Вы находитесь на странице: 1из 48

INTERACTION ASSESSMENT IN LINEAR

MULTIVARIABLE PROCESSES USING DIRECTED


SPECTRAL DECOMPOSITION
Arun K. Tangirala
Dept. of Chemical Engineering, IIT Madras
Chennai, Tamilnadu, India

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

OUTLINE

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

OUTLINE
Motivation
Review
RGA, Dynamic RGA, Sensi0vity func0ons

Directed Analysis
Main results
Benchmark for interac0on assessment
Interac0on quan0ca0on

Simulation study
Concluding Remarks
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

INTERACTIONS

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

INTERACTIONS

eu1

u1
-Gc1

G11

y1
ey1

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

INTERACTIONS

eu1

u1
-Gc1

eu2

G21

G11
11

u2
G22

-Gc2

G12

Arun K. Tangirala (IIT Madras)

y1

y2

ey1

ey2

Lecture Series, University of Alberta

June 10, 2011

INTERACTIONS
Interaction: Effect felt in one loop due to changes (disturbances /
setpoint changes) in other loops (typical of all MIMO system)
eu1
eu2
Impact of interaction:
Reduces performance
Can lead to instabili0es
Not necessarily harmful!

u1
-Gc1

G21

G11
11

u2
G22

-Gc2

G12

y1

y2

ey1

ey2

Problems of interest:
1. Quantify interactions

2. Relate interactions to a performance metric

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

WHATS THE USE?


Controller Design
Design controllers for multivariable systems such that interactions are
at a minimum or even better, beneficial
For MIMO systems, two possible control congura0ons - (i) decentralized (mul0loop)

controllers and (ii) mul0variable controllers

Control Loop Performance Assessment


For a given multivariable closed-loop system, assess the strength of
interactions and relate it to performance
A typical performance metric is variance. Can we determine the contribu0ons of interac0ons

to variance?

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

MEASURING INTERACTIONS
Relative Gain Array (Bristol, 1966)

(yi /uj )all loops open

ij =

(yi /uj )all other loops closed except loop (y - u )


i
j steady state

= K KT

Dynamic RGA (Witcher & McAvoy, 1977, Other researchers, later years)

(s) = G(s) G(s)T


Loop under study remains open
Outputs of all other loops held at their set-points

Both measures assume perfect control


Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

INTERACTION MEASURES
Generalized Dynamic Relative Gain (Huang et al, 1993)

Actual controllers rather than perfect controllers are used

Loop decomposition method (Zhu and Jutan, 1996)

Takes into account all perturba0ons; Rela0ve interac0on, Absolute interac0on

Joint stationary representation approach (Seppala et al, 2002)

Use the mul0variate impulse response func0on (the H matrix in the VMA representa0on)

Semi-quan0ta0ve approach to interac0on

Performance RGA (Skogestad et al)

Deni0on based on exact factoriza0on of sensi0vity func0on; limited interpreta0on

Several other methods


Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

WHAT DO WE DESIRE?
The interaction measure should quantitatively correspond to a
performance metric (e.g., variance)

We should be able to compute from models (for controller design)

We should be able to estimate it from data (for performance assessment)

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

INTERACTION & DIRECTIONALITY


Interaction is a directed phenomenon (direction matters)
Effect felt (by an output) through indirect pathways arising due to
connections with other loops
Dierence between a SISO loop and a MIMO control system

The direct & indirect transfer functions play a key role in quantifying
interaction
Basic question:
What are the contributions of the indirect pathways to the variance
of a closed-loop output?

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

DIRECTED VARIANCE DECOMPOSITION


Variance decomposition in frequency domain: y2i
Set up

x=
so that

xx ()

y1

y1 y1 ()
u1 y1 ()
..
.
..
.
ym y1 ()
um y1 ()

ym

y1 u1 ()
u1 u1 ()
..
.
..
.
ym u1 ()
um u1 ()

u1

..
.
..
.

1
=
2

..
.
..

um

y1 ym ()
u1 ym ()
..
.
..
.
ym ym ()
um ym ()

yi yi () d

T
y1 um ()
u1 um ()
..
.
..
.
ym um ()
um um ()

Spectral Factorization (jointly stationary process driven by white-noise):

xx () = H()e H ()
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

WHAT DOES H CONTAIN? DTF


Directed Transfer Function (DTF)

hij() is the net transfer function from the (white-noise) innovations in xj to


the ith variable xi (in that direction)

ey1
ey2
eu1
eu2

y1
y2
u1
u2

The DTF (Saito and Harashima, Kaminski and Blinowska) is the


normalized hij() and is a non-parametric quantity by definition
Its estimation, however, is carried out using a vector auto-regressive (VAR)
modelling of the time-series
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

10

ESTIMATION OF DTF
Construct VAR / VMA model
p

x[k] =

r=1

Ar x[k r] + e[k]

OR

x[k] =

r=1

Hr e[k r] + e[k]

VAR models are easier to construct since LS es0ma0on methods can be used. Then,

H() = A

() =

h11 ()
h21 ()
..
.

...
...
..
.

h1m ()
h2m ()
..
.

hm1 () . . .

hmm ()


rj
Ar e
=
; A() = I

r=1

a
11 ()
a
21 ()
..
.

...
...
..
.

a
1m ()
a
2m ()
..
.

a
m1 () . . .

a
mm ()

Each element hij(): Total effect of the jth source on the ith variable
The transfer function hij() consists of direct and indirect components
ji ())

aij ()det(M
hij () = hD,ij () + hI,ij () ; hD,ij () =

det(A())
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

i = j
June 10, 2011

11

DIRECT(ED) ENERGY TRANSFERS


The total energy transfer is the sum of direct, indirect and an interference
terms
|hij ()|2 = |hD,ij ()|2 + |hI,ij ()|2 + 2|hD,ij ()||hI,ij ()|cos(D () I ())
Interferences occur due to phase dierences between direct and indirect transfers
They can be either construc0ve or destruc0ve depending on the phase dierence

The term |hii()|2 quantifies the fraction of energy received by xi[k] from
its own driving force (and due to unaccounted sources)
For analysis purposes, fix (or force) e = Inn

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

12

PATHWAYS FOR TRANSFER OF ENERGY


Total transfer function

Ej()

Xi()

Direct transfer function

Ej()

Xi()

Indirect transfer function

Indirect energy transfer

Direct energy transfer

Sejej ()

Sxixi()

Interference
effect
Interference effect
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

13

A 2X2 SYSTEM
eu1

u1
-Gc1

G21

G11

y1
ey1

Arun K. Tangirala (IIT Madras)

eu2

u2
G22

G12

-Gc2

y2
ey2

Lecture Series, University of Alberta

June 10, 2011

14

A 2X2 SYSTEM
eu1

u1
-Gc1

eu2

G21

G11

y1
ey1

u2
G22

G12

-Gc2

y2
ey2

Decompose spectrum of y1:

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

14

A 2X2 SYSTEM
eu1

u1
-Gc1

eu2

G21

G11

y1
ey1

u2
G22

G12

-Gc2

y2
ey2

Decompose spectrum of y1:


y1 y1 () = |hy1 y1 ()|2 + |hy1 u1 ()|2 + |hy1 y2 ()|2 + |hy1 u2 ()|2

Interaction eects

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

14

A 2X2 SYSTEM
eu1

u1
-Gc1

eu2

G21

G11

y1
ey1

Decompose spectrum of y1:

u2
G22

G12

-Gc2

y2
ey2

Due to past of y1 and


unaccounted variables

y1 y1 () = |hy1 y1 ()|2 + |hy1 u1 ()|2 + |hy1 y2 ()|2 + |hy1 u2 ()|2

Interaction eects

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

14

A 2X2 SYSTEM
eu1

u1
-Gc1

eu2

G21

G11

y1
ey1

Decompose spectrum of y1:

u2
G22

G12

-Gc2

y2
ey2

Due to past of y1 and


unaccounted variables

y1 y1 () = |hy1 y1 ()|2 + |hy1 u1 ()|2 + |hy1 y2 ()|2 + |hy1 u2 ()|2

Interaction eects
|hy1 u1 ()|2 = |hD,y1 u1 ()|2 + |hI,y1 u1 ()|2 + hIF,y1 u1 ()

Interaction eects
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

14

A 2X2 SYSTEM
eu1

u1
-Gc1

eu2

G21

G11

y1
ey1

Decompose spectrum of y1:

u2
G22

G12

-Gc2

y2
ey2

Due to past of y1 and


unaccounted variables

y1 y1 () = |hy1 y1 ()|2 + |hy1 u1 ()|2 + |hy1 y2 ()|2 + |hy1 u2 ()|2

Interaction eects
Feedback and interaction dependent
|hy1 u1 ()|2 = |hD,y1 u1 ()|2 + |hI,y1 u1 ()|2 + hIF,y1 u1 ()

Interaction eects
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

14

A 2X2 SYSTEM
eu1

u1
-Gc1

eu2

G21

G11

y1
ey1

Decompose spectrum of y1:

u2
G22

G12

-Gc2

y2
ey2

Due to past of y1 and


unaccounted variables

y1 y1 () = |hy1 y1 ()|2 + |hy1 u1 ()|2 + |hy1 y2 ()|2 + |hy1 u2 ()|2

Interaction eects
Feedback and interaction dependent
|hy1 u1 ()|2 = |hD,y1 u1 ()|2 + |hI,y1 u1 ()|2 + hIF,y1 u1 ()

Interaction eects

Idea is to arrive at an interaction and feedback invariant term


Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

14

DIRECT AND INDIRECT TRANSFER FUNCTIONS


eu1

u1
-Gc1

eu2

G21

G11

y1

u2
G22

G12

-Gc2

y2

ey1

hy 1 y 1 =
hy 1 y 2

1 + G22 Gc2

G12 Gc2
=

where

hD,y1 u1
hI,y1 u1
Arun K. Tangirala (IIT Madras)

;
;
=

ey2
G11 (1 + G22 Gc2 ) G12 G21 Gc2
hy 1 u 1 =
+

G12
hy 1 u 2 =

(1 + G11 Gc1 )(1 + G22 Gc2 ) G12 G21 Gc2 Gc1

G11 (1 + G22 Gc2 )


=
= G11 hy1 y1

G12 G21 Gc2


G12 G21 Gc2
=
=
hy y

1 + G22 Gc2 1 1
Lecture Series, University of Alberta

June 10, 2011

15

normalized
output
spectrum
into
componen
0
0e
Thetwo
LHS
of (34)
=H
(28)
u
S
S
e
trum
uy
uu
u
0
0
Relationship with Sensitivity Function u
an
interaction
and
feedback
invariant
(first
two
te
Suy Suu eu
trum of the outpu
of th
2
RHS i.e. (1 + |Gij ()| ) which of
depends
on theof
the
diagonal
The
VAR
model
for
the
process
in
terms
of
transfer
funcENSITIVITY
FUNCTION
THE
MATRIX
control
loopofpairing
and
(ii ) an interaction and fee
observe the connection
between
multivariate
sensiThe VAR
model
for
the
process
in
terms
transfer
functions
is
given
by
dependent (last term on RHS i.e. i ()) term w
ty function matrix S0 and Htions
matrix,
we begin
is given
by with
Yi,f ()sett
=

pairing
aswell as the controller
depended
on

I Gp y
ey
0
0
=
=
(29)
y
Syy Syu ey
I Gp y
ey
The
nicely interpreted
as the
= yif yif ()
=
=
=He
(28)
G
u canebe
=cLHSI of (34)
(29)
u
0
0
th
u
Suy Suu eu
trum
ofeuthe output of the i loop filtered by the i
Gc I
u
Thu
of1the
sensitivity functi
ofthediagonal
multiloop

Thus, i () repre
of th
y 1 I Gp
ey
func
e VAR model for the process in terms of transfer
= H e of the
(30)
filtered out
y
I=
Gp = ey
1
ns is given by
u
G
I
e
=
=
=
H
e
(30)
c
Yi,f () = 0 u Yii ()
Equ
S
()
u
G
I
e
c
ii
u


Assuming all output disturbances
to be white, by
virtue (34)wide
Equation
at
1
0
I Gp y Assuming
ey
all output
disturbances
toy be
white,
by virtue
of definition,
the
sensitivity
function
Syy
can beknownsitiv
widely
tha
=
() =
y()
()
=
(29)
i yi
if yif 0
0
2
related
to the multivariate
response
matrixfunction
the
sensitivity
function Syyfrequency
()|Scan
be
this
sitivity
ii ()|
Gc I
u of definition,
eu
By comparing
(28)response
and (30),
the multiloop
related to the H().
multivariate
frequency
matrix
t
this senwork, theon
out
sitivity (28)
function
given
Thus,
()multiloop
representssenthe contribution
to the
va
H().
and
(30),by
func
on the inverse
of
ithe

1
By comparing
of the filtered output.
y
I Gp sitivity
ey function given by
function. Thisloop
am
0
1
=
=
=He
(30)
Syy = (Imm + Gp Gc )
loop(31)
while keepin
u
Gc I
eu
0
1
Syy = (Imm +
Gp Gc ) (34) attracts(31)
Equation
some nice interpretation
Pro
Here
we consider
a multivariable
a deuming all output disturbances to be white, by
virtue
widely
known thatprocess
filteringGwith
inverse of th
p andthe
0
Proposition
1
controller
Gc .a de- the open
definition, the sensitivity function
can
be (diagonal)
sitivity
function
provides
loop system
Here weSconsider
a multivariable
process
Gp and
yy ()centralized
ted to the multivariate frequency
response
matrixcontroller
thisG
work,
the output spectrum filtering is implem
centralized
(diagonal)
c.
inverse of the diagonal

). By comparing (28) and (30), the multiloop senon the


matrix of the sens
G11This
G12amounts
G1mto opening up of that par
vity function given by
function.

loops closed.
G11 G12

G
loop while
other
1m
G
G
G
21 keeping
22 all
2m

Sy0i y

0
1

Gp =
(32)
Syy = (Imm + Gp Gc )
(31)
G
..
.. . .
..
0
G

2m
21 G22

equi
Syi yj = ij th elem
. .

.
.

Gp = .
(32)

.. Proposition
.. 1
=
..
S
. University
16
K. Tangirala
(IIT Madras)
Lecture
Series,
of
June equivalent
10, 2011
proces

e we Arun
consider
a multivariable
process G and
ade. Alberta
=

&

.
Gm1
Gm2 Gmm

normalized
output
spectrum
into
componen
0
0e
Thetwo
LHS
of (34)
=H
(28)
u
S
S
e
trum
uy
uu
u
0
0
Relationship with Sensitivity Function u
an
interaction
and
feedback
invariant
(first
two
te
Suy Suu eu
trum of the outpu
of th
2
RHS i.e. (1 + |Gij ()| ) which of
depends
on theof
the
diagonal
The
VAR
model
for
the
process
in
terms
of
transfer
funcENSITIVITY
FUNCTION
THE
MATRIX
control
loopofpairing
and
(ii ) an interaction and fee
observe the connection
between
multivariate
sensiThe VAR
model
for
the
process
in
terms
transfer
functions
is
given
by
dependent (last term on RHS i.e. i ()) term w
ty function matrix S0 and Htions
matrix,
we begin
is given
by with
Yi,f ()sett
=

pairing
aswell as the controller
depended
on

I Gp y
ey
0
0
=
=
(29)
y
Syy Syu ey
I Gp y
ey
The
nicely interpreted
as the
= yif yif ()
=
=
=He
(28)
G
u canebe
=cLHSI of (34)
(29)
u
0
0
th
u
Suy Suu eu
trum
ofeuthe output of the i loop filtered by the i
Gc I
u
Thu
of1the
sensitivity functi
ofthediagonal
multiloop

Thus, i () repre
of th
y 1 I Gp
ey
func
e VAR model for the process in terms of transfer
= H e of the
(30)
filtered out
y
I=
Gp = ey
1
ns is given by
u
G
I
e
=
=
=
H
e
(30)
c
Yi,f () = 0 u Yii ()
Equ
S
()
u
G
I
e
c
ii
u


Assuming all output disturbances
to be white, by
virtue (34)wide
Equation
at
1
0
I Gp y Assuming
ey
all output
disturbances
toy be
white,
by virtue
of definition,
the
sensitivity
function
Syy
can beknownsitiv
widely
tha
=
() =
y()
()
=
(29)
i yi
if yif 0
0
Multi-loop sensitivity
for
a
2
x
2
system:
2
related
to the multivariate
response
matrixfunction
the
sensitivity
function Syyfrequency
()|Scan
be
this
sitivity
ii ()|
Gc I
u of definition,
eu
By comparing
(28)response
and (30),
related to the H().
multivariate
frequency
matrix
t
this senwork,
theon
out

the0 multiloop
0
1sitivity
function
given
Sthe
Son
Thus,
by
()
representssento the
va
1 +(28)
Gc2 Gand
G
Gmultiloop
H().
(30),
func
of
ithe
22
12
c2
S
0= (I + G1
)1By=comparing
y1 y1contribution
y1 y2the inverse
=
G
yy
p sitivity
c
0
0
y
I Gp
ey function
Gby
1 +filtered
Gc1 G11output. 1
Sy2 y1 Sfunction.
Thisloop
am
given
21 Gc1 of0the
y 2 y2
=
=
=He
(30)
Syy = (Imm + Gp Gc )
loop(31)
while keepin
where

=
(1
+
G
G
)(1
+
G
G
)

G
G
G
G
u
Gc I
eu
0
11 c1
22 c21
12 21 c1 c2
Syy = (Imm +
Gp Gc ) (34) attracts(31)
Equation
some nice interpretation
Pro
Here
we consider
a multivariable
a deuming all output disturbances to be white, by
virtue
widely
known thatprocess
filteringGwith
inverse of th
p andthe
0
Proposition
1
controller
Gc .a de- the open
definition, the sensitivity function
can
be (diagonal)
sitivity
function
provides
loop system
Here weSconsider
a multivariable
process
Gp and
yy ()centralized
ted to the multivariate frequency
response
matrixcontroller
thisG
work,
the output spectrum filtering is implem
centralized
(diagonal)
c.
inverse of the diagonal

). By comparing (28) and (30), the multiloop senon the


matrix of the sens
G11This
G12amounts
G1mto opening up of that par
vity function given by
function.

loops closed.
G11 G12

G
loop while
other
1m
G
G
G
21 keeping
22 all
2m

Sy0i y

0
1

Gp =
(32)
Syy = (Imm + Gp Gc )
(31)
G
..
.. . .
..
0
G

2m
21 G22

equi
Syi yj = ij th elem
. .

.
.

Gp = .
(32)

.. Proposition
.. 1
=
..
S
. University
16
K. Tangirala
(IIT Madras)
Lecture
Series,
of
June equivalent
10, 2011
proces

e we Arun
consider
a multivariable
process G and
ade. Alberta
=

&

.
Gm1
Gm2 Gmm

normalized
output
spectrum
into
componen
0
0e
Thetwo
LHS
of (34)
=H
(28)
u
S
S
e
trum
uy
uu
u
0
0
Relationship with Sensitivity Function u
an
interaction
and
feedback
invariant
(first
two
te
Suy Suu eu
trum of the outpu
of th
2
RHS i.e. (1 + |Gij ()| ) which of
depends
on theof
the
diagonal
The
VAR
model
for
the
process
in
terms
of
transfer
funcENSITIVITY
FUNCTION
THE
MATRIX
control
loopofpairing
and
(ii ) an interaction and fee
observe the connection
between
multivariate
sensiThe VAR
model
for
the
process
in
terms
transfer
functions
is
given
by
dependent (last term on RHS i.e. i ()) term w
ty function matrix S0 and Htions
matrix,
we begin
is given
by with
Yi,f ()sett
=

pairing
aswell as the controller
depended
on

I Gp y
ey
0
0
=
=
(29)
y
Syy Syu ey
I Gp y
ey
The
nicely interpreted
as the
= yif yif ()
=
=
=He
(28)
G
u canebe
=cLHSI of (34)
(29)
u
0
0
th
u
Suy Suu eu
trum
ofeuthe output of the i loop filtered by the i
Gc I
u
Thu
of1the
sensitivity functi
ofthediagonal
multiloop

Thus, i () repre
of th
y 1 I Gp
ey
func
e VAR model for the process in terms of transfer
= H e of the
(30)
filtered out
y
I=
Gp = ey
1
ns is given by
u
G
I
e
=
=
=
H
e
(30)
c
Yi,f () = 0 u Yii ()
Equ
S
()
u
G
I
e
c
ii
u


Assuming all output disturbances
to be white, by
virtue (34)wide
Equation
at
1
0
I Gp y Assuming
ey
all output
disturbances
toy be
white,
by virtue
of definition,
the
sensitivity
function
Syy
can beknownsitiv
widely
tha
=
() =
y()
()
=
(29)
i yi
if yif 0
0
Multi-loop sensitivity
for
a
2
x
2
system:
2
related
to the multivariate
response
matrixfunction
the
sensitivity
function Syyfrequency
()|Scan
be
this
sitivity
ii ()|
Gc I
u of definition,
eu
By comparing
(28)response
and (30),
related to the H().
multivariate
frequency
matrix
t
this senwork,
theon
out

the0 multiloop
0
1sitivity
function
given
Sthe
Son
Thus,
by
()
representssento the
va
1 +(28)
Gc2 Gand
G
Gmultiloop
H().
(30),
func
of
ithe
22
12
c2
S
0= (I + G1
)1By=comparing
y1 y1contribution
y1 y2the inverse
=
G
yy
p sitivity
c
0
0
y
I Gp
ey function
Gby
1 +filtered
Gc1 G11output. 1
Sy2 y1 Sfunction.
Thisloop
am
given
21 Gc1 of0the
y 2 y2
=
=
=He
(30)
Syy = (Imm + Gp Gc )
loop(31)
while keepin
where

=
(1
+
G
G
)(1
+
G
G
)

G
G
G
G
u
Gc I
eu
0
11 c1
22 c21
12 21 c1 c2
Syy = (Imm +
Gp Gc ) (34) attracts(31)
Equation
some nice interpretation
Pro
Here
we consider
a multivariable
a deuming all output disturbances to be white, by
virtue
widely
known thatprocess
filteringGwith
inverse of th
p andthe
0
Proposition
1
controller
Gc .a de- the open
definition, the sensitivity function
can
be (diagonal)
sitivity
function
provides
loop system
Here weSconsider
a multivariable
process
Gp and
yy ()centralized
The multi-loop
output
sensitivity
iscontroller
identical
toc .thethe
output
sub-block
offiltering
H
ted to the multivariate
frequency
response
matrix
thisG
work,
output
spectrum
is implem
centralized
(diagonal)
inverse of the

). By comparing (28) and (30),


multiloop senon the
diagonal
matrix of the sens

the
0
0
G12
11
vity function given by
Sy1 y 1 Sy1 y2
h
hamounts
function.
yG
y1y 2 G1mtoopening up of that par
1 yThis
1

loops closed.
0
0
G11 =
G12
h
G
loop
while
all
other
1m
G
G

G
21 keeping
22
2m

h
S
S
Sy0i y
y 2 y1 y 2 y 2

0
1
y
y
y
y
2
1
2
2

Gp =
(32)
Syy = (Imm + Gp Gc )
(31)
G
..
.. . .
..
0
G

2m
21 G22

equi
Syi yj = ij th elem
. .

.
.

Gp = .
(32)

.. Proposition
.. 1
=
..
S
. University
16
K. Tangirala
(IIT Madras)
Lecture
Series,
of
June equivalent
10, 2011
proces

e we Arun
consider
a multivariable
process G and
ade. Alberta
=

&

.
Gm1
Gm2 Gmm

INTERACTION FACTORIZATION
Spectrum can now be expressed as
y1 y 1

0 2
|S11
| (1 + |G11 |2 )
0 2
+ |S11
| |1 + G22 Gc2 |2 |G12 |2

G
21
2
|G21 Gc2 |2 + 2|G11 |
|G
|cos
+
1
+
|G
|
c2
c2
G12 S2

Interaction eects 01 ()

The quantity 01 () vanishes at all if and only if G12 = 0


01 () = 0,

G12 = 0

The quan0ty
01 ()
can be nega%ve, posi%ve or zero depending on the interference term

The interference term can cancel out the remaining terms only at select (nite) frequencies but not
over a range of frequencies

Main Result:
1 ()

0 2
01 () = |S11
| 1 ()

|1 + G22 Gc2 |2 |G12 |2

Arun K. Tangirala (IIT Madras)

Factorization

G
21
2
|G21 Gc2 |2 + 2|G11 |
|G
|cos
+
1
+
|G
|
c2
c2
G12 S2

Lecture Series, University of Alberta

June 10, 2011

17

GraphicalGrepresentation
RAPHICAL REPRESENTATION
eu1
eu1

ey1

eu2

Direct
transfer

G11

u1
-Gc1

G21

G11

y1

u2
G22

G12

0
S11

y1

G12
-Gc2

S2

y2
-G21Gc2

ey1

Indirect
transfer

Loop 2

-Gc2

eu2

ey2

ey2

Signal flow graph of 2 2 system

Transmission of eects of noise source through


direct and indirect paths

G. Sebastian and A. K. Tangirala (IIT-M)


Arun K. Tangirala (IIT Madras)

Interaction Analysis
Lecture Series, University of Alberta

July 28, 2010

15 / 26

June 10, 2011

18

INTERACTION AND FEEDBACK INVARIANCE

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

19

INTERACTION AND FEEDBACK INVARIANCE


For any ith loop of a decentralized control loop
1
yi yi ()
0
2
|Sii ()|

Arun K. Tangirala (IIT Madras)

Interaction
Interaction
feedback dependent
feedback invariant


2
1 + |Gii ()|
+
i ()

Lecture Series, University of Alberta

June 10, 2011

19

INTERACTION AND FEEDBACK INVARIANCE


For any ith loop of a decentralized control loop
1
yi yi ()
0
2
|Sii ()|

Interaction
Interaction
feedback dependent
feedback invariant


2
1 + |Gii ()|
+
i ()

For a SISO loop, the interac0on term vanishes and the result is an iden0ty

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

19

INTERACTION AND FEEDBACK INVARIANCE


For any ith loop of a decentralized control loop
1
yi yi ()
0
2
|Sii ()|

Interaction
Interaction
feedback dependent
feedback invariant


2
1 + |Gii ()|
+
i ()

For a SISO loop, the interac0on term vanishes and the result is an iden0ty

LHS is the spectrum of the output filtered by its sensitivity function

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

19

INTERACTION AND FEEDBACK INVARIANCE


For any ith loop of a decentralized control loop
1
yi yi ()
0
2
|Sii ()|

Interaction
Interaction
feedback dependent
feedback invariant


2
1 + |Gii ()|
+
i ()

For a SISO loop, the interac0on term vanishes and the result is an iden0ty

LHS is the spectrum of the output filtered by its sensitivity function

It can be calculated both from the knowledge of transfer func0ons as well as from data

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

19

INTERACTION AND FEEDBACK INVARIANCE


For any ith loop of a decentralized control loop
1
yi yi ()
0
2
|Sii ()|

Interaction
Interaction
feedback dependent
feedback invariant


2
1 + |Gii ()|
+
i ()

For a SISO loop, the interac0on term vanishes and the result is an iden0ty

LHS is the spectrum of the output filtered by its sensitivity function

It can be calculated both from the knowledge of transfer func0ons as well as from data

The invariant term is useful in detecting and quantifying the extent of


interactions for design and assessment.
yif yif () = 1 + |Gii |2 = interaction eects are absent;
yif yif () 1 + |Gii |2 = interaction eects are detrimental / beneficial
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

19

FILTERING BY THE INVERSE OF SENSITIVITY

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

20

FILTERING BY THE INVERSE OF SENSITIVITY


Filtering the output by inverse of diagonal of sensitivity function
Opening up the corresponding loop while keeping other loops closed

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

20

FILTERING BY THE INVERSE OF SENSITIVITY


Filtering the output by inverse of diagonal of sensitivity function
Opening up the corresponding loop while keeping other loops closed
eu1

u1
-Gc1

eu2

G21

G11

u2
G22

-Gc2

G12

y1

y2

ey1

ey2

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

20

FILTERING BY THE INVERSE OF SENSITIVITY


Filtering the output by inverse of diagonal of sensitivity function
Opening up the corresponding loop while keeping other loops closed
eu1

u1
-Gc1

eu2

G21

G11

u2
G22

G12

y1

y2

ey1

ey2

Arun K. Tangirala (IIT Madras)

y1 [k]
Sy01 y1 (q 1 )
-Gc2

Y1 ()
Sy01 y1 ()

Lecture Series, University of Alberta

June 10, 2011

20

FILTERING BY THE INVERSE OF SENSITIVITY


Filtering the output by inverse of diagonal of sensitivity function
Opening up the corresponding loop while keeping other loops closed
eu1

u1
-Gc1

eu1

eu2

G21

G11

u2
G22

G12

y1

y2

ey1

ey2

Arun K. Tangirala (IIT Madras)

y1 [k]
Sy01 y1 (q 1 )

u1

eu2

G21

G11

-Gc2

Y1 ()
Sy01 y1 ()

Lecture Series, University of Alberta

u2
G22

-Gc2

G12

y1

y2

ey1

ey2

June 10, 2011

20

FILTERING BY THE INVERSE OF SENSITIVITY


Filtering the output by inverse of diagonal of sensitivity function
Opening up the corresponding loop while keeping other loops closed
eu1

u1
-Gc1

eu1

eu2

G21

G11

y1 [k]
Sy01 y1 (q 1 )

u2
G22

y1

y2

ey1

ey2

G21

G11

-Gc2

G12

u1

eu2

Y1 ()
Sy01 y1 ()

u2
G22

-Gc2

G12

y1

y2

ey1

ey2

To understand this, recall

Filtering the output of a SISO loop by the inverse of sensi0vity opens up the loop (discounts
for / cuts o the feedback)

Filtering the output vector of a MIMO system by the inverse of sensi0vity matrix is equivalent
to opening up all loops

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

20

ILLUSTRATION: 2 X 2 SYSTEM
0.40.2z 1
1z 1

0
0.40.2z 1
1z 1

Gp =

'()*+,-./

!%

!"#

$"#

1 () d = 3.5422
0

z 1
10.2z 1
z 1
10.3z 1

Gc =

!$
!

z 1
10.4z 1
z 1
10.1z 1

%"#

&

%
!
!%

!"#

$"#

2 () d = 3.9767
0

%"#

&

5!

6!

,+'/

Arun K. Tangirala (IIT Madras)

7!

$!!

$%!

'()*+,-./

%!

0.20.1z 1
1z 1

!
!%

!"#

$"#

1 () d = 3.1479
0

%"#

&

01/2-/*34
!%
%
!
!%

!"#

8/9:;*9/<;0<=;;:<%

$"#

2 () d = 3.9767
0

%"#

&

01/2-/*34

!"#

0.40.2z 1
1z 1

01/2-/*34
'()*+,-./

!$

01/2-/*34
!%
'()*+,-./

'()*+,-./

'()*+,-./

Gc =

8/9:;*9/<;0<=;;:<%
!"#

%!

5!

6!

7!

$!!

$%!

,+'/

Lecture Series, University of Alberta

June 10, 2011

21

ILLUSTRATION: 2 X 2 SYSTEM
0.40.2z 1
1z 1

0
0.40.2z 1
1z 1

Gp =

'()*+,-./

!%

!"#

$"#

1 () d = 3.5422
0

z 1
10.2z 1
z 1
10.3z 1

Gc =

!$
!

z 1
10.4z 1
z 1
10.1z 1

%"#

&

%
!
!%

!"#

$"#

2 () d = 3.9767
0

%"#

&

5!

6!

,+'/

7!

$!!

$%!

'()*+,-./

%!

0.20.1z 1
1z 1

!
!%

!"#

$"#

1 () d = 3.1479
0

%"#

&

01/2-/*34
!%
%
!
!%

!"#

8/9:;*9/<;0<=;;:<%

$"#

2 () d = 3.9767
0

%"#

&

01/2-/*34

!"#

0.40.2z 1
1z 1

01/2-/*34
'()*+,-./

!$

01/2-/*34
!%
'()*+,-./

'()*+,-./

'()*+,-./

Gc =

8/9:;*9/<;0<=;;:<%
!"#

%!

5!

6!

7!

$!!

$%!

,+'/

Interaction effects are negative (reduced variance) in frequency ranges at the cost of positive (larger variance)
effects in other frequency ranges
Lesser interaction effect in loop 1 is achieved at the cost of larger settling times in loop 2
Change in loop 2 controller produces interaction effects only in loop 1 (in the invariance domain)
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

21

NORMALIZED INTERACTIONS
Kc = 0.4 and

KI = 0.2 in both loops

Kc = 0.2

and

loop 1

loop 1
6

4
2
0

K=K0/|hy y |2
1 1

0.5

1.5

loop 2

0
0.5

1.5

2.5

0
0

0.5

1.5

loop 2
magnitude

zero line
2.5
3
K0 = absolute interaction
K/(1+|G112

magnitude

magnitude

magnitude

KI = 0.1 in loop 2

K=K0/|hy y |2
2.5 1 1
3
zero line
K0= absolute interaction
K/(1+|G |2)
11

4
2
0
2

0.5

frequency

1.5

2.5

frequency

The normalized interaction term should be as close to zero as possible


Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

22

CONCLUDING REMARKS

Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

23

CONCLUDING REMARKS
Quantification of interaction with respect to variance
Factoriza.on of interac.on term has been derived

Existence of an invariant term in the filtered domain has been established


Contributions of interaction to output variance involves indirect energy
transfer as well as interference
Interference term can cause posi0ve or nega0ve interac0on depending on the phase

dierence

Sensitivity function and the conditional sensitivity function are the keys
Road ahead: A single index for interaction, quantify margins, proof that
negative valued interaction is beneficial,
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

23

REFERENCES
Kaminski M. and Blinowska, K. (1991). A new method of the description of the information flow
in the brain structures. Biological Cybernetics, 65, 203-210.
Gigi, S. and Tangirala, A.K. (2010). Quantitative analysis of directional strengths in jointly
stationary linear multivariate processes. Biological Cybernetics, 103(2), 119-133.
Priestley, M. (1981). Spectral analysis and time series. Academic Press, London
Gevers, M. and Anderson, B. (1981). Representations of jointly stationary stochastic feedback
processes. Int. J. Control 33(5), 777-809.
Lutkepohl, H. (2005). New introduction to multiple time series analysis. Springer, New York.
Gigi, S. and Tangirala, A.K. (2010). Frequency-domain quantification of interactions in MIMO
systems using directed variance decomposition. In: 5th international symposium on Design,
Operation and Control of Chemical Processes, PSE-Asia, Singapore.
Seppala, C.T., Harris, T.J. and Bacon, D.W. (2002). Time series methods for dynamic analysis of
multiple controlled variables. Journal of Process Control, 12:257276.
Zhu, Z-X., and Jutan, A. (1996). Loop decomposition and dynamic interaction analysis of
decentralized control systems. Chemical Engineering Science., 51(12), 3325-3335.
Arun K. Tangirala (IIT Madras)

Lecture Series, University of Alberta

June 10, 2011

24

Вам также может понравиться