitivity of eigenvalueswith respect to additive unstructured uncertaintyaffecting the state matrix (not detailed here).
Advanced Feedback Control – p.
tate ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 56/170
Pole Placement Problem: find K ∈ IRm
×n
such that λ(Ac = A + BK ) equals some specified set.
The computation of feedforward matrix H will beconsidered later.
There is always some solution provided the pair(A, B) is controllable.
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 57/170
tate ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 58/170
Characterisitc subspacesBecause for one λ, the associated v and w = Kvcomply with
(A − λII n)v + Bw = O,
then v
∈S (λ) where
S (λ) = {v ∈ lC n | ∃w ∈ lC m | (A−λII n)v +Bw = O}(A, B) controllable ⇒ dim(S (λ) = m.
⇒only (m
−1) should be exploited to assign
v ∈ S (λ), exactly what is offered by K .Advanced Feedback Control – p.
tate ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 59/170
Define:
T λ =
A − λII n B
∈ lC n×(n+m),
Rλ = N λ
M λ
= Ker(T λ) with N λ ∈ lC n×m, M λ ∈ lC m
and with some parameter vector z ∈ lC m, it comes
π = v
w = N λz
M λz ,
leading to admissible eigenvector v ∈ S (λ) and
associated input direction w.Advanced Feedback Control – p.
tate ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 60/170
Choice of z
Assume vd is ome desired eigenvector (with forinstance zero entries to try to reach some decouplingproperties). One has to assign an admissible v as closeas possible to vd. Solving a classical least squareproblem leads to
z = (N ′λN λ)−1N ′λvd. (30)
Remark: It is possible to rather give specifications onvarious ui and then to deduce suitable vdi.
Advanced Feedback Control – p.
tate ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 61/170
Theorem 1 There exists K ∈ IRm
×n
solving the problem iff
(i) vectors vi are linearly independent;
(ii) vi = v j when λi = λ j;
(iii) vi
∈S (λi).
In this event the unique solution is given by
K = W V −1 (31)
where
W = w1 . . . wn .Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 62/170
utput ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 63/170
Pole Placement Problem: find F ∈ IRm
×n
such that λ(Ac = A + BF C ) equals some specified set.
Remark: it is possible but dangerous to assign only part of the spectrum following the same kind of reasoning as for state feedback.
Nessary condition for solving the problem : (A,B,C )minimal.
Advanced Feedback Control – p.
utput ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 64/170
About the dof:• At first fight, m × p entries in F so problem can
be solved if mp
≥n but not so simple.
• In 1975, Kimura proved that m + p > n ⇒generic assignability.
• Later (1981), it was proved that condition ismp ≥ n but in the field of complex matrices...but no need for a complex F !
• In 1996, Wang proved that sufficient condition inthe field of real matrices is mp > n but theassociated design method is not very tractable.
• In practice, the tractable (or even non iterative)techniques require that Kimura’s condition holds.
Advanced Feedback Control – p.
utput ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 65/170
What to do if Kimura’s condition does not hold?• Assign only part of the spectrum,
•or apply a dynamic feedback.
If it holds, there are several techniques available withdifferent restrictions:
• "Polynomial" design;
• Parametric approach;
• Geometric approach (very elegant);• Coupled Sylvester Equations (my favourite !),
•and many others I may not know or that still have
to be found.Advanced Feedback Control – p.
utput ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 66/170
Principle of the "Sylvester approach":
Solve the system:
AV − V Λ = −BW (right eigenstructure) (32)
U ′A − ΛU ′ = −L′C (left eigenstructure) (33)
Ker(U ′) = Im(V ) (orthogonality) (34)
The main idea: assign {λi, i ∈ {1,...,p}} and the
associated vi as well as {λi, i ∈ { p + 1,...,n}} andthe associated ui, while respecting the three aboveequations.
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 67/170
utput ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 68/170
Simplified algorithm (cont’d):• Compute
N λi
M λi
= Ker( N λi) ∈ lC (n+m)×ri)
(it generically exists when m + p > n);• Choose p parameter vectors zi ∈ lC ri such that
V p = [Cv1,...,Cv p] = [CN λ1z1, ... , CN λpz p] (37)
is a full rank matrix;
Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 69/170
utput ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 70/170
With direct transmission D:
Just find F by the above technique to assign thespectrum of
Ac = A + BF C
and then deduce
F = F (II p + DF )−1.
Advanced Feedback Control – p.
utput ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 71/170
m + p ≤ nIt is possible to design a dynamic feedback or order lin order to assign n + l poles but one has to satisfy
l ≥ n − m − p + 1. (40)
Hence, Kimura’s condition holds for the "augmentedsystem" (see part of the various feedback structures)and then one computes a static gain for thisaugmented system that corresponds to a dynamic gainfor the original system.
Some special cases can also be handled with static
gain (since 2006 !)Advanced Feedback Control – p.
utput ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 72/170
Pole placement and feedforward
It might be possible (depending on dimensions) to
computeˆ
H and H = (II m − F D)ˆ
H such that
(
−(C + DF C )(A + BF C )−1B + D)H = II p. (41)
to ensure a unit static gain otherwise add integratorsbefore to solve the problem... but with integrators,
Kimura’s condition is harder to satisfy.
Advanced Feedback Control – p.
utput ee ac ass gnment
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 73/170
Simple example with MATLAB
Model and desired spectrum:
>> A=[1 4 5;0 2 6;1 0 3];
>> B=[1 1;1 0;0 0];
>> C=[1 0 0;0 1 0];
>> D=eye(2);
>> n=3;m=2;p=2;>> lambda=[-1 -2 -3];
Advanced Feedback Control – p.
utput ee ac ass gnment
Choice of L solution to "left" Sylvester equation:
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 74/170
Choice of Ln− p, solution to left Sylvester equation:
>> Lam_nmoinsp=diag(lambda(p+1:n))
Lam_nmoinsp =
-3
>> L_nmoinsp=[1;1];
>> U_nmoinsp=sylv(A’,-Lam_nmoinsp’,-C’*L_nmoinsp)
U_nmoinsp =
-0.3025
0.0420
0.2101
Advanced Feedback Control – p.
utput ee ac ass gnment
Computation of 1 its kernel v1 and w1:
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 75/170
Computation of
N 1, its kernel, v1 and w1:
>> NN1=[A-lambda(1)*eye(3) B;U_nmoinsp’ zeros(n-p,m)]
NN1 =
2.0000 4.0000 5.0000 1.0000 1.0000
0 3.0000 6.0000 1.0000 0
1.0000 0 4.0000 0 0
-0.3025 0.0420 0.2101 0 0
>> R1=null(NN1)
R1 =
-0.0364
-0.3074
0.0091
0.8675
0.3892
>> v1=R1(1:3);w1=R1(4:5);
Advanced Feedback Control – p.
utput ee ac ass gnment
The same for λ2.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 76/170
The same for λ2.
>> NN2=[A-lambda(2)*eye(3) B;U_nmoinsp’ zeros(n-p,m)]
NN2=
3.0000 4.0000 5.0000 1.0000 1.0000
0 4.0000 6.0000 1.0000 0
1.0000 0 5.0000 0 0
-0.3025 0.0420 0.2101 0 0
>> R2=null(NN2)
R2 =
-0.0305
-0.2499
0.0061
0.9629
0.0975
>> v2=R2(1:3);w2=R2(4:5);Advanced Feedback Control – p.
utput ee ac ass gnment
Computation of W V and F :
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 77/170
Computation of W p
, V p
and F :
>> Wp=[w1 w2];Vp=C*[v1 v2];
>> F_hat=Wp*inv(Vp)
F_hat =
-285.8000 31.0000
242.8000 -30.0000Verification of the cloded-loop spectrum:
>> eig(A+B*F_hat*C)
ans =
-3.0000
-2.0000
-1.0000
Advanced Feedback Control – p.
utput ee ac ass gnment
Deduction of F and computation of H :
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 78/170
p
>> F=F_hat*inv(eye(p)-D*F_hat)
F =
-0.9773 0.0227
0.1780 -0.7897
>> H_hat=inv(-(C+D*F_hat*C)*inv(A+B*F_hat*C)*B+D)
H_hat =
58.5529 -84.1059
-49.6706 71.3412
>> H=(eye(p)-F*D)*H_hat
H =
-0.2161 0.3106
-0.0962 0.1406Advanced Feedback Control – p.
utput ee ac ass gnment
Construction of the closed-loop model and verification of the static gain
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 79/170
p g
>> Ac=(A+B*inv(eye(m)-F*D)*F*C);
>> Bc=B*inv(eye(m)-F*D)*H;
>> Cc=(C+D*inv(eye(m)-F*D)*F*C);
>> Dc=D*inv(eye(m)-F*D)*H;
>> -Cc*inv(Ac)*Bc+Dc
ans =
1.0000 -0.0000
0.0000 1.0000
Advanced Feedback Control – p.
References• B. C. Moore
On the flexibility offered by state feedback in multivariable systems
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 80/170
On the flexibility offered by state feedback in multivariable systems
beyond closed-loop eigenvalue assignment .
IEEE Transactions on Automatic Control, Vol 21, p.689-692, 1976, for
the state feedback
• H. Kimura.
Pole assignement by gain output feedback . IEEE Transactions on
Automatic Control, Vol 20, p.509-516, 1975, for m + p > n
• J. Rosenthal et F. Sottile.Some remarks on real and complex output feedback . Systems and
Control Letters, Vol 33, p.73-80, 1997, for other conditions including
Wang’s one
Advanced Feedback Control – p.
References (2)• V. L. Syrmos et F. L. Lewis.
Output feedback eigenstructure assignment using two Sylvester
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 81/170
p f g g g y
equations.
IEEE Transactions on Automatic Control, Vol 38(3), p.495-499,
1993.for static output feedback (Sylvester approach)
• A. N. Andry, E. Y. Shapiro et J. C. Chung.
Eigenstructure assignment for linear systems
IEEE Transactions on Aerospace and Electronic Systems, Vol 19(5),
p.711-729, 1983, for the meaning of eigenstructure• O. Bachelier, J. Bosche et D. Mehdi
On pole placement via eigenstructure assignment approach.
IEEE Transactions on Automatic Control, Vol 51(9), p.1554-1558,
2006, for the case m + p = n
... and all the references therein.
Advanced Feedback Control – p.
Model reduction
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 82/170
The main idea: Approximate a high order (n) linearmodel by a reduced order (r) model to make thedesign simpler.
S
= x = Ax + Bu
y = Cx + Du →S
r = xr = Arxr + Bru
yr = C rxr + Dru(42)
Advanced Feedback Control – p.
ua ty o r
Several criteria can be considered:
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 83/170
• Preserve the dominant poles (i.e. neglect the fast(high frequency) dynamics);
• Approximate the input/ouput behaviour: for asame input vector, yr should be as close aspossible to y.
Advanced Feedback Control – p.
ome ex s tng met o s
According to these criteria, a non complete list of
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 84/170
g pexisting techniques is as follows:
• By modal approach;
• By "agregation";• By Schur decomposition;
•By minimization of norm (ex:
H∞-norm);
• By balancing transformation (the only onepresented here and maybe the most known!).
Advanced Feedback Control – p.
Balanced reduction
Only valid for asymptotically stable minimal models
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 85/170
y y p ybut there exists a (not well known) extension tounstable models.
The idea: neglect the dynamics of the state entries thatare the less controllable and observable in S.
But how to quantify controllability and observability?
Answer: through the grammians (or Gramm
matrices).
Advanced Feedback Control – p.
Balanced reduction
Controllability grammian
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 86/170
W c =
∞
0
eAτ BB′eA′τ dτ. (43)
which solves Lyapunov equation
AW c + W cA′ =
−BB′. (44)
Observability grammian
W o = ∞0
eA′
τ C ′
CeAτ dτ ⇒ (45)
A′W o + W oA =−
C ′C. (46)
(hence the stability assumption). Advanced Feedback Control – p.
Balanced reduction
The controllability grammian can be interpreted in
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 87/170
terms of energy.
There exists a basis in IRn in which W c is diagonal. In
this basis each diagonal entry wci of W c is thereciprocal of the minimum energy required to(asymptotically) bring the state vector to
[0, . . . , 0, 1, 0 . . . , 0]′. Thus, it can be seen as acontrollability index of xi.
For observability, the reasoning is based upon dualityto conclude that in the "diagonal" basis, woi is anobservability index of xi.
Advanced Feedback Control – p.
Balanced reduction
LS
b d ib d b h i l f i (A B C)
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 88/170
Let be decribed by the triplet of matrices (A,B,C )assuming D = 0 for the sake of conciseness.
Theorem: There exists a a full rank matrix T such that the realisationS = (T −1AT,T −1B,CT ) = (A, B, C ) is balancedi.e. both grammians equal to the same diagonal
matrixW o = W c = Σ.
It means that in this basis, for each entry xi, thecontrollability and observability indices are the same
⇒one has to neglect the dynamics of the less
controllable and observable xi.Advanced Feedback Control – p.
Balanced reduction
In the balanced basis, the model is (here, D is kept)
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 89/170
˙x1 = A11x1 + A12x2 + B1u
˙x2 = A21x1 + A22x2 + B2u
y = C 1x1 + C 2x2 + Du.
(47)
The limit between the preserved dynamics (that of x1
and the neglected ones (that of x2) depends on apossible gap in the diagonal entries of Σ.
Advanced Feedback Control – p.
Balanced reduction
The idea is then to neglect the dynamics of x2
, the lessll bl d b bl f ( h h l
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 90/170
2controllable and observable part of x (thus the lessinfluent on the input/ouput behaviour) by imposing˙x2 = 0. This technique is sometimes called the"singular perturbations approximation"
It is also possible to simply truncate x by imposing
x2 = 0 but it does not preserve the static gain so it israther rough as a reduction.
Advanced Feedback Control – p.
Balanced reduction
So, the reduced model S is given by
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 91/170
r
Sr
= xr = Arxr +Bru
yr = C rxr + Dru.
(48)
where xr = x1 and
Ar = A11 − A12A−122 A21
Br = B1
−A12A−1
22 B2
C r = C 1 − C 2A−122 A21
Dr = D − C 2A−122 B2.
(49)
Advanced Feedback Control – p.
Balanced reduction
MATLAB corresponding functions
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 92/170
• minreal: Compute the minimal realization of original system S;
• balreal: Compute the balanced realization
(A, B, C, D) of S;
• modred: Compute the final realization(Ar, Br, C r, Dr) of reduced system Sr.
Advanced Feedback Control – p.
References• E. J. Davison.
A method for simplifying linear dynamic systems.
IEEE T ti A t ti C t l V l 11(1) 93 101 1966
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 93/170
IEEE Transactions on Automatic Control, Vol. 11(1), p. 93-101, 1966,
for a first glimpse to (non presented) modal approach
• B. Moore.
Principal component analysis in linear systems: controllability,
observability and model reduction
IEEE Transactions on Automatic Control, Vol. 26, p. 17-31, 1981, for
the bases of balanced reduction• L. Fortuna et G. Muscato
Model reduction via singular perturbation approximation of
normalized right coprime factorizations.
Proceedings of "European Control Conference ECC’95", Roma, Italy,
September 1995. for the balanced reduction of unstable models.
Advanced Feedback Control – p.
Norms
Definition and properties
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 94/170
A norm enables ones to compare an element withanother in a set which does not necessarily own a
relation of order. It is usually denoted by ||u||• whereu is the concerned element and • stands for theconsidered norm.
(i) ||u||• ≥ 0
(ii) ||u||• = 0 ⇔ u = 0
(iii) ||au||• = |a|.||u||•, ∀a ∈ lC(iv) ||u + v||• ≤ ||u||• + ||v||• (triangular inequality)
(50)
Advanced Feedback Control – p.
Vector norms
Euclidean norm
I d t f l { } ∈ { lC n}2
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 95/170
Inner product of a couple {x; y} ∈ { lC n}2:
< x, y > =
ni=1
x∗i yi = x∗y (51)
From this inner product, one can define the Euclidean
norm of 2-norm (the most natural):
||x||2 = √< x, x > =
ni=1
x2i = √x∗x. (52)
There are many other vector norms not detailed here.Advanced Feedback Control – p.
Vector function norms
Now the vectors depend on some real or complexvariable (t or s)
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 96/170
variable (t or s).
L2 and
H2-norms
Let Ln2 be the set of vector functions X (s) ∈ lC n, with
s ∈ lC whose square can be "summed" along theimaginary axis:
||X
||2 = 1
2π ∞
−∞X (iω)∗X (iω)dω
1/2
<∞
.
(53)
||X
||2 is called the
L2-norm of X (
Lfor Lebesgue).
It can be shown that Ln2 is an Hilbert space.Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 97/170
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 98/170
Vector function norms
L2-gain
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 99/170
LLet a mathematical operator R be defined over thefollowing sets:
R : Lnw
2 → Lnw
2
w(t)
→e(t)
Then,
GL2(R) = supw∈Hnw
2
||e||2||w||2(55)
is the
L2-gain of
Rwhich corresponds to the highest
energy gain associated with R.Advanced Feedback Control – p.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 100/170
Matrix norm
and Σ, if q = min
{m, n
}, complies with
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 101/170
{ }
Σ =
σ1 0 · · · 0 0
0 σ2 · · · 0 0
... ... . . . ... ...
0 0 · · · σq 0
si q = m,
Σ =
σ1 0 · · · 0
0 σ2 · · · 0...
.... . .
...
0 0 · · · σq
0 0 · · · 0
si q = n,
Σ = diag{σ1, · · · , σq} si q = m = n.
(58)
Advanced Feedback Control – p. 1
Matrix norm
σi are the singular values of M :
¯(M ) ≥ ≥ ≥ (M ) ≥ 0
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 102/170
σ(M ) = σ1 ≥ σ2 ≥ · · · ≥ σq = σ(M ) ≥ 0.(59)
• rank (M ) = number of non-zero singular values= number of linearly independant rows orcolumns.
• M such that rank (M ) < min{m; n} is rank deficient, otherwise full rank .
•M square and rank deficient cannot be inverted
and owns n − r zero singular values.
• σi = eigenvalues of M M ∗ (si m ≤ n) or M ∗M (si n
≤m).
• M Hermitian ⇒ σi = |λi|. Advanced Feedback Control – p. 1
Matrix norm
σ(M ) is a norm called 2-norm because it is inducedby the Euclidean vector norm in the following way:
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 103/170
by the Euclidean vector norm in the following way:
σ(M ) = ||M ||2 = max
x = 0 ∈ lC n
||M x
||2
||x||2 = max
x = 0 ∈ lC n
x∗M
∗
M xx∗x
(60)
Besides
σ(M )
≤
||M x||2
||x||2 ≤σ(M ). (61)
shows that the gain from x to (M x) lies in the range
[σ(M );σ(M )]
Advanced Feedback Control – p. 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 104/170
L∞ − H∞-norm of a transfer
LetRL
ne×nw
∞(resp.
RHne×nw
∞), be the set of proper
transfer matrices G(s) ∈ lC ne×nw (i e with finite
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 105/170
RL RHtransfer matrices G(s) ∈ lC (i.e. with finitedirect transmission) and with no pole on the
imaginary axis I
(resp. with no pole over lC +
∪ I ).
Then the L∞ (resp. H∞-norm) simply corresponds tothe frequency for which the transfer is the highest inthe sense of the 2-norm. Hence:
||G||∞ = supω ||G(iω)||2 = sup σ(G(iω)).ω
(63)
Advanced Feedback Control – p. 1
Transfer matrix norm
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 106/170
20log(σ(G(iω)))
20log(σ(G(iω)))
Pulsation ω (rad/s)
G a i n ( d B
)
20log(||G||∞)
Figure 1: Gain Bode diagramm in the MIMO caseAdvanced Feedback Control – p. 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 107/170
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 108/170
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 109/170
H2-norm
Stochastic interpretation
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 110/170
If the wi are white noises scaled such thatW (iω)W (iω)∗ = II nw
, then the expectation of the
inner product of the induced input vector checks
ne
i=1 E (e∗i (t)ei(t)) = ||G||
2
2.(68)
Remark: For this reason the so-called
H2-problem can
be related to the celebrated LQG-problem
Advanced Feedback Control – p. 1
2 2-norm
H2-norms in terms of gain
R i di h h H i h h L i h
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 111/170
Reminding that the H∞-norm is the the L2-gain thenthe
H2-norm checks
||G||2 = supW (s)
∈Hnw∞
||E ||2
||W
||∞. (69)
which, unlike for the H∞-norm, is not verymeaningful.
Advanced Feedback Control – p. 11
H2-norm
H2-norms and grammians
R i d f h ll bili d b bili
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 112/170
Remind of the controllability and observabilitygrammians:
W c =
∞0
eAtBBT eAT tdt ; W o =
∞0
eAT tC T CeAtd
(70)that satisfy
AW c + W cA
T
= −BB
T
,AT W o + W oA = −C T C. (71)
Advanced Feedback Control – p. 11
H2-norm
H2-norms and grammians (cont’d)
Th
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 113/170
Then
||G||2 = trace(BT W oB) = trace(CW cC T ).(72)
This provides an analytical expression of the
H2-norm
which is valid for calculation.
Unlike the
H∞-norm, the
H2-norm can be computed
directly through its analytic expression.
Advanced Feedback Control – p. 11
approac
For simplicity, only real matrices are considered.
M t i i liti
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 114/170
Matrix inequalities
This is related to the notion of sign definition (partialorder of Löwner).
M ∈ IRn×
n
is positive definite (M > 0) (resp.semi-positive definite (M ≥ 0)) iff
xT M x > 0 (resp.≥ 0) ∀x = 0 ∈ IRn. (73)
M is negative definite (M < 0) (resp. semi-negative
definite (M ≤ 0)) iff (−M ) > 0 (resp. (−M ) ≥ 0).Advanced Feedback Control – p. 11
atr x nequa t es
In practice mostly symmetric matrices are handled so
sign definition is now considered only for thosematrices
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 115/170
matrices.With this assumption,
M < (≤) 0 ⇔ λmax(M ) < (≤) 0
M > (≥) 0 ⇔ λmin(M ) > (≥) 0
(74)
A straightforward notation isM > (≥) N ⇔ M − N > (≥) 0
M < (≤) N ⇔ M − N < (≤)0.(75)
Advanced Feedback Control – p. 11
LMI bases
Example of MI:
M = M T = AX3 + (X3)T AT + eBY Y T (eB)T < 0
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 116/170
M = M = AX + (X ) A + e Y Y (e ) < 0
with, for instance, X and Y that are unknown.
Among all possible MI only two will be consideredbecause they are often encountered:
• LMI : Linear matrix inequalities, that can besolved,
• BMI : Bilinear matrix inequalities.
Advanced Feedback Control – p. 11
ropert es o
•If M 1 < 0 and M 2 < 0 then one can stack these
properties in one single MI:
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 117/170
M 1 O
O M 2 < 0. (76)
• If M = M T is such that
M =
M 1 M 2M T
2 M 3
< 0, (77)
then M 1 < 0 and M 2 < 0.
Advanced Feedback Control – p. 11
LMI
LMI are interesting because they can be solved !
The most famous LMI come from
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 118/170
The most famous LMI come from...
Theorem 2 Let the autonomous continuous (resp.
discrete) model
x = Ax (resp. xk+1 = Axk)
This model is asymptotically stable iff ∃P = P T > 0such that
AT P + P A < 0, (resp. − P + AT P A < 0).
(Lyapunov’s inequality and its discrete counterpartdue to Stein). Advanced Feedback Control – p. 11
BMI
Consider the next MI with respect to X and Y :
AX + XT AT + XBY + Y T BT XT > 0
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 119/170
AX + X A + XBY + Y B X > 0
Because of the two last terms, it is bilinear.
Unfortunately those BMI are very difficult to solve inspite of some existing software.
Some crucial control problems are unfortunately veryeasily formulated as BMI, not as LMI (ex: static
output feedback stabilization).
Advanced Feedback Control – p. 11
A useful tool !
Schur’s lemma
Let S Q = QT and R = RT be matrices
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 120/170
Let S , Q = Q and R = R be matrices.
Q S
S T R < 0 ⇔
R < 0
Q − SR−1
S T
< 0(78)
This lemma enables to handle Stein’s inequality as an
LMI wrt A:
−P + AT P A O
O −P < 0 ⇔ −P AT P
P A −P < 0.Advanced Feedback Control – p. 11
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 121/170
e process
• u: control vector issued from control law;
• w: disturbance to be rejected (in practice, notalways actual exogeneous signals);
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 122/170
always actual exogeneous signals);
• y: measured ouput for the purpose of control;• e: vector of signals to be controlled.
E (s)
Y (s)
= P (s)
W (s)
U (s)
, with P (s) =
P ew(s) P eu(s)
P yw(s) P yu(s)
.
The idea is to reduce the transfer from w to e.
Advanced Feedback Control – p. 1
e process
P (s) = D + C (sII − A)−1
B, with (79)
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 123/170
B = Bw Bu ; C = C e
C y ; D = Dew Deu
Dyw Dyu
.
(80)
Or in other words
x(t) = Ax(t) + Bww(t) + Buu(t)
e(t) = C ex(t) + Deww(t) + Deuu(t)y(t) = C yx(t) + Dyww(t) + Dyuu(t)
x(t) ∈ IRn
, w(t) ∈ IRnw
, u(t) ∈ IRnu
, e(t) ∈ IRne
et y(t) ∈ IRny
.Advanced Feedback Control – p. 1
e process
Assumptions
• A1 : (A; Bu) and (A; C y) are respectively
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 124/170
A1 : (A; Bu) and (A; Cy) are respectively
stabilisable and detectable;
• A2 : Dyu = Ony,nu;
• A3 : Dew = One,nw(only for H2-problem).
A1 is rather classical and completely compulsory.A2 is just technical and induces no loss of generality.A3 is necessary for the
H2-norm to be defined.
Advanced Feedback Control – p. 1
Static controller
One considers a static state feedback controller
u = Kx. (81)
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 125/170
u Kx. (81)
In such a case, since y = x, the process model reducesto
x(t) = Ax(t) + Bww(t) + Buu(t)e(t) = C ex(t) + Deww(t) + Deuu(t).
(82)
Advanced Feedback Control – p. 1
ynam c contro er
One considers a dynamic output feedabck controller
xc(t) = Acxc(t) + Bcy(t)(83)
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 126/170
c( ) c c( ) + cy( )
u(t) = C cxc(t) + Dcy(t)(83)
where xc(t) ∈ IRn. Thus,
K (s) = Dc + C c(sII n − Ac)−1
Bc. (84)In the H2-case (not studied in these slides), one has toconsider a strictly proper controller i.e. Dc = Onu,ny
.
Advanced Feedback Control – p. 1
ose - oop mo e
What is the state-space model of F(P (s), K (s)) ?
With static controller
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 127/170
Under Assumption A2:
x
e = A
f B
f C f Df
x
w = A + B
uK B
wC e + DeuK Dew
In the
H2-case, Df = 0.
Advanced Feedback Control – p. 1
ose - oop mo e
With dynamic controller
Still under Assumption A2:
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 128/170
p
xxc
e
=
Af Bf
C f Df
xxc
w
=
A + BuDcC y BuC c Bw + BuDcDyw
BcC y Ac BcDyw
C e
+ Deu
Dc
C y
Deu
C c
Dew
+ Deu
Dc
Dyw
x
xc
w
.
In the H2-case, Df = 0.
Advanced Feedback Control – p. 1
•-pro em
Problem 1 Let P (s) and γ •
> 0 be given. Also let
assumptions A1 to A3 hold. Find a stabilizing (staticor dynamic) feedback such that
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 129/170
y ) f
||F(P (s), K (s))
||•< γ
•.
If • = ∞, then Assumption A3 can be omitted.
With no additional constraints (such as weightingmatrices), the problem is referred to as standard.
Advanced Feedback Control – p. 1
•-pro em
H2 or
H∞?
... not exactly the same philosophy.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 130/170
In the H∞-case, one looks after the L2-gain i.e. thehighest possible energy transfer or, from the frequencyviewpoint, the energy transfer at the worst frequency.
In the H2-case, one considers energy transfer over thewhole frequency range, not focusing on the worst.
Advanced Feedback Control – p. 1
•-synt es s
In the following slides, some solutions are given for
• the H2-design by state static feedback,
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 131/170
• the
H∞-design by state static feedback,
• the H∞-design by output dynamic feedback,
Advanced Feedback Control – p. 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 132/170
2 stat c es gn
The idea is that X 2 is a matrix upper bound of either
the observability grammian (X 2 > W o: primalversion) or of the controllability grammian (X 2 > W c:d l i ) S h L i d
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 133/170
dual version). So the Lypaunov equations used to
calculate the grammians are here replaced by LMIs.
The dual version enables ones to derive some K .
Advanced Feedback Control – p. 1
2 stat c es gn
Theorem 3 There exists u = Kx , K
∈IRnu×n such
that the H2-norm of F(P (s), K (s)) is less than γ 2 iff there exist two symmetric positive definite matrices
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 134/170
{X 2; T
} ∈ {IRn×n
}2 , and a matrix L
∈IRnu×n such
that
AX 2 + BuL + X 2AT
+ LT
BT u Bw
BT w −II nw
< 0,
−T C eX 2 + DeuLX 2C T
e + LT DT eu −X 2
< 0,
trace(T ) = γ 22 . Advanced Feedback Control – p. 1
2 stat c es gn
In this event, K is given by
K = LX −12 .
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 135/170
• Notice that with various LMI solvers, it ispossible to minimize γ 2 while satisfying the LMI
constraints.• Also notice that X 2 can be inverted since it is
positive definite.
Advanced Feedback Control – p. 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 136/170
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 137/170
∞ stat c es gn
Theorem 4 There exists u = Kx , K
∈IRnu×n such
that the H∞-norm of F(P (s), K (s)) is less thanγ ∞ > 0 iff there exists a symmetric positive definite
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 138/170
matrix X ∞ ∈
IRn×n , and a matrix L
∈IRnu×n such
that
AX ∞ + B
uKX ∞ + X ∞AT + X ∞K T BT
u(•
) (•
)
BT w −γ ∞II nw (•)
C eX ∞ + DeuKX ∞ Dew −γ ∞II ne
< 0.
(87)
Advanced Feedback Control – p. 1
∞ stat c es gn
In this event, K is given by
K = LX −1∞ .
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 139/170
• Notice that with various LMI solvers, it ispossible to minimize γ ∞ while satisfying the LMI
constraints.• Also notice that X ∞ can be inverted since it is
positive definite.
Advanced Feedback Control – p. 1
∞ ynam c es gnCondition for solvability
Under assumptionsA
1-A
2, the H∞ dynamic problemcan be solved iff there exist R = RT and S = S T suchthat
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 140/170
that
N R O
O IInw
T
AR + RAT RC T e Bw
C eR −γ ∞II ne Dew
BT w DT ew −γ ∞II nw
N R O
O IInw
<
N S O
OII ne
T
AT S + SA SBw C T e
BT wS −γ ∞II nw DT
ew
C e Dew −γ ∞II ne
N S O
OII ne
< 0
R II n
II n S
≥ 0
(88)Advanced Feedback Control – p. 1
∞ ynam c es gn
where Span(N R) = Ker([BT u DT
eu]) and
Span(N R) = Ker([C y Dyw]).
M d t ll i t iff
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 141/170
Moreover, a n-order controller exists iff
rang(II n − RS ) = n. (89)
• It is also possible to achieve
min γ ∞ under the LMI constraints.R=RT ;S =S T
Advanced Feedback Control – p. 1
∞ ynam c es gn
How to recover K (s) ?
Achieve the singular value decomposition of
(II RS) in order to obtain M ; N IRn×n 2
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 142/170
(II n
−RS ) in order to obtain
{M ; N
} ∈ {IR
}such that
M N T = II n
−RS.
Then X ∞ = S N
N T
−M −1RN
is solution to the condition of the bounded real lemmawhich therefore becomes an LMI (thus solvable) wrt
(Ac, Bc, C c, Dc). Advanced Feedback Control – p. 1
∞ ynam c es gn
Example
e =
¾
z
u
¿
x 1 ( )+ x
w =
¾
b
v
¿
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 143/170
K (s)
u y
u
x
+
+
1
s
v
+
z P (s)
K (s)
+b x
y
Find the state-space model, write the LMI system,deduce the minimum value of γ ∞ and explain how to
recover K (s).
Advanced Feedback Control – p. 1
∞ ynam c es gn
State-space model:
b
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 144/170
x = [0]x + 1 0 b
v w
+ [1]u
z
u
e
=
1
0
x +
0 0
0 0
b
v
+
0
1
u
y = [1]x +
0 1
b
v
+ [0]u.
(90)
Advanced Feedback Control – p. 1
∞ ynam c es gn
which correspond to these matrices:
A = 0 Bw =
1 0
Bu = 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 145/170
C e = 1
0
Dew =
0 00 0
Deu =
01
C y = 1 Dyw = 0 1 Dyu = 0.
Assumptions A1 and A2 are easily verified.
Advanced Feedback Control – p. 1
c ∞ ynam c es gn
[BT u DT
eu] = [1
|0 1] = [C y Dyw]
⇒N R = N S
1 00 1
.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 146/170
−1 0 A is scalar then so are R and S . The first LMI to besolved is
1 0 0 0
0 1 0 0
−1 0 0 0
0 0 1 0
0 0 0 1
T
0 R 0 1 0
R −γ ∞ 0 0 0
0 0 −γ ∞ 0 0
1 0 0 −γ ∞ 0
0 0 0 0
−γ ∞
1 0 0 0
0 1 0 0
−1 0 0 0
0 0 1 0
0 0 0 1
Advanced Feedback Control – p. 1
∞ ynam c es gn
which becomes
⇔
γ ∞ −R −1 0
−R γ ∞ 0 0
> 0
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 147/170
⇔ −1 0 γ ∞
00 0 0 γ ∞
> 0.
and, by Schur’s lemma, is equivalent to
γ ∞ > 0,
γ 2∞ − 1 − R2 > 0.
Advanced Feedback Control – p. 1
∞ynam c es gn
In a totally similar way, the 2nd inequality reduces to
γ 2∞ − 1 − S 2 > 0.
R 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 148/170
whereas the 3rd one , i.e. R 1
1 S ≥ 0 leads to
R ≥ 0
RS − 1 ≥ 0
⇔
S ≥ 0
RS − 1 ≥ 0.
Advanced Feedback Control – p. 1
∞ynam c es gn
The whole of constraints yields
γ 2∞ − 1 > min (max{R2; S 2}).R,S
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 149/170
which shows that the optimum is reached for
R = S = 1 ⇒ γ ∞ =
√2.
But in this case, the optimal controller is not of ordern = 1. Anyway, for a suboptimal case RS = 1, one
gets
M = −N =√
RS − 1,
Advanced Feedback Control – p. 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 150/170
o e p acement
LMI region
Any set D ⊂ lC defined by
T
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 151/170
D = {z ∈ lC | α + βz + β T
z < 0} (91)where α = αT ∈ IRl×l and β ∈ IRl×l is an openLMI-region of order l.
These regions are always convex and, if α and β arereal (as in the above definition), symmetric wrt the
real axis.
Advanced Feedback Control – p. 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 152/170
o e p acement
LMI formulation of a disc
... of center ρ and radius r.
|z − ρ| < r ⇔ (z − ρ)(z − ρ) − r2 < 0
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 153/170
⇔ −r + (z − ρ)1r
(z − ρ) < 0.
Applying Schur’s lemma, it comes
−r z − ρ
z − ρ −r = α + βz + β T z < 0
⇔α =
−r −ρ
−ρ
−r < 0 ; β =
0 1
0 0 .
Advanced Feedback Control – p. 1
-sta ty
A matrix (or by extension a model) is said
D-stable
when its eigenvalues lie inside D.
Theorem 5 Let be an LMI-region. A matrix A is
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 154/170
DD-stable iif ∃X D = X T D > 0 such that
M D
(A, X D
) = α⊗
X D
+ β ⊗
(AX D
) + β T
⊗(X
DAT ) < 0.
This is an LMI wrt X D or xrt A.
Advanced Feedback Control – p. 1
D-stabilizationOne considers a static state feedback control law.
Theorem 6 Let D be an LMI-region. P (s) is
-stabilizable by state feedback iif X = X T > 0
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 155/170
Dy f f
∃ D Dand L such that
M D(A, Bu, X D, L) = α ⊗ X D + β ⊗ (AX D) + β T
⊗ (X DAT
)+
β ⊗ (BuL) + β T ⊗ (LT BT u ) < 0.
In this event, K is given by K = LX −1
D .
Advanced Feedback Control – p. 1
xt synt es s
One still considers a static state feedback control law.
Theorem 7 Let D be an LMI-region. P (s) is-stabilizable by state feedback that ensures
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 156/170
D||F(P (s), K )||• < γ • ∀ • ∈ {∞; 2} if ∃{X = X T > 0; T = T T ; L} such that
Z ∞(P (s), X , L , γ ∞) =
AX + BuL + XAT + LT BT u (•) (•)
BT w −γ ∞II nw (•)
C eX + DeuL Dew
−γ ∞II ne
Z 21(P (s), X , L) =
AX 2 + BuL + X 2A′ + L′B′u Bw
BT w
−II nw
< 0
Advanced Feedback Control – p. 1
xt synt es s
Z 22(P (s), X , L , T ) = −T C eX 2 + DeuL
X 2C T e + LT DT eu −II n
< 0
2
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 157/170
trace(T ) < γ 2
2
M D(A, Bu, X , L) = α⊗X +β ⊗(AX )+β T ⊗(XAT )+β ⊗(BuL)+β T ⊗(LT BT u ) <
In this event, K is given by K = LX −1.
•A great interest in LMI is that one can stack several
LMI constraints with preserving the LMI nature of theproblem but...
Advanced Feedback Control – p. 1
xt synt es s
• ...The condition is only sufficient because one
imposes the constraint
X = X 2 = X ∞ = X D.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 158/170
This is referred to as the Lyapunov ShapingParadigm.
•In such a mixt synthesis only one criterion is
minimized (either γ 2 or γ ∞) and the other one isarbitrarily chosen. Another possibility is tominimize a weighted objective function.
• The design of dynamic controllers is alsopossible but harder.
Advanced Feedback Control – p. 1
Introduction to robust controlPolytopic uncertainty
(...still considering state feedback)
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 159/170
The process model is assumed to be uncertain (notprecisely known) but also assumed to belong to afamily of models defined by
M = M (τ ) = A(τ ) Bw(τ ) Bu(τ )
C e(τ ) Dew(τ ) Deu(τ ) =N
j=1
(τ jM j
Advanced Feedback Control – p. 1
o ytop c uncerta nty
in which τ = [τ 1,...,τ N ]T contains the coefficients of
a convex combination (i.e. τ j ≥ 0 andN
j=1
(τ j) = 1)
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 160/170
and M j are the vertices of a so-called polytope of matrices:
M j = A j B jw B ju
C je D jew D jeu
. (92)
Remarque 1 There are many possible descriptions of uncertainties either in the frequency domain (i.e.affecting the transfer matrix) or in the time domain(i.e. affecting the state-space model). Here only few
time-domain uncertainties are introduced. Advanced Feedback Control – p. 1
o ytop c uncerta nty
Why polytopic uncertainty?
It has a very interesting special case i.e the affineparametric uncertainty:
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 161/170
M = M 0 +
p
i=1
(δiN i), (93)
M 0 is the nominal part, the matrices N i are known andthe δi are unknown parameters obeying
δimin≤ δi ≤ δimax
∀i ∈ {1,...,p}. (94)
In this case, the polytope has N = 2 p vertices.
Advanced Feedback Control – p. 1
o ytop c uncerta nty
Example:
M =
−1 + δ1 δ2 3
1 −2 + 2δ1 0
where
|δ1| ≤ 0, 5
|δ2| ≤ 0, 2.⇔
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 162/170
M 0 =
−1 0 3
1 −2 0
; N 1 =
1 0 0
0 2 0
; N 2 =
0 1 0
0 0 0
which corresponds to the polytopic description:
[M 1
|M 2
|M 3
|M 4] =
−0, 5 0, 2 3 −0, 5 −0, 2 3 −1, 5 0, 2 3 −1, 5 −0, 2 3
1 −1 0 1 −1 0 1 −3 0 1 −3 0
Advanced Feedback Control – p. 1
o ytop c uncerta nty
Yes, but why polytopic uncertainty?
Simply because it is easily handled through LMImachinery.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 163/170
For example, assume one wants to analyze the robuststability of a polytopic matrix A with two vertices A1
and A2. This matrix is robustly stable if
∃P = P T > 0 such that
AT
1
P + P A1 < 0 & AT
2
P + P A2 < 0.
The condition is only sufficient since only one uniqueP is considered and it is not a function of the
ucnertainty (one talks about quadratic stability).Advanced Feedback Control – p. 1
o ust m xt synt es s
Theorem 8 Consider an uncertain process model
P (s, τ ) and some LMI-region D. There exists aD-stabilizing state feedback K that ensures||F(P (s), K )||• < γ • ∀ • ∈ {∞; 2} if
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 164/170
∃{X = X T > 0; T = T T ; L} such that
Z ∞j
=Z
(P j
(s), X , L , , γ ∞
) < 0
Z 21j= Z (F(P j(s), X , L)) < 0
Z 22j= Z (F(P j(s), X , L , T )) < 0 ∀ j ∈ {1,...,N },
trace(T ) < γ 22M jD = M D(A j + B ju, X , L) < 0
In this event, K is given by K = LX −1. Advanced Feedback Control – p. 1
o ust m xt synt es s
There are two reasons why the condition is
conservative:
• Lyapunov shaping paradigm,
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 165/170
• Quadratic stability.
Avoiding the shaping paradigm is quite difficult butthere exist techniques that consider a matrix X (τ )(i.e. which is dependent on the uncertainty).
Advanced Feedback Control – p. 1
Robustness andH∞
Many authors consider that
H∞ approach belongs to
the realm of robust control whereas it is simplydisturbance rejection.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 166/170
The reason is as follows:Consider the uncertain matrix (Linear FractionalTransform (LFT)-based uncertainty):
A = A + B∆C (95)
where ∆ = ∆(II − D∆)−1 and ||∆|| ≤ ρ, with ∆
complex (this special LFT is called norm-boundeduncertainty).Problem: find the largest value of ρ such that A is
Hurwitz for all ∆ .This value is the so-called complexstabilit radius. Advanced Feedback Control – p. 1
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 167/170
ReferencesThere are so many (I used quite many in French but...)
• S. Boyd, L. El Ghaoui, E. Feron and V. Balakrishnan
Linear Matrix Inequalities in System and Control.
Volume 15 of SIAM Studies in Applied Mathematics, USA, 1994, still
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 168/170
the ultimate reference for LMI in control• P. Gahinet and P. Apkarian
A LMI approach To H∞ control.
International Journal of Robust and Nonlinear Control, Vol 4,
p.421-448, 1994 for the solution to dynamic H∞-problem
• M. Chilali and P. Gahinet
H∞ design with pole placement constraints.
IEEE Transactions on Automatic Control, Vol 41(3), p.358-367, 1996.,
for the pole placement in LMI-regions
Advanced Feedback Control – p. 1
References (2)
• P. Khargonekar, I. R. Petersen and K. Zhou,.
Robust stabilization of uncertain linear systems: Quadtric
stabilizability and H∞ control.IEEE Transactions on Automatic Control, Vol 35, p.356-361, 1990, for
the robust stability against norm-bounded uncertainty)
C. Scherer and S. Weiland.
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 169/170
• Lectures Notes DISC Course on Linear Matrix Inequalities in Control
available for downloading from the web, very general and elegant
approach covering many of the aspects of these slides and even much
more.
... and all the references therein.
See also the very good slides proposed by D. Henrion on his web page:
http://www.laas.fr/ henrion
Advanced Feedback Control – p. 1
Hoping you enjoyed theseslides, the control communitynow needs you to investigate
8/3/2019 Advanced Feedback Control
http://slidepdf.com/reader/full/advanced-feedback-control 170/170
many of the problems that arestill unsolved !
Advanced Feedback Control – p. 1