6.3 LPV model identiation of the mass-spring-damper system
7.1.1 Nonlinear and linearized dynami models
The rst degree of freedom (# 1) of the arm is used for the vertial motion. The
seond and third ones (# 2 and # 3) are linked to the relative position of the
target enter of mass (blue ross in Figure 7.2b) and the laser spot (red ross in
Figure 7.2b). Forthis validationstep, the studied motionis restrited to these two
lastdegrees of freedom (# 2 and #3).
Asexplainedpreviously,the spotpositionmeasurementsaremadewiththe help
of a CCD amera. In pratie, this amera must be loated so that it does not
disturb the surgeon and his sta. Furthermore, it is diult to move during the
operation mainly for safety reasons. These pratial onditions highly redue the
workingeldoftherobot. Beauseoftheseonstraints,theuseofaglobaltehnique
[135℄requiringapersistentexitationoftheinputsaswellastheshedulingvariables
is not oneivable. On the ontrary, a loal approah [135℄ seems to be well-suited
for the LPV model identiationof suh asystem.
The
x
andy
positions of the end-eetoran bewrittenfrom the geometrimodel,resultingin the nonlinear measurement equation
z = g(q)
,i.e.,z 1 =(ℓ 1 − 2
3 ℓ 3 1 v 2 1 ) cos(φ 1 ) − ℓ 2 1 v 1 sin(φ 1 ) + (ℓ 2 − 2
3 ℓ 3 2 v 2 2 ) cos(φ 12 ) − ℓ 2 2 v 2 sin(φ 12 )
(7.6a)
z 2 =(ℓ 1 − 2
3 ℓ 3 1 v 2 1 ) sin(φ 1 ) + ℓ 2 1 v 1 cos(φ 1 ) + ℓ 2 2 v 2 cos(φ 12 ) + (ℓ 2 − 2
3 ℓ 3 2 v 2 2 ) sin(φ 12 )
(7.6b)
where
φ 12 = φ 1 + φ 2 + 2ℓ 1 v 1
.In order tox the struture of the globalLPV modelrequired by the tehnique
introdued in Setion 5.5, a standard Jaobian linearization an be applied to the
generalized seond order model given in Eq. (7.4). More speially, for a set of
working points
(q 0 , q ˙ 0 )
, weget3M(p)¨ q = D(p) ˙ q + K(p)v + Hu,
(7.7)where
p = p ( q 0 , q ˙ 0 )
isthe sheduling variable vetor and whereM(p(q 0 , q ˙ 0 )) = M (q 0 ),
(7.8a)D(p(q 0 , q ˙ 0 )) = ∂ F (q 0 , q ˙ 0 )
∂ q ˙ ,
(7.8b)K(p(q 0 , q ˙ 0 )) = ∂ F (q 0 , q ˙ 0 )
∂q − M −1 (q 0 ) M (q 0 )
∂q M −1 (q 0 ) F (q 0 , q ˙ 0 ),
(7.8)are the inertia, the damping and the stiness of the linearized modelrespetively.
Notie that
H
does not depend on the sheduling variablep
. In this work, wefous onthe identiationofa modelwhihinludesthe variability of the behavior
with respet to the positions
φ k
,k = 1, 2
. Then, the other phenomena an benegleted and, onsequently 4
,
q 0 =
φ ⊤ 0 0 1×n v ⊤
and
q ˙ 0 = 0 n q ×1
. Like for most ofthe methods onsidered inthe literature, the inertia matrix
M(p)
is inverted. Thisproperty is generally used in the literature (see, e.g., [54℄) leading to the following
loalequation
¨
q = M −1 ( p ) D ( p ) ˙ q + M −1 ( p ) K ( p ) q + M −1 ( p ) Hu .
(7.9)By looking loser at the equations available in [68℄, it is lear that the matrix
M
is an ane funtion of
cos(φ 2 )
andsin(φ 2 )
. Then, the sheduling variable an behosen as
cos(φ 2 )
, i.e.,p = cos(φ 2 )
. Now, the inner matries an be modeled [68℄asfollows,
M(p) = M 0 + M 1 p,
(7.10)while the other matries
D
andK
are onstant. Notie that this angular positionis easy to measure on a exible robot beause an enoder is generally loated at
the motor side of the joints. This availabilityis paramount when the experimental
3. Inthefollowing,theequilibriumvaluesareomittedin ordertoshortenthenotations.
4. This hoie means that the Coriolis eets are negleted. The validity of this assumption
modeling of the LPV model is onsidered. So, the matrix inversion involved in
Eq. (7.9)leadstomatries
M −1 (p)D(p)
,M −1 (p)K(p)
andM −1 (p)H
whihsatisfya frational dependeny on the shedulingvariable
p = cos(φ 2 )
. This hint justiesthe use of alinear frationalLPV desription(see Eq. (2.15))of thesystem instead
of a more standard ane LPV model(see Eq. (2.21)). Now, by onsidering
χ =
q ⊤ q ˙ ⊤ ⊤
∈ R 8
(7.11)as state vetor, the following loal linearized dark-gray-box state-spae model an
bededued 5
˙
χ = A 0 (p)χ + B 0 (p)u
(7.12)with
A 0 ( p ) =
0 n q ×n q I n q M −1 (p)K(p) M −1 (p)D(p)
(7.13a)
B 0 (p) =
0 n q ×n u M −1 (p)H
.
(7.13b)Forthisstate-spaeform,thestateequationonlydependson
φ 2
6
whereastheoutput
equation depends onboth
φ 1
andφ 2
. As shown in Eq. (7.6), the output equationsImageplane
rigidarms
exiblearms
Figure 7.3: Flexible manipulator(gray) and virtual rigid manipulator (blak) with
the same imagepositionof the end-eetor.
are highly nonlinear with respet to
φ 1
andφ 2
. This omplexity an be reduedby using eiently the Jaobian of the rigid geometri model (see Eq. (7.6) for a
denition of
g
), i.e.,J(φ 0 ) = ∂ g
φ ⊤ 0 0 1×n v ⊤
∂φ .
(7.14)Exept for the singular positions, this Jaobian is invertible and it is possible to
dene a new measurement vetor
α = α 0 + J −1 (φ 0 )(z − z 0 )
(7.15)5. Inthesequel,
z •
,φ •
,v •
,q •
,...,areonsideredasvariationsaroundanominalvalue.6. Thesystem behavioris invariantby arotationof angle
φ 1
. Therefore, thematriesA
andB
donotdependonφ 1
.where, in the following,
α 0
andz 0
are assumed to be equal to 0 without anylim-itation for the appliability of the approah. The entries of this new measurement
vetor
α
are the angular positions of a titious rigid arm whih would have thesame geometryand the samemeasurement
z
than the exible one (see Figure7.3).The use of
α
instead ofz
allows the simpliation of the measurement equation.Letus nowonsider
y = φ ˙ α ˙ ⊤
,
(7.16)assystem outputs, i.e., the jointveloities (
φ ˙
)measured by enoders and theti-tious joint veloities (
α ˙
). On top of that, the industrialmanipulatorsare equipped with low-level joint-veloity ontrol loops in order to redue the eets of thefri-tions whih our in the gear-boxes and therefore obtain a simpler behavior (see
Controller in Figure7.2a). In this partiular ase, the inner loop is assumed to be
asa standard stati output feedbak. More preisely,
u = Λ
φ ˙ ∗ − φ ˙
,
(7.17)where
φ ˙ ∗ = φ ˙ ∗ 1 φ ˙ ∗ 2 ⊤
is the vetor of the speed referenes. The state-spae
repre-sentationgiven in Eq. (7.12)beomes
˙ χ =
0 n q ×n φ M −1 (q)HΛ
φ ˙ ∗ +
0 n q ×n q I n q M −1 (q)K(p) M −1 (p) D(p) − HΛ
I n φ 0 n φ ×n v
χ,
(7.18)where
Λ = diag(λ 1 , λ 2 )
. By analyzing the relations given in Eq. (7.18), it appearsthatthisstate-spae formisnotminimalbeausethetworst states
φ 1
andφ 2
haveno eets on the output. The hange of output signal allows the redution of the
modelorder. Thus, aminimalrealizationoforder6anbeextrated by onsidering
¯ χ =
v ⊤ φ ˙ ⊤ v ˙ ⊤ ⊤
.
(7.19)More preisely,
˙¯
χ = A(p, θ g ) ¯ χ + B(p, θ g ) ˙ φ ∗ ,
(7.20)y =
˙ α φ ˙ ⊤
= C χ, ¯
(7.21)where 7
A(p, θ g ) =
0 n v ×n φ 0 n v ×n v I n v (M −1 (p, θ g )K(p, θ g )) (:, n φ + 1 : end) − M −1 (p, θ g )HΛ 0 n q ×n v
,
(7.22a)B(p, θ g ) =
0 n v ×n φ M −1 (p, θ g )HΛ
, C =
0 2∗n φ ×n v C 1
,
(7.22b)7. For oursystem,
D = 0 n
q ×n q
.with
C 1 =
1 0 0 0 0 1 0 0 1 0 ℓ 1 0 0 1 ℓ 1 ℓ 2
.
(7.22)with
θ g ∈ R 16
ontaining the unknown gray-box parameters. This linearized state-spae representation an now be relatedtothe LPV given inEq. (2.13).Indeed, with straightforward alulations, it an be shown that suh a
dark-gray-box LPVdesriptionan be transformedinto anLFR suhasthe one given in
Eq.(2.15)byapplyingthestruturedmatriesofEq.(7.22)andtheinnerstrutures
dened by Eq. (7.10) asfollows,
M(θ g ) =
D zw (θ g ) C z (θ g ) D zu (θ g ) B w (θ g ) A(θ g ) B u (θ g ) D yw (θ g ) C y (θ g ) D yu (θ g )
=
− M −1 0 M 1 M −1 0 M 1 − M −1 0 M 1 0 2×2 M −1 0 M 1 0 2×2 0 2×2 0 2×2 I 2 0 2×2
− M −1 0 K M −1 0 K − M −1 0 HΛ 0 4×2 M −1 0 HΛ 0 2×2 0 2×2 I 2 0 2×2 0 2×2
,
(7.23)where
θ g ∈ R 28
. More preisely,M −1 0 K ∈ R 4×2 , M −1 0 M 1 ∈ R 2×2 , M −1 0 HΛ ∈ R 4×2 .
Notie that the matrix produts dened above ontains the unknown parameters
to estimate. In this ase, the
∆ p
blok-diagonal matrix ontaining the sheduling variable(p = cos(φ 2 )
) has the following form∆ p = I 2 cos(φ 2 ).
(7.24)Moreover, in the blak-box ase, one an onsider
M LP V (θ b ) =
D zw (θ b ) C z (θ b ) D zu (θ b ) B w (θ b ) A (θ b ) B u (θ b ) D yw (θ b ) C y (θ b ) D yu (θ b )
∈ R (n x +n w +n u )×(n x +n z +n y ) ,
(7.25)where
θ b ∈ R (n x +n w +n u )×(n x +n z +n y )
withn u = n y = 2
. As a onsequene, thegray-box LPV model an be alulated by the following upperLFT, i.e.,
F u ( M LP V (θ b ), ∆ p ),
(7.26)while the blak-box one by using,
F u (M(θ g ), ∆ p ).
(7.27)Notie that, beause of the spei test-bed set-up presented in Setion 7.1, the
amera measurements are available only loally. So, in the LTI framework, the
system understudy has fouroutputs
y LT I = φ ˙ α ˙ ⊤
.
(7.28)Namely,thejointveloitiesofthetitiousrobotiarmand themeasured realones.
However, whentherobothastomoveinalargerworkspae,likeintheLPVase,the
outputsofthe systemare restritedtothe jointveloitiesmeasuredby themounted
enoders and