354

Journal of Power Electronics, Vol. 8, No. 4, October 2008

JPE 8-4-8

Precision Position Control of PMSM using Neural Observer and

Parameter Compensator
Jong-Sun Ko, Young-Ger Seo* and Hyun-Sik Kim*

Department of Electronic and Electric Eng., Dankook University, Yongin, Korea

ABSTRACT
This paper presents neural load torque compensation method which is composed of a deadbeat load torque observer and
gains compensation by a parameter estimator. As a result, the response of the PMSM (permanent magnet synchronous
motor) obtains better precision position control. To reduce the noise effect, the post-filter is implemented by a MA
(moving average) process. The parameter compensator with an RLSM (recursive least square method) parameter estimator
is adopted to increase the performance of the load torque observer and main controller. The parameter estimator is
combined with a high performance neural load torque observer to resolve problems. The neural network is trained in online phases and it is composed by a feed forward recall and error back-propagation training. During normal operation, the
input-output response is sampled and the weighting value is trained multi-times by the error back-propagation method at
each sample period to accommodate the possible variations in the parameters or load torque. As a result, the proposed
control system has a robust and precise system against load torque and parameter variation. Stability and usefulness are
verified by computer simulation and experiment.
Keywords : Permanent magnet synchronous motor, Neural deadbeat observer, Parameter compensator, Back-propagation
method

1. Introduction
Recently, precision position control has become more
and more important in LCD inspection of LM drives, chip
mounted machines, semiconductor production machines,
precision milling machines, high resolution CNC
machines, precision assembly robots, high speed hard disk
drivers, etc. It is merging with nanotechnology, as a part
of nano-fabrication, and is also spreading to the
Manuscript received July 8, 2008; revised Aug. 27, 2008
Corresponding Author: [email protected]
Tel: +82-31-8005-3602, Fax: +82-31-897-6104, Dankook Univ.
*
Dept. of Electronic and Electric Eng., Dankook Univ.

bio-engineering
field
and
optical
equipment.
Additionally, it is very important in direct drive systems.
A PMSM has replaced many DC motors since industry
applications require smaller and more powerful actuators.
The PMSM has low inertia, large power-to-volume ratio,
and low noise as compared to permanent magnet DC
servomotors having the same output rating [1,2] . However,
the disadvantages of this machine are high cost and the
need for a more complex controller because of its
nonlinear characteristic.
The proportional-integral (PI) controller usually used in
PMSM control is simple to realize but it is difficult to
obtain a sufficiently high performance in tracking

Precision Position Control of PMSM using Neural Observer and Parameter Compensator

applications. A new systematic approach was done in state

space using digital position information in a PMSM
system[3-6]. However, the machine flux linkage is not
exactly known for a load torque observer thus creating the
problem of uncertainty[7,8]. With the cogging effect, some
damage on the permanent magnet over the current can
affect the value of k t . This causes small position or speed
errors and increases the chattering effect, which should be
reduced as much as possible. It also makes miss-estimated
load torque in deadbeat observer systems. In this paper a
parameter compensator with an RLSM parameter
estimator is suggested to increase the performance of the
load torque observer and main controller. This
compensator makes the system work as if was in a
nominal system parameter. Therefore the deadbeat load
torque observer has a good performance as if there was no
parameter variation. Finally, this controller can be used in
robot or vestibular systems which are simulators. These
systems need exact sinusoidal speed control even when an
unbalanced load is injected. Other production equipment
can use this controller to increase production quality.

355

A new state is defined for the tracking controller as Eqn.

(3). Where r is the rotor speed reference [2]. The control
input becomes Eqn. (4).

z& = r

(3)

iqc1 = k1 k 2 k3 z .

(4)

The augmented system for the speed control of a PMSM

is expressed as follows:

p
p
0 0 kt
&
2J
0
2J
J
+ 0 i 0 T
=
1
0
0


qs L 0 r
z& 0 1 0 z 0
0
1

(5)

y = [0 1 0] .
z

(6)

If the load torque TL is known, an equivalent current

command i qc 2 can be expressed as

2. Modeling of PMSM
The system equations of a PMSM model can be
described as

commands the output controller and compensates for load

torque effect. However, disturbances are unknown or
inaccessible in a real system.

(2)

3.2 Load torque observer and MA process

It is well known that an observer is available when input
is unknown and inaccessible. For simplicity, a 0-observer is
selected [4]. The system equation can be expressed as

3 p
m i qs .
2 2

where
p : number of poles

m : flux linkage of permanent magnet

: angular velocity of rotor

J :inertia moment of rotor

B :viscous friction coefficient.

3. Control Algorithm
3.1 Position controller

(7)

(1)

= r
Te =

1
TL .
kt

Then, the feeding forward equivalent q axis current

p
3 1 p
B
=
TL
m i qs
2 J 2
J
2J

i qc 2 =

p
p
B
0 kt
&

2J
2J

& J

0 y + 0 iqs + L y [0 1 0] y .
y = 1 0

T& 0 0
TL
0 TL 0

(8)
To reduce the disadvantages of a deadbeat observer

356

Journal of Power Electronics, Vol. 8, No. 4, October 2008

which is too noise sensitive a moving average (MA) filter

is considered [9].
1
~
T L (k ) = (TL (k ) + TL (k 1))
2

(9)

] , T (k ) = [ (k ) y (k ) iqs (k ) ] .

where T = [

An RLSM can estimate the real parameter. The resultant

equations are as follows [14-16]:

3.3 Parameter estimator and compensator

The discrete dynamic equation of the PMSM can be
written as

y (k + 1) = (k ) + y (k ) + iqs (k ) + TL (k )
where =

(10)

B
h
P J
J J Bh
J
(h + e J ) ,
(1 e J ) , = 1 , = kt
2J B
B B
B

~ ~
F (k ) (k ) (k ) T F (k )
~
~
1 + (k ) T F (k ) (k )

~
E (k + 1) = y (k + 1) (k ) T (k )
where T (k ) = [
~

[10]

. The control input,

to compensate for parameter variation and to make the

system an equivalent nominal system, is defined as
follows:
iqc (k ) = C1 (k ) (k ) + C2 (k ) y (k ) + C3 (k ) iqc (k )

(11)

Therefore the resultant compensated system is equal to

the nominal equivalent system.
y(k +1) = (k) + y(k) + (C1 (k)(k) + C2 (k) y(k) + C3 (k)iqc (k))
= n (k) + n y(k) + n iqs (k)

(12)
and n , n , n

F (0) =

TL
,
k t

(0 < << 1)

3.4 Proposed neural network observer

The approximation of the multi-variable neural network
can be done by the Hornick function method. This neural
network can compensate for the effects of system
parameter variations [17-19]. Fig. 1 shows a controller with a
back-propagation neural network (BPNN) used which is
based on an augmented state feedback [20-23].
This BPNN uses rotor speed, rotor speed reference,
speed error, and equivalent q-phase current commands of
the torque observer as input nodes to learn the optimal
current command. The error between the real controller
output and neural network propagate back to a hidden
layer under the following bipolar activation function.

parameters and nominal parameters, respectively. These

values can be obtained easily with Eqn. (12) as
( n )
( n )

C1 =
, C2 =
, C 3 = n respectively.

Parameter compensation requires real parameter

estimation [11-13]. Using a discrete system equation without
disturbance Eqn. (13) a parameter and a measured
parameter can be separated.
(13)

(15)
(16)

are actual

y (k + 1) = (k ) + y (k ) + iqs (k ) = T (k )

(14)

],

Based on the assumption that there is no effect, the load

torque, feed back gain and feed forward gain are defined

where , ,

F (k + 1) = F (k )

T (k ) = (k ) y (k ) iqs (k )

P J J
J Bh
=
( h e J )
2J B B
B

as C1 , C 2 and C3 respectively

(k + 1) = (k ) + F (k + 1) (k ) E (k + 1)

Fig. 1 Diagram of the neural network using the error

back-propagation

Precision Position Control of PMSM using Neural Observer and Parameter Compensator

f (netk ) =

2
1
1 + exp(netk )

357

(17)

netk = kj y j

(18)

ok = f (netk )

(19)

The slope of the activation function is considered as 1

for simplicity and weight is changed according to the delta
learning rule. Eqn. 19 represents the neural network output
of the each node.

E=

1 n
(d k ok ) 2
2 k =1

kj =

E
kj

Fig. 2 Block diagram of the proposed algorithm

(20)

4. Configurations of Overall Systems

(21)

The total block diagram of the proposed controller is

shown in Fig. 2. The C-Language program and a
TMS320C31 DSP implement the digital control.

where d k is iqc . A connection weight wkj is a weight

between j-th and k-th hidden layer is of NN. This delta
rule can guarantee weight movement to the negative line
of the error variation. The error signal is known as Eqn. 22
applying a chain rule. Therefore from Eqn. 17 and Eqn. 22
the error signal can be obtained as Eqn. 23.

ok =
ok =

E
E o k
=
net k
ok net k

1
2
( d k ok )(1 o k )
2

(22)
Fig. 3

From these results, the weighting functions are adapted

by the delta learning rule as follows [24]:

kj (k + 1) = kj (k ) + ok y j

Block diagram of the proposed control system

(23)
Experimental load systems directly coupled to the motor
axis are depicted in Fig. 3. This system creates time
varying load torque to show the effectiveness of the
proposed algorithm.

(24)
PMSM

Load

Hidden layer weights are also changed by the same

method.

yj =

n
1
2
(1 y j ) ok kj
2
k =1

ji ( k + 1) = ji ( k ) + yj x i

PMSM
r2

r1

(25)

Bar Load
t1

(26)
(a) Inertial load

The resultant block diagram of the proposed controller is

shown in Fig. 2.

(b) bar load

Fig. 4 The figure of load for parameter and load variation

358

Journal of Power Electronics, Vol. 8, No. 4, October 2008

[A]

3
2.5
2
1.5
1
0.5
0
-0.5 12
-1
-1.5
-2
-2.5
-3

13

14

15

16

17

18

[sec]

(a) Augmented state feedback

Fig. 5

The configuration of the experiment system

Position

1
0.8
0.6

5. Simulation and Experimental Results

0.4

The parameters of a PMSM motor used in this

simulation and this experiment are shown in Table 1.

[rad]

0.2
0
12

13

14

15

16

17

18

16

17

18

-0.2
-0.4
-0.6
-0.8
-1

Table 1

Power : 400 watt

PMSM parameters
0.363 104

Inertia :

Rated torque : 1.3 Nm

[sec]

kgm2

Iqc

3
2.5

Stator resistance : 1.07

2
1.5

Phase inductance : 4.2 mH

1
0.5

[A]

Rated current : 2.7 A

0
-0.5

12

13

14

15

-1

The hysteresis band gap is chosen as 0.01[ A ] and the

sampling time h is determined as 0.2 [ ms ]. The weighting

-1.5
-2
-2.5

matrix is selected as Q = diag [0.1 80 30000] , R = 1

and

the

optimal

gain

matrix

becomes

-3

[sec]

(b) Dead beat observer and parameter compensator algorithm

k = [0.0773 4.9807 62.5080]. The deadbeat observer

L = [9623.9 2.7000 275.20]

[25]

Position

1
0.8
0.6
0.4
0.2

[rad]

and the gain matrices are calculated from nominal values.

The gain is obtained using the pole placement method at
origin
in
z
domain
and
becomes

0
12

. The simulation

-0.2

results are shown in Fig. 5 and Fig. 6. Fig. 5 shows the

speed response of the conventional controller. There is a
small speed ripple and small overshoot caused by a current
ripple of the hysteresis band gap and parameter variation.

-0.6

13

14

15

16

17

18

16

17

18

-0.4

-0.8
-1

[sec]

Iqc

3
2.5
2
1.5
1

Position

0.5

[A]

0.8
0.6
0.4

12

13

14

15

-1
-1.5

0.2

[rad]

0
-0.5

-2

0
12

13

14

15

-0.2

16

17

18

-2.5
-3

-0.4

[sec]

-0.6
-0.8

(c) Parameter compensation algorithm

-1

[sec]

359

Precision Position Control of PMSM using Neural Observer and Parameter Compensator

Position

Position

0.8

0.8
0.6

0.78

0.4

0.76

0
12

13

14

15

16

17

18

[rad]

[rad]

0.2

-0.2

0.74

-0.4
-0.6

0.72

-0.8
-1

0.7

[sec]

12

12.5

13

13.5

14

14.5

15

[sec]
Iqc

(c) Proposed algorithm

2.5
2

Fig. 7

1.5
1

[A]

0.5
0
-0.5

12

13

14

15

16

17

18

Performance comparison of three controllers for the

parameter variation, Zoom in the rotor position(0.7rad~
0.8rad)

-1
-1.5
-2
-2.5
-3
[s ec]

(d) Proposed algorithm

Fig. 6

Simulation results of the rotor position, q phase current

command for load

The inertial parameter has 100 times the permanent

magnet value and two times the R and L value. This
conventional algorithm makes a large current ripple due to
parameter variation. Fig. 6 shows the results of a proposed
algorithm that has the same position command and same
disturbance condition as Fig.4. The load effects are reduced
by the proposed algorithm of the parameter compensation.
In the proposed system with NN, the error is markedly
decreased, more than in the case of the parameter
compensation.

Fig. 7 presents 0.1[rad] scaled simulation results to

show a comparison between the two controllers. The
conventional controller in Fig. 7(a) has a large position
ripple compared with the proposed system Fig 7(c). The
parameter compensation is not as good as in the parameter
compensation plus NN case. This is the result of a
parameter compensator with a neural network.

(a) Augmented state feedback algorithm

Position

0.8

0.78

[rad]

0.76

0.74

0.72

0.7
12

12.5

13

13.5

14

14.5

15

[sec]

(b) Load torque observer algorithm

(a) Load torque observer algorithm

Position

0.8

0.78

[rad]

0.76

0.74

0.72

0.7
12

12.5

13

13.5

14

14.5

15

[sec]

(b) Parameter compensation algorithm

(c) Parameter compensation algorithm

360

Journal of Power Electronics, Vol. 8, No. 4, October 2008

(d) Proposed algorithm

Fig. 8

Experimental results of the rotor position and

q-phase current command with a inertia load

The experimental results are depicted in Fig. 8 and Fig.

9. In this experiment, real observable gains are reduced to
about 30% to obtain some effectiveness in parameter
compensation. The parameter compensator calculates the
real parameter and compensates for current miss turned
gains. Fig. 8 shows experimental results of the position
with current command at about 3 seconds in duration.
There is a current ripple in a steady state and large position
overshoot in transient state shown in Fig. 8(a) and 8(b).
However, after 20 minutes, there is no current ripple and
the position error in the proposed system as shown in Fig.
8(b). A more detailed figure is shown in Fig. 9 from 0 to
20 minutes. This figure has a scale between 0.1[rad]
(0.7[rad] ~ 0.8[rad]) with a zooming data and shows that
position error decreases gradually as time goes on.

(c) Disturbance observer and parameter compensator

(after twenty minutes)
Fig. 9 Experiment results of zoom in the rotor position for
inertia load

6. Conclusions
A new deadbeat load torque observer with a system
parameter compensator was proposed to obtain better
performance from the PMSM in a precision position
control system. This compensator makes a real system
work as in a nominal parameter system. Therefore the
deadbeat load torque observer demonstrated a good
performance and acted as if there were no parameter
variation. To reduce of the effect of noise a post-filter was
implemented by an MA process. The system response
comparison was conducted between the deadbeat gain
observer and the parameter compensated system with a
deadbeat observer. Since the parameter compensated
system acted as if there was no parameter variation, the
conventional deadbeat load torque is well adapted to real
systems. It can be used to cancel out steady state and
transient position errors due to external disturbances, such
as friction, load torque and the small chattering effect of
the deadbeat control.

Acknowledgment
(a) Disturbance observer algorithm

This
work
was
supported
by
grant
No.(R01-2006-000-10118-0) from the Basic Research
Program of the Korea Science & Engineering Foundation.

References

(b) Disturbance observer and parameter compensator

(after one minute)

[1] D. W. Novotny, R. D. Lorentz, Introduction to field

orientation and high performance AC drives, IEEE-IAS
Tutorial Course, 1986.
[2] P. C. Krause, Analysis of electric machinery, McGraw-Hill,

Precision Position Control of PMSM using Neural Observer and Parameter Compensator

1984.
[3] Nandam P. K., Sen P. C., A Comparative Study of a
Luenberger Observer and Adaptive Observer-Based Variable
Structure Speed Control System using a Self-Controlled
Synchronous Motor, IEEE Trans. Industrial Electronics,
Vol. 37, No. 2, pp. 127132, 1990.
[4] K. J. Astrm, B. Wittenmark, Computer controlled system,
Prentice Hall, International, 1997.
[5] J. S. Ko, T. H. Lee, B. L. Park, C. W. Jeon, Precision Speed
Control of PMSM for Stimulation of the Vestibular System
Using Rotatory Chair, Transaction on KIPE, Vol. 5, No. 5,
pp. 459466, Oct. 2000.
[6] J. S. Ko, T. H. Lee, C. W. Jeon, S. S. Lee, Precision Speed
Control of PMSM Using Disturbance Observer and
Parameter Compensator, Transaction on KIPE, Vol. 6, No. 1,
pp. 98106, 2001.
[7] D. Alazard, P. Apkarian, Exact Observer-Based Structure
for Arbitrary Compensators, Int. J. Control, Vol. 41, No. 6,
pp. 15651575, 1999.
[8] J. S. Ko, S. K. Youn, A Study of Adaptive Load Torque
Observer and Robust Precision Position Control of BLDD
Motor, Transaction on KIPE, Vol. 4, No. 2, pp. 138143,
April 1999.
[9] C. T. Chen, Linear System Theory and Design, Holt,
Rinehart and Winston, Inc., 1984.
[10] C. Y. Huang, T. C. Chen, C. L. Huang, Robust Control of
Induction Motor with A Neural-Network Load Torque
Estimator and A Neural-Network Identification, IEEE
Transaction on Industrial Electronics, Vol. 46, No. 5, pp.
990998, 1999.
[11] K. S. Narendra, K. Parthasarathy, Identification and
Control of Dynamical Systems using Neural Network,
IEEE Trans. on Neural Networks, Vol. 1, No. 1, pp. 427,
1990.
[12] S. Chen, S. A. Billings, Neural Networks for Nonlinear
Dynamic System Modeling and Identification, Int. J.
Control, Vol. 56, No. 2, pp. 319346, 1992.
[13] Tanaka K, Yuzawa T, Moriyama R, Miki I., Initial Rotor
Position Estimation for Surface Permanent Magnet
Synchronous Motor, Record of Thirty-Sixth Industry
Applications Conference, Vol. 4, No. 30, pp. 25922597,
2001.
[14] J. D. Landau, System Identification and Control Design.
Englewood Cliffs, NJ, Prentice-Hall, 1990.
[15] G. C. Goodwin, K. S. Sin, Adaptive Filtering Prediction
and Control. Englewood Cliffs, NJ, Prentice-Hall, 1984.
[16] Faa-Jeng Lin, Po-Hung Shen, Ying-Shieh Kung, Adaptive
Wavelet Neural Network Control for Linear Synchronous

361

Motor Servo Drive, IEEE Transaction on Magnetics, Vol.

41, No. 12, pp. 44014412, 2005.
[17] Fukuda, T., Shibata, T., Theory and Application of Neural
Networks for Industrial Control Systems, IEEE Trans. on
Industrial Electronics, Vol. 39, No. 6, 1992.
[18] J. O. P. Pinto, B. K. Bose, L. E. B. da Silva, A Stator Flux
Oriented Vector-Controlled Induction Motor Drive with
Space Vector PWM and Flux Vector Synthesis by Neural
Networks, IEEE Trans. Ind. Applicat., Vol. 37, pp.
13081318, 2001.
[19] Yang Yi, D. Mahinda Vilathgamuwa, Azizur Rahman,
Implementation of an Artificial-Neural-Network-Based
Real-Time Adaptive Controller for an Interior
Permanent-Magnet Motor Drive, IEEE Trans., Ind.
Applic., LA-39, No. 1, pp. 96104, 2003.
[20] Naomitsu Urasaki, Tomonobu Senjyu, Toshihisa Funabashi,
Hideomi Sekine, An Adaptive Dead-time Compensation
Strategy for a Permanent Magnet Synchronous Motor Drive
Using Neural Network, Journal of Power Electronics, Vol.
6, No. 4, pp. 279289, Oct. 2006.
[21] Naomitsu Urasaki, Tomonobu Senjyu, High Efficiency
Drive Technique for Synchronous Reluctance Motors
Using a Motors Using a Neural Network, Journal of
Power Electronics, Vol. 6, No. 4, pp. 340346, Oct. 2006.
[22] Mona N. Eskander, Neural Network Controller for a
Permanent Magnet Generator Applied in Wind Energy
Conversion System, Journal of Power Electronics, Vol. 2,
No. 1, pp. 4654, January 2002.
[23] Mona N. Eskander, Minimization of Losses in Permanent
Magnet Synchronous Motors Using Neural Network,
Journal of Power Electronics, Vol. 2, No. 3, pp. 220229,
July 2002.
[24] Fayez Fahim El-Sousy, A Vector-Controlled PMSM Drive
with a Continually
On-Line Learning Hybrid
Neural-Network Model-Following Speed Controller,
Journal of Power Electronics, Vol. 5, No. 2, pp. 129141,
April 2005.
[25] J. S. Ko, J. H. Lee, S. k. Chung, M. J. Youn, A Robust
Position Control of Brushless DC motor with Dead Beat
Load Torque Observer, IEEE Transaction on Industrial
Electronics, Vol. 40, No. 5, pp. 512520, 1993.
Jong-Sun Ko was born in South Korea in
1960. He received his B.S. degree from Seoul
National University, Seoul, South Korea, in
1984, and his M.S. degree from Korea
Advanced Institute of Science and
Technology (KAIST), Taejon, South-Korea,

362

Journal of Power Electronics, Vol. 8, No. 4, October 2008

in 1989. He received his Ph.D. in Electrical Engineering at

KAIST in 1994. He has been with the Production Engineering
R&D Center of Samsung Electronics Co. from 1983 to 1996. He
was a visiting scholar at the University of Tennessee in 1999. He
was a Prof. of Electric & Electronic Engineering in Wonkwang
University, Iksan, South Korea from 1996 to 2003. He has been a
professor of Electronic and Electric Engineering in Dankook
University, Seoul, South Korea since 2003. His research interests
are in drive systems, digital control, robust control, and factory
automation. He is a member of IEEE, KIEE and KIPE. He has
done projects with the Samsung Company and other universities
in the field of automatic production, automatic controllers,
PMSM controllers, and Torque Observers for PMSM.
Young-Ger Seo was born in South Korea in
1982. He received his B.S. degree from
Dankook University, Seoul, South Korea, in
2007. He is currently a M.S. student at
Dankook University. His research interests
are in motor drive controls, development of
simulators and wind power systems.
Hyun-Sik Kim was born in South Korea in
1978. He received his B.S. degree in
Electrical Engineering from Wonkwang
University, South Korea, in 2004.
He
received his M.S. degree from Dankook
University. He is currently working towards
his Ph.D. at Dankook University. His research interests are in
drive systems, digital control, and control of AC motors.