Vous êtes sur la page 1sur 9

Outline

1. Motivation
2. Tools
Interval algebra

Bounding regions

Zonotopes
DC-functions

J.M Bravo, T. Alamo


and E.F. Camacho

3. Robutst MPC
4. Identification
5. State estimation
6. Fault detection
E.F. Camacho EECI MPC course , Paris

Why region bounding in MPC?


Need to compute regions for stability and
robustness:

Random error
e~N(0,)

Invariant terminal regions,


Robust control invariants,
Uncertain evolution sets,
Uncertain parameter sets, ...

Measurements
error
Bounded error

Difficulty in computing regions for long horizons


and nonlinear systems.
Bounding by approximate regions easier to
compute (although somewhat conservative)
E.F. Camacho EECI MPC course , Paris

|e|<=

E.F. Camacho EECI MPC course , Paris

Mathematical Tools

Interval
Arithmetic

DC
Functions

Zonotope
Method

Mathematical
Mathematical
Tools
Tools

Interval
Interval Arithmetic
Arithmetic
Defines operations with interval variables

Methods
Methodsbased
basedon
onBounded
BoundedParameters
Parametersand
andBounded
BoundedError
Error
System
Identification
Robust
Predictive
Control

State
Estimation

Sum, substraction, multiplication,


Applications
Applications

Powerful tool to compute the range of a function


with interval variables

Fault
Diagnosis and
Detection
E.F. Camacho EECI MPC course , Paris

E.F. Camacho EECI MPC course , Paris

Mathematical Tools

Mathematical Tools

Interval
Interval Arithmetic
Arithmetic

Interval
IntervalArithmetic
Arithmetic

Interval: X = [a, b] is the set: { x : a x b} .


I It is the set of all intervals.
Unitary interval [-1, 1].
Box: Vector with interval components.
Unitary box Bn : Box with unitary intervals.
Interval Arithmetic: Arithmetic defined on interval.

n
Range of a function f : R n R on a set XR
:
f ( X ) = { f ( x) : x X }
n
Inclusion function: A function F : I I
n
inclusion function of f : R R if :
.
f ( X ) F ( X ), X I n

[ a , b ] + [c , d ] = [ a + c , b + d ]
a,b,c,d R
[ a , b] [c , d ] = [a d , b c]
[ a , b] [c , d ] = min ( a c , a d , b c , b d ) , max ( a c , a d , b c , b d )
[ a , b] [c , d ] = [ a , b] [1 d , 1 c ] , if 0 [c , d ]
E.F. Camacho EECI MPC course , Paris

Examples:

Natural Interval Extension: replace each real variable by


its corresponding interval and apply the interval
arithmetic rule.
E.F. Camacho EECI MPC course , Paris

Mathematical Tools

Mathematical Tools

Interval
IntervalArithmetic
Arithmetic

Interval
IntervalArithmetic
Arithmetic

Wf ( X ) = X + X = [1, 2] + [1, 2] = [ 2,6]


2

g ( x ) = exp( x

Fundamental Theorem (Moore66). The Natural Interval


Extension of a real function is a inclusion function.

x [1, 2] X

f ( x ) = x2 + x

is a

sin( x )

Wg ( X ) = exp( X

sin( X )

Interval Extension of the Medium Value Theorem:

x [1, 2] X
sin([1,2] )

) = exp([1, 2]

f ( X ) f ( c ) +W( x f ( X ) ) ( X c )

where W( x f ( X ) ) is a interval matrix that bounds the


gradient:
x f ( X )

= [ 3.011,15.12]

E.F. Camacho EECI MPC course , Paris

Mathematical Tools

10

Mathematical Tools

Interval
IntervalArithmetic
Arithmetic

Zonotopes
Zonotopes

Zonotope is a geometric figure. It


can be used to bound model
parameters o system states.

Pros:
An efficient way to compute Set-Value functions.
Guaranteed computation of ranges.

The Zonotope arithmetic provides


sum, subtraction, intersection,
operations with zonotopes.

Cons:
Overestimation.

f ( x) = x x

E.F. Camacho EECI MPC course , Paris

x [1, 2]

Wf ( x ) = [1, 2] [1, 2] = [ 1,1]


E.F. Camacho EECI MPC course , Paris

Powerful tool to compute the range of


a function and the system evolution
11

E.F. Camacho EECI MPC course , Paris

12

Mathematical Tools

Mathematical Tools

Zonotopes
Zonotopes

Zonotopes
Zonotopes

n
nxm
Given a vector p R and a matrix H R
, It is denominated
zonotope of order m to the set:

Z = x : b R m with b

1 and x = p + Hb

Minkowski sum of two sets: A,B A B= {a+b : a A, b B}

Interpretations:

Z = p HB m where B m is the unitary interval in R m .


0 1 0 2
Z=
B
0 0 1

Affine transformation of the unitary hypercube.

Minkowski sum of the segments defined by the columns of the


matrix.

E.F. Camacho EECI MPC course , Paris

0 1
Z=
0 1

13

1 3
B
0

-1

0 4
B
1

14

Zonotopes-Strip Intersection:

B n

p [H

E.F. Camacho EECI MPC course , Paris

0 1
Z=
0 1

Zonotopes
Zonotopes

Sum of zonotopes:

-1

Mathematical Tools

Zonotopes
Zonotopes

p HB r

-1 2
B
1

0 1
Z=
0 1

E.F. Camacho EECI MPC course , Paris

Mathematical Tools

Examples:

Z = p HB r

]B r + n

15

E.F. Camacho EECI MPC course , Paris

Mathematical Tools

Mathematical Tools

Zonotopes
Zonotopes

Zonotopes
Zonotopes

Zonotopes-Strip Intersection:

F = x : cT x d 1

Z = p HB r

E.F. Camacho EECI MPC course , Paris

Zonotopes-Strip Intersection:

16

F = x : cT x d 1

Z = p HB r

17

E.F. Camacho EECI MPC course , Paris

18

Mathematical Tools

Mathematical Tools
Zonotopes
Zonotopes

Zonotopes
Zonotopes

Zonotopes-Strip Intersection:

F = x : cT x d 1

Z = p HB r

Zonotopes-Strip Intersection:
r
Objective: Bound the intersection of zonotope Z = p H B and strip:

F = x : cT x d 1

Z ( ) = p ( ) + H ( ) B r +1

Given Z , F and a vector , it is defined:

p ( ) = p + ( d c T p )

H ( ) = ( I c T ) H

then: Z F Z ( ) = p ( ) + H ( ) B

Compute *that minimizes the volume of the intersection zonotope.

r +1

N (n ,r )

Volume Z ( ) =

i =1

2 n 1 c T det [ Ai ] +

N (n ,r )

i =1

2 n det [ Bi

vi ] v i
T

19
E.F. Camacho EECI MPC course , Paris

19

E.F. Camacho EECI MPC course , Paris

Mathematical Tools

Mathematical Tools

DC
DCFuntions
Funtions

Zonotopes
Zonotopes

f ( x ) = x 3 + x 2 + 1,

A DC function is a function
compoused of the difference
of two convex funtions.

Range of a function:

Using zonotopes and the Interval Extension of the Mean Value


(.) can be obtained:
Theorem, a boundding operator

f ( X ) ( X )

Where X and ( X ) are zonotopes

21

g ( x) = x3 + x 2 + 1 + 2 x 2
h( x) = 2 x 2

f ( x) = g ( x) h( x)

E.F. Camacho EECI MPC course , Paris

DC
DCFuntions
Funtions

DC
DCFuntions
Funtions

f ( x) = g ( x) h( x)

Range DC function

)
f ( x ) = g m ( x ) h( x )

g ( x)

f ( x) = g ( x) h( x)

g ( x)
g m (x)

h( x)

E.F. Camacho EECI MPC course , Paris

22

Mathematical Tools

Mathematical Tools

Range DC function

x [ 1,1]

f ( x) = x3 + x 2 + 1 + 2 x 2 2 x 2

A bound of the range of a


DC function can be obtained
by the linearization of the
convex functions

E.F. Camacho EECI MPC course , Paris

20

23

h( x)

E.F. Camacho EECI MPC course , Paris

24

Mathematical Tools

Mathematical Tools

DC
DCFuntions
Funtions

DC
DCFuntions
Funtions

(
f ( x) = g ( x) hm ( x)
Range DC function

(
f (x)

f ( x) = g ( x) h( x)

Range DC function

f ( x) = g ( x) h( x)
)
f (x)

g ( x)

)
(
f ( X ) min( f ( x)), max( f ( x))

h( x)

hm (x)
Vertices evaluation

E.F. Camacho EECI MPC course , Paris

25

E.F. Camacho EECI MPC course , Paris

Robust Predictive Control

26

Robust Predictive Control

Problem definition
Given a system, design predictive
controllolers with guaranteed stability
and moderated computational cost.
Consider models of the system with
bounded uncertainties and
0.4
parameters.
0.2

Proposed solutions

Key idea

X%
N

X%
4
X%
3

uk

X%
2

SYSTEM

xk

xk +1 = f ( xk , u k , wk )

0
-0.2

xk X , uk U , wk W

x2

-0.4
Controllers that use Interval Arithmetic
-0.6
and the Zonotope method to bound -0.8
-1
the uncertain evolution of the system.-1.2-0.2 -0.15-0.1 -0.05 0

0.05 0.1 0.15

x1

E.F. Camacho EECI MPC course , Paris

27

E.F. Camacho EECI MPC course , Paris

28

Robust Predictive Control

Robust Predictive Control


Key idea

Reachable sets in k:

Bounding operator (, , ) such that:

Xi +1 = f ( Xi , ui , W ) , X0 = xk

XN

X3

u ( N 1| k )

f ( X ,U ,W ) ( X ,U ,W )

X2
X1
xk

u (0 | k )

Approximated Reachabe sets:

u (1 | k )

X%i +1 = X%i , ui , W , X%0 = xk

Reachable sets:
X j +1 = f ( X j , u ( j 1 | k ), W )

E.F. Camacho EECI MPC course , Paris

X 0 = xk
29

Xi X%i
E.F. Camacho EECI MPC course , Paris

30

Robust Predictive Control

Robust Predictive Control


Key idea

Application to a CSTR

Bounding operators:

X%N

Interval Arithmetic

Zonotope method

Given a model of a chemical reactor,


Designe a MPC that stabilize the plant.
dCA q
E
= ( CAf CA ) k0 exp
CA
dt V
RT

X%3

DC functions

xk

X%
1

Bounding Operator:
Approximated Reachable sets:

dT q
UA
H k0
E
= (T f T )
exp
(Tc T )
CA +
dt V
V Cp
Cp
RT

X%2

f ( X ,U ,W ) ( X ,U ,W )
~
~
~
X j +1 = ( X j , u j , W ) X 0 = xk

E.F. Camacho EECI MPC course , Paris

Additive uncertainty in the states has been


considered.

31

System Identification

E.F. Camacho EECI MPC course , Paris

System Identification

uk

Problem definition
Given a parametric model, a set of
input/output measurements and a bounded
error in the measurements.
Goal. Obtain a guaranteed bound of the set
of parameters consistent with the previous
hypothesis.

SYSTEM

yk

HYPOTHESIS

OBJECTIVE

A parametric model of a
system

Proposed solutions

What is the set of consistent


parameters?

A set of measurements

Use zonotopes and Strip intersections to


bound the parameters.
Use DC functions to discard parameters that
are not consistent.

A bounded error of each


measurement
Iterative algorithm

E.F. Camacho EECI MPC course , Paris

33

System Identification

E.F. Camacho EECI MPC course , Paris

34

System Identification

Problem definition

F( y

Problem definition

CSFAk

, rk )

yk rkT k

CSFAk

Proposed solution

Proposed solution

Use zonotopes to bound the parameters


Use zonotopes and Strip intersections to
improved the bound the parameters.
E.F. Camacho EECI MPC course , Paris

32

Use zonotopes to bound the parameters


Use zonotopes and Strip intersections to
improved the bound the parameters.
35

E.F. Camacho EECI MPC course , Paris

36

System Identification

System Identification
AFSSk

F( y

Problem definition

, rk )

yk r k
T
k

Problem definition (Nonlinear)


yk = f ( rk , ) + ek

{( y1 r1 ),( y2

CSFAk
F( y

, rk )

, rk )

FSS = {

CSFAk

Proposed solution

,k = 1,...,N}

Proposed solution

Use zonotopes to bound the parameters


Use zonotopes and Strip intersections to
improved the bound the parameters.

R n : y k - f(rk ,)

r2 ),...,( y N rN )}

ek E = {e R : e }

y k rkT k

C SF Ak I F( y

FSS

Use DC functions to obtain convex


relaxations

C S F Ak +1

C S F Ak

E.F. Camacho EECI MPC course , Paris

37

E.F. Camacho EECI MPC course , Paris

38

System Identification

System Identification
AFSSk

Problem definition (Nonlinear)


yk = f ( rk , ) + ek

{( y1 r1 ),( y2

r2 ),...,( y N rN )}

ek E = {e R : e }

r2 ),...,( y N rN )}

ek E = {e R : e }

FSS = {

Use DC functions to obtain convex


relaxations

39

Convex
relaxation of
constraints

AFSSk

,k = 1,...,N}

Proposed solution

E.F. Camacho EECI MPC course , Paris

R n : y k - f(rk ,)

FSS

,k = 1,...,N}

Proposed solution

{( y1 r1 ),( y2

AFSSk

Convex
relaxation of
constraints

Problem definition (Nonlinear)


yk = f ( rk , ) + ek

R n : y k - f(rk ,)

FSS


FSS = {

AFSSk

Separating plane

Use DC functions to obtain convex


relaxations

E.F. Camacho EECI MPC course , Paris

40

State Estimation

System Identification
AFSSk

Problem definition (Nonlinear)


yk = f ( rk , ) + ek

{( y1 r1 ),( y2

r2 ),...,( y N rN )}

ek E = {e R : e }
R n : y k - f(rk ,)

Convex
relaxation of
constraints

AFSSk

FSS = {

Problem definition

FSS

,k = 1,...,N}

Proposed solutions
Use the Zonotope method to bound the
system evolution and the states that are
consistent with the measurements.
Use DC functions to bound the evolution of
the states.

Proposed solution

Use DC functions to obtain convex


relaxations

E.F. Camacho EECI MPC course , Paris

Given a model, a set of input-output


measurements and a bounded error of the
measurements.
Goal. Obtain a guaranteed bound of the set of
states that are consistent with the model,
measurements and bounded error.

AFSS k +1
41

E.F. Camacho EECI MPC course , Paris

42

State Estimation

State Estimation

Problem definition

Problem definition

X%k 1

X%k 1

X k = X%k 1 , W f X%k 1 , W

X%k 1

Proposed solutions

Proposed solutions
Use zonotopes and strips and
zonotopes intersections to
obtain the states.
Use DC functions.

Use zonotopes and strips and


zonotopes intersections to obtain the
states.
Use DC functions.

E.F. Camacho EECI MPC course , Paris

43

E.F. Camacho EECI MPC course , Paris

State Estimation

State Estimation

Problem definition

Problem definition

X%k 1

X%k 1

X k = X%k 1 , W f X%k 1 , W

X%k 1

e
yk

X%k X k X

Proposed solutions
Use zonotopes and strips and
zonotopes intersections to
obtain the states.
Use DC functions.

e
yk

X k = X%k 1 , W f X%k 1 , W

44

X%k 1

e
yk

X%k X k X

Proposed solutions
Use zonotopes and strips and
zonotopes intersections to
obtain the states.
Use DC functions.

X%k 1

X%k

X%k 1

E.F. Camacho EECI MPC course , Paris

45

State Estimation

E.F. Camacho EECI MPC course , Paris

46

Fault detection

Problem definition

Problem definition

Given a model of the system, a set of


input-output measurements and a
bounded error of the measurements.
Goal. Detect possible faults of the
system and identify their origin.

Proposed solutions

Proposed solution

Use zonotopes and strips and


zonotopes intersections to
obtain the states.
Use DC functions.

E.F. Camacho EECI MPC course , Paris

Apply developed bounded error


method of system identification and
state estimation.
47

E.F. Camacho EECI MPC course , Paris

e
yk

X k = X%k 1 , W f X%k 1 , W

X%k 1

X k = X%k 1 , W f X%k 1 , W

48

Fault detection

uk

SYSTEM

References

yk
Robust MPC of constrained nonlinear systems based on interval arithmetic
IEE Proceedings - Control Theory and Applications, (2005)
Robust MPC of constrained discrete-time nonlinear systems based on
approximated reachable sets Automatica (2006)

Are the measured sequence of inputs U and


outputs Y consistent with the model of the
faultless system?
If the measurements are inconsistent with
the model of the faultless system, the
existence of a fault is proved concluding the
fault detection task.

E.F. Camacho EECI MPC course , Paris

49

Thanks
E.F. Camacho EECI MPC course , Paris

51

Bounded error identification of systems with time-varying parameters.


IEEE Transactions on Automatic Control (2006)
An algorithm for bounded-error identification of nonlinear systems based on DC
Functions Automatica (2008)
Guaranteed state estimation by zonotopes Automatica (2005)
A set-membership state estimation algorithm based on DC programming
Automatica (2008)
Robust fault detection using zonotope-based set-membership consistency test
International Journal of Adaptative Control and Signal Processing (2008)

E.F. Camacho EECI MPC course , Paris

50

Vous aimerez peut-être aussi