Professional Documents
Culture Documents
Synthesis and Analysis of Robust Flux Observers For Induction Machines
Synthesis and Analysis of Robust Flux Observers For Induction Machines
GE 44, CRTT - Bd. de I'Universite - B.P. 406 - 44602 Saint Nazaire cedex - France
Ph. (33) 24017 26 35, E-mail.' mickael.hilairet@ge44.univ-nantesfr
tlRCCyN - UMR CNRS 6597, 1 rue de La Noe, 44072 Names Cedex 3 - France
E-mail: pchevrel@emnfr
Abstract: lIDs paper presents and compares two approaches for the design of robust
flux observers for induction motors. Both take the measurement noises into account.
The first one, based on Linear Matrix Inequalities (LMI) formulation, leads to a
quadratically stable flux observer. The second one is a Kalman filter with a
particular tuning method, which reduces the effects of some parameter uncertainties.
The dynamical performances, robustness and algorithmic complexity of both
observers are shown and compared. Copyrighte;2000 IFAC
573
two proposed design methodologies. Lastly, their random disturbance Wsl, so as to take inverter effects
respective performances and algorithmic complexity into account. The resulting measured vector is made
are analyzed. of the stator current (weighted as shown on fig 1) and
the stator voltage (Us) provided by the controller.
~}-
electrical part. The differential equation of the
mechanical subsystem is not considered here, and 'I
we have chosen the stator reference frame to obtain a
(a,~) LPV (Linear Parameter-Varying) electrical
subsystem (1), where the rotor speed Q is considered
l' Q(O)
as an external time-varying parameter.
Fig 1: Standard Plant for the observer design
o
o 0 Finally let us defme cl> , the estimation of cl>r and cl> r
o
the observation error <1», - «1>,. The weighting
3. H.. LPV FLUX OBSERVER DESIGN 'VUe Sn•• 3ae [0;1) / Q= a.Q I11III....
. +(l-a).Q
34
{'VQ,Sp(Q)e Sp 3ae[O;I)/Sp =a.S p_ +(I-a).Sp_ ( , )
Let us consider the LPV model represented by the
state space equation (2) corresponding to (1).
Then, the H.. LPV flux observer design problem may
be written as: find a dynamical system Q(Q) (5)
providing, from the stator current and voltage
= ~~~~-~Jx,.]
i,
[ C; 0I
I
c.io'
V
<=> ~~ 1~~~+~1
C 0
y
c.io
I
I
(2) measurements, the flux estimation ci>,
estimation error cl> r is minimized in the Lrnorm
such that the
574
This problem is the classical problem of finding a
LPV feedback to control a LPV plant according to
quadratic H.. performance. A solution to this problem
can be derived from Apkarian and Gahinet (1995),
formulating it in terms of convex optimization
problem under LMI constraints. It is provided by the
LMI-solver (Gahinet, et al., 1995) available in the As shown in Hilairet and Auger (2000), the two
Matlab® package. The resulting observer is a matrices Aci(Q) and Bd contain many symmetries and
quadratically stable continuous-time system of 611> anti-symmetries, which when taken into account,
order (7), as for the standard plant (P) (Darengosse et provide an important reduction of the complexity of
al., 1999; Apkarlan and Gahinet, 1995): the Kalman filter. The random signals v[k] and w[k]
modelize the discretization error, the effects of the
inverters, the parameter uncertainties and the
measurement noise. Ka1man filters (Grewal and
(7)
Andrews, 1995) are classical algorithms, and their
equations are shortly recalled here (see equation 11).
Compared to Luenberger observers, the main
Starting with the linear fractional representation differences are that the correction uses the predicted
(LFR) of the observer (Darengosse and Chevrel, state rather than the previous estimated state, with a
1999), the main idea of the discretization consists in gain deduced from the statistics of the noises, rather
replacing the integration operator (S-I) by a than a pole assignment approach (Verghese and
generalized Euler operator. Being strictly causal, this Sanders, 1988).
approximation may be applied straightforwardly.
Moreover, its time prewarping coefficients insure
accuracy whatever the frequency of internal signals. X(k+~k] = A.s(O).xo<Ik]+Bd.luc]
This property is interesting because the electrical P[k+~k]=A.s(O)Jl[kjk]Atd(O)+Q
variables are defined in a stator reference framework,
Iqk+l]=P[k+~klC.(CP[k+~k].c +R)~
and these signals are sinusoidal with a frequency
depending on the rotor speed. Finally the observer X(k+~k+l]=X(k+~k]+Iqk+l].(Y[k+l]-CX[k+~k])
algorithm is given by the following expression with P[k+~k+I]=P[k+~k]-Iqk+l].CP[k+~k]
the sampling period (Ts=400 J.l.s): -
Wc] = (llk]+Wc+l])/2
qk]=(qk]+qk+l])/2
(11)
(12)
o <Xl 0 <X2
575
2 2 this Kalman filter are the classical ones. But the
.
WIth 0. 1 =a.
-2-'
a
0. 2
a·
=_"-
a
and a =
2
p.a•.a"
a J 2 proposed implementation takes the structural
properties of the transition matrix into account,
yielding a lower computational cost. This observer
This choice makes the implementation of the Kalman has three degrees of freedom, and a determination
filter easier. Decreasing the elements of Q means that procedure is proposed, based on an optimization of
the model provides a good state prediction, and the the filter in a given operating mode, and for the
sensors yield noisy measures, leading to small values worst-case parameter uncertainty.
of the correction gains. On the opposite, increasing
the elements of Q means that the state prediction is 2)1 tOol
.,.
corrupted by parameter uncertainties, and/or the
sensor noise is negligible, leading to an important
correction of the prediction state thanks to the ·2
5. ALGORITHM COMPLEXITY
As a conclusion, the Kalman observer is based on a For the evaluation of the total number of arithmetic
discrete-time model, deduced from the continuous- operations, the trigonometric functions required for
time model by an approximation. The equations of the H.. LPV flux observer are supposed to be
576
computed by fIfth degree polynomials, as done for Figure 7 and 8 are two simulations with respectively
example on DSP's (Sitton, 1997). The three an underestimation and overestimation of 25 % of the
trigonometric functions are therefore equivalent to 15 stator resistance. The two observers have a nearly
multiplications and 15 additions. equivalent rotor flux estimation error and they are not
sensible to stator resistance uncertainties unlike most
of the previously proposed observers.
6. SIMULATION RESULTS AND
PERFORMANCE ANALYSIS Figure 9 is a simulation with an overestimation of
10% of the mutual inductance, which may result
The following simulation results show the behavior from a magnetic saturation of the induction machine.
of the observers in a open-loop scheme. The rotor The two observers have the same behavior whereas
flux estimation (solid line) is not used in the rotor the H.. LPV flux observer seems to be more accurate.
speed control and noises are added in the
measurements (voltages, currents and rotor velocity).
,~
, --~--~----------------~~
The simulation setup fIrst starts-up the motor by
applying a sinusoidal three-phase voltage on the a.• ~
Figure 4 is a simulation without parameters 'a l'--~--~-- __- -__- -____ --~~
an underdamped response, whereas the one estimated ·,~.3 0:. 0.5 U 0.7 0.8 O~,
by the H.. LPV flux observer has an important titNo (t )
(dashed lines) is more robust against overvalued rotor a1.,' a., a.. D.' 0.7 0.'
577
considered. Simulations performed with uncertainties
," --~--~----------------~--
in the rotor resistance, the stator resistance or the
mutual inductance have pointed out that both
observers nearly have the same robustness. Future
extensions of these approaches may lead to a better
control of the dynamical performances, and to the
estimation of the angular velocity for sensorless
,.---~--~----------~--- control.
.-
~ r---'·--·---.---·-..·- -. --------.~ REFERENCES
.1~3 04 05 01 01 01 O' 1
"""(S' Apkarian. P. and P. Gahinet (1995). A convex
characterization of gain-scheduled H.. control-
lers. IEEE Trans. on Auto. Control. Vol. 40, N°5,
Fig 7: Open-loop scheme.iRslR. =+25% and .iRr/R. =0% pp 833-864.
Bodson, M., J. Chiasson and R. Novotnak (1994).
,~
! --~--~----------------~-- High performance induction motor control via
O.• ~
Input-Output linearization. IEEE Control
i •• ~ Systems Magazine, Vol. 14, No 4, pp 25-33.
.- , .7
~ /---'------------ Bornard, G. and H. Hammouri (1991). A high gain
t
..•
0.7 t-
.; observer for a class of uniformly observable
·~'3' •• ... ... ..7 ..• systems. In Proc. IEEE CDC. pp 1494-1496 .
Darengosse, C., P. Chevrel, I. S. Guelle and S. Siala
,.,~
, --~------------------~-
(1999). A LMI-based observer for the induction
'i
motor. In Proc. ofECC'99. ID N°770.
I ·f....-..-.-.---..--...-.-.•.------_~---..-.----_1 Darengosse, C. and P. Chevrel (1999). Synthese et
.. ··r ~ experimentattion d'un Observateur de Flux LPV.
• 1~3 04 05 oe 01 06 01 1
In Proc. ofJDA'99 (Nancy-France). ppI15-119 .
tir'n.(s)
Delmotte, E. (1997). Observateur robuste de flux
pour la commande vectorielle d'une machine
Fig 8: Open-loop scheme.iRslR. =-25% and .iRr/R. =0% asynchrone. Universite des sciences et
technologies de Lille (France), N°2138.
Kinpara, Y., S. Doki, S. Okuma and S.
Sangwongwanich (1995). Slip-Frequency-Type
0.11 :'-
vector control equivalent to Flux-Feedback
.-I." '
0.7 ~/'-- .. --~---- ...- - - - - - - - - - -..-....... -.~ .. ~-~-~- .•• ,. ...~~ .. -~~.~--- ~
vector control using flux observer. Elect. Eng. in
Japan. Vol. 115, n02. pp 123-134.
, Gahinet, P., A. Nemirosvsky, A. J. Laub and M.
·~:3! 0.' ... ..• 07 ..• •. Chilali (1995). L.M.I. Control Toolbox.
'."--~------------------~- Mathworks Inc.
., Grewal, M.S. and A.P. Andrews (1995). Kalman
I-v· . rJ,,----.-..-_.-------.----,, . ________ _
~ filtering, theory and practice (Thomas Kailath,
578