Beruflich Dokumente
Kultur Dokumente
Michele TARAGNA
Dipartimento di Automatica e Informatica
Politecnico di Torino
michele.taragna@polito.it
In matrix terms:
vR,1 iR,1 e1
vR,2 iR,2 e2
. = . [R ] + ..
.. ..
.
|{z}
vR,N iR,N o eN
| {z } | {z } | {z }
y e
Pna Pnb m
yj = i=1 ai yji + b u
i=1 i ji + ej , j = 1, . . . , N
T
o = a1 a2 ana b1 b2 bnb Rn , n = na + nb
T
y = yna +1 yna +2 yN RN , N = N na
y = o + e
" #
yna y1 una una +1nb
.. .. .. .. .. ..
= . . . . . . RN n
yN 1 yN na uN 1 uN nb
Experimental modeling: model building from experimental data 3
Politecnico di Torino - DAUIN M. Taragna
e2 e2
2 %e %
% %e2
2
e1 e1
e3 e3
Assumption: Be is symmetric with respect to the origin of RN
Experimental modeling: model building from experimental data 5
Politecnico di Torino - DAUIN M. Taragna
T)-1T 2
A = (
2
y
^
e
o
o 2
y Be T)-1T
A = ( EUS
y1 1
Parameter space n
y3 Measurement space N
Note that o EU S and that the distance between o and y is not greater than
Experimental modeling: model building from experimental data 6
Politecnico di Torino - DAUIN M. Taragna
EUI2
^
2
^
o
because EU S is the image of a 2m
^
EUI1
Experimental modeling: model building from experimental data 7
Politecnico di Torino - DAUIN M. Taragna
Evaluation of EUS
the vertices of M U S
= 1, . . . , 2N
are denoted by yk , k
Theorem: EU S = A[M U S ] = conv Ayk , k = 1, . . . , 2 Rn
N
y2 y7 y8
conv{ 1 , . . . , p } : y6 y5 T)-1T 2
A = (
EUS
convex hull of the set { 1 , . . . , p} y
Ay6
Ay2
^
Evaluation of EUIj
h m M
i
Theorem: EU Ij = j , j R 2
m
where j =
N
P
ajk yk sign ajk ,
EUS
2M
^
EUI2
k=1
^
2
^
M m
j = 2 j j , o
2m
^
A = ajk = (T )1 T ,
h i
= j = Ay
1m ^1 1M 1
^ ^
m
= min j = min (Ay)j =
Proof: j
EU S yM U S EUI1
N
P N
P N
P
= min ajk yk = min ajk yk = min ajk yk
y:|yi yi |k=1 y:yi yi k=1 y:yi yi yi +k=1
i=1,...,N i=1,...,N i=1,...,N
= yk if ajk > 0, or by yk = yk + if ajk < 0.
and such a minimum is achieved by yk
Since M U S = y Be is symmetric with respect to the data vector y , then
Description of ellipsoids
Let x be an ellipsoid in RN centered in xo :
n o
T
x = x RN : (x xo ) 1
x (x x o
)1
The form matrix x RNN is symmetric and positive definite it is invertible
The directions of the main axes of x are given by the eigenvectors ui of x ,
which are orthogonal because x is positive definite
p
The lengths of the semi-axes of x are given by i (x ),
where i (x ) is the i-th eigenvalue of x
u1
O1
x2 O2
u2 u1
u2
xo
:x
x1
z o = P xo Rn , z = P x P T Rnn
x2
x P z2
xo z
Pxo
x1 z1
x3 Space N Space n , n<N
2
Evaluation of EUS
T)-1T
A = ( 2
2
^
M EUS 2
y
EUI22
^
2
^
o
o
y Be2 T)-1T
A = (
2m
^
1m ^1 1M 1
^ ^
y1
y3 Measurement space N
EUI12
Parameter space n
n o
Theorem: EU S 2 = A y Be2 = Rn: ( )T T ( ) 2 Rn
is an ellipsoid in Rn with = Ay as symmetry center and (T )1 as form matrix
y2
T)-1T
A = ( 2
2
^
M EUS 2
y
EUI22
^
2
^
o
o
y Be2 T)-1T
A = (
2m
^
1m ^1 1M 1
^ ^
y1
y3 Measurement space N
EUI12
Parameter space n
Proof: by definition, EU S 2 is the linear mapping of M U S 2
= y Be2 through the matrix A:
2
2
n n T T 1 2
o
EU S = A y Be = R : ( Ay) [AA ] ( Ay)
But Ay = , A = (T )1 T and then:
n o
AAT = (T )1 T [(T )1 T ]T = (T )1 T [(T )1 ]T =
= (T )1 T [(T )T ]1 = (T )1
| {z }
I
2
Evaluation of EUIj
h m Mi
Theorem: EU Ij2 = j j , j + j = j , j R
| {z } | {z }
m M
j j
rh i
1
j = (T )
jj
2
2M EUS 2
^
EUI22
2
^
o
2m
^
1m ^1 1M 1
^ ^
EUI12
Experimental modeling: model building from experimental data 14
Politecnico di Torino - DAUIN M. Taragna
Theorem:
F P S EU S
F P S , then
Proof: if h i
(y ) Be y Be A A [y Be ] = EU S
h i
But A = (T )1 T = and then EU S .
Evaluation of FPS and PUIj
N
If F P S , then (y
) Be = e R : |ei | , i = 1, . . . , N
Ti : i-th row of
T
(y )i = y
i i , i = 1, . . . , N
n o
F P S = Rn : yi Ti , i = 1, . . . , N
2 2
Evaluation of FPS and PUIj
n o
Theorem: F P S 2 = Rn : ( )T T ( ) 2 2
2 = (y )T (y ) = k y k22 2
= fitting error between measured outputs and estimated outputs
a greater fitting error a smaller F P S 2 a lower uncertainty on parameters
h p p i h i
Moreover, P U Ij2 = j j 2 2 , j + j 2 2 = jm , jM R
| {z } | {z }
jm jM T2
q EUS 2
j = [(T )1 ]jj TM
PUI22
EUI22
T T
2 ^
T2
^
o
T 2m FPS 2
Optimal estimates
Definition: given an estimate , the estimate error E() is given by:
E() = sup k k
F P S
opt
Definition: an estimate is optimal if:
opt
E( ) E(), Rn
Central estimate:
C
h Ci C
.
= j , where j = (jm + jM ) 2, j = 1, . . . , n
the central estimate is optimal both if Be = Be2 and if Be = Be , since:
C
.
M
jm ) 2, j = 1, . . . , n
[
oj ] j ( j
LS
2
if Be = Be , the least squares estimate = (T )1 T y is central
LS
is optimal if Be = Be2 , but in general it is not optimal if Be = Be
2
Voltage Vz in V
8
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045
Position z in m
The static characteristic of the position-voltage transducer is nearly linear in the range
between 1.3 e 3.5 cm the characteristic can be linearly approximated by:
Vz = Kt z + Vo
Experimental modeling: model building from experimental data 20
Politecnico di Torino - DAUIN M. Taragna
In matrix form:
z V 1 e1
1 z,1
1 1 e2
z2 Vz,2
.. = .. .. + ..
. . . 2 .
zN Vz,N 1 eN
i.e., the estimation problem is in the standard form:
y =+e
where y RN , RN 2 , e RN and the unknown is R2
Using the Least Squares estimation algorithm:
T 1 T
= A" y , #with A
" = ( ) #
1 1.8194 103
= =
2 2.2791 102
1 2
Kt = = 549.62 V/m, Vo = = 12.526 V
1 1
Experimental modeling: model building from experimental data 22
Politecnico di Torino - DAUIN M. Taragna
Evaluation of the Estimate Uncertainty Intervals EUIj
N
N N
e Be= e R : |ei | , i = 1, . . . , N , = 5 104
h m M
i
EU Ij = j = min j , j = max j , j = 1, 2
EU S EU S
m PN
k=1 ajk [yk sign (ajk )]
j = min j =
EU S
M PN m
j = max j = k=1 ajk [yk + sign (ajk )]
= 2 j j
EU S
h m Mi
1 , 1 = [1.6996 103 , 1.9392 103 ]
h m Mi
2 , 2 = [2.2291 102 , 2.3291 102 ]
h i h M m
i
Ktm , KtM = 1/1 , 1/ 1 = [515.67, 588.36] V/m
h i h M m m M
i
Vom , VoM = 2 /1 , 2 /1 = [13.703, 11.495] V
Envelope of the static characteristics of models whose parameters are taken as the
extremes of the Estimate Uncertainty Intervals EU Ij , j = 1, 2
Position transducer
10
Experimental data (z, Vz)
z min(M * V , m * V ) + m
8 1 z 1 z 2
z = 1 * Vz + 2
z max(M 1
* Vz , m
1
* Vz) + M
2
6
2
Voltage V in V
z
linearity interval
10
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
Position z in m
Evaluation of the Parameter Uncertainty Intervals PUIj
n o
dim()
F P S = R : yi [ ]i , i = 1, . . . , N
h i
P U Ij = min j , max j EU Ij , j = 1, 2
F P S F P S
The extremes of P U Ij , j = 1, 2, are solutions of the linear programming problems
min j = min cT
" # " #
F P S M b
y
M= , b= + , c = j -th column of I22
max j = min (c)T y
F P S M b
h i
m
1 = min ,
1 1 M
= max 1 = [1.7909 103
, 1.8484 103
]
F P S F P S
h i
m
2 = min 2 , M
2 = max 2 = [2.2596 102
, 2.2807 102
]
F P S F P S
m M h i
Kt , Kt = 1/ M 1 , 1/ m
1 = [541.01, 558.38] V/m
m M h M m i
Vo , Vo = 2 /1 , m 2 / M
1 = [12.735, 12.225] V
2
Voltage V in V
z
linearity interval
10
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
Position z in m
Feasible Parameter Set F P S (continuous line) and set of estimates given by the
extremes of Parameter Uncertainty Intervals P U Ij , j = 1, 2
Feasible Parameter Set FPS
0.0229
0.0228
0.0228
LS
(y)
0.0227
2
0.0227
0.0227
0.0226
0.0226
1.79 1.8 1.81 1.82 1.83 1.84 1.85
3
1 x 10
0.0228
0.0228
LS
(y)
0.0227
2
0.0227
C(y)
0.0227
0.0226
0.0226
1.79 1.8 1.81 1.82 1.83 1.84 1.85
3
1 x 10
Essential references
F. C. Schweppe, Uncertain Dynamics Systems. Englewood Cliffs, NJ: Prentice Hall, 1973.
M. Milanese, R. Tempo, A. Vicino (editors), Robustness in Identification and Control. New York:
Plenum Press, 1989.
M. Milanese, A. Vicino, Optimal estimation theory for dynamic systems with set membership
uncertainty: an overview, Automatica, vol. 27, no. 6, pp. 9971009, 1991.
Special Issue on System Identification for Robust Control Design, IEEE Transactions on
Automatic Control, vol. AC-37, no. 7, pp. 899974, 1992.
R. S. Smith, M. Dahleh (editors), The Modeling of Uncertainty in Control Systems, vol. 192 of
Lecture Notes in Control and Information Sciences. London, UK: Springer-Verlag, 1994.
M. Milanese, J. Norton, H. Piet-Lahanier, E. Walter (editors), Bounding Approaches to System
Identification. New York: Plenum Press, 1996.
J. R. Partington, Interpolation, Identification, and Sampling, vol. 17 of London Mathematical
Society Monographs New Series. New York: Clarendon Press - Oxford, 1997.
A. Garulli, A. Tesi, A. Vicino (editors), Robustness in Identification and Control, vol. 245 of
Lecture Notes in Control and Information Sciences. Godalming, UK: Springer-Verlag, 1999.
J. Chen, G. Gu, Control-Oriented System Identification: An H Approach. New York: John
Wiley & Sons, Inc., 2000.