Position Control of a PM Stepper Motor

Using Neural Networks

Gang Feng
School of Electrical Engineering, University of New South Wales
Sydney, NSW 2052, Australia
Dept. of MEEM, City University of Hong Kong
Tat Chee Ave., Kowloon, Hong Kong
Email: [email protected]
Abstract
This paper considers position control of a PM stepper
motor. A new control scheme is proposed based on a kind
of exact linearization controller and a neural network
based compensating controller. This scheme takes
advantages of simplicity of the model based control
approach and uses the neural network controller to
compensate for the motor modeling uncertainties. The
neural network is trained on line based on Lyapunov
theory and thus its convergence is guaranteed.
1.

Introduction

Due to the various disadvantages of DC motors,

positioning systems are more and more implemented by
using induction motors or stepper motors. Originally,
stepper motors were designed to provide precise
positioning control within an integer number of steps
without using any sensors. This is an open-loop operation.
However, using the stepper motors in an open-loop
configuration results in very low performance. In
particular, the PM stepper motors have a step response
with significant overshoot and a long settling time.
Therefore, feedback control has been proposed for stepper
motor position systems. A number of feedback control
techniques have been developed during the last few years.
One of them is the so-called feedback linearization
controller [1-3]. The basic idea is to design a feedback
controller so that the nonlinear stepper motor system
becomes a linear system. Then the techniques of linear
control systems can be used to design the controller for the
linearized stepper motor system so that the required
performance can be achieved.
It is understandable that such a control technique is
capable of providing excellent positioning results if the
complete dynamics of the stepper motor are known.
However it actually results in no better performance than
that of the conventional fixed gain controllers due to the
fact that it is very hard to obtain perfect dynamic model for
the stepper motors in practice. Moreover, there also exist
uncertainties, as well as time-varying effects in practice.
All these factors lead to the difficulty of achieving higher
precision positioning performance.
Recently, increasing attention has been paid to the use of
artificial neural networks in nonlinear control [4-6]. One

possible way to use the neural networks is to replace either

the entire control system or the feedforward controller with
neural networks. Such research work was based on the
desire to obtain the benefits of model-based control
without a priori knowledge of system dynamics. However,
in many cases, the nominal dynamic model of the stepper
motor can be found a priori indeed. Therefore, a priori
knowledge of stepper motor dynamic model should be
appropriately used rather than totally discarded.
In this paper, a new scheme for stepper motor positioning
control is proposed. This scheme takes advantage of
simplicity of the conventional methods, and incorporates a
compensating controller to achieve high positioning
performance. The compensating controller is based on an
RBF neural network, which is trained on-line to identify
the stepper motor modeling uncertainties.
The rest of the paper is organized as follows. Section 2
presents the problem formulation and Section 3 proposes a
new control scheme which is followed by some concluding
remarks in section 4.
2.

Problem formulation

The equations describing the stepper motor can be given as

below [7],
di a
= [v a Ri a + K m sin( N r )] / L
dt
dib
(1)
= [vb Rib K m cos( N r )] / L
dt
d
= [ K m i a sin( N r ) + K m ib cos( N r ) G
dt
K D sin( 4 N r )] / J L / J
d
=
dt
where ia , ib and v a , vb are the currents and voltages in
phase A and B respectively, L and R are the selfinductance and resistance of each phase winding, K m is
the motor torque constant, N r is the number of rotor
teeth, J is the rotor inertia, G is the viscous friction
constant, is the rotor speed, is the motor position, L
is the load torque, and the term K D sin( 4 N r ) represents

the detent torque due to the permanent rotor magnet

interacting with the magnetic material of the stator poles.
As shown in [1], if the modeling is perfect, with an
appropriate nonlinear coordinate transformation and exact
feedback linearization, the stepper motor equations can be
rewritten as,
x = Ax + Bu
(2)
where
0 1 0 0
0 0
0 0 1 0
0 0

A=
B=
0 0 0 0
1 0

0 0 0 0
0 1

and x 4 is related with the maximum rated phase current.

Then a possible linear control law for u can be chosen as,
x1 x1d

0 x 2 x 2 d r (t )
a 30 a 31 a 32
+
(3)
u =

0
0 a 40 x3 x3d 0
0

x4
where [ x1d x 2 d x 3d ] := [ d d d ] / K 3 , d , d , d are
the desired position, speed and acceleration profiles
respectively, and r (t ) =  d / K 3 . It can be seen that with
the above controller, the closed loop system can be
expressed as,
~
x = Ac ~
x
(4)
where
~
x := [x x
x x
x x
x ]T
0
0
Ac =
a 30

1d

2d

0
a 31

1
a 32

3d

It can be clearly seen that the performance of the stepper

motor position control system will degrade due to this
uncertainty term f(x). Our objective is to design a
compensating controller to improve the stepper motor
positioning performance. If the nonlinear function f(x)
were known a priori, then a modified controller
(6)
u = uo + uc
a30
uo =
0

[ x1 x 2 x3 ] := [ ] / K 3 ,
:=  , K 3 := K m / J ,

the motor state variables. It is supposed that the uncertainty

term satisfies the matching condition, that is, the term is in
the range of the control signal u(t). In such a case, the
system can be expressed as,
(5)
x = Ax + B[u + f ( x)]

0
0
.
0

a 40

Therefore, the error dynamics of the positioning control

system can be well designed by suitably choosing the
coefficients a 30 , a 31 , a 32 and a 40 .
However, if the modeling is not perfect or there exist
uncertainties which is always the case in practice, then the
above design might lead to the significant degrade of the
performance. In this case, we propose a new control
scheme which includes a compensating controller. It is
supposed that the nominal model of the stepper motor is
known a priori. Then the feedback linearization control
discussed above can be designed based on this nominal
model. However, in this case, the stepper motor control
system equation will not be expressed as in eqn.(2).
Instead, there will exist an uncertainty term f(x) in the
equation, which will be the unknown nonlinear function of

a31 a32
0

x1 x1d
0 x2 x2 d r (t )
+
a40 x3 x3d 0

x4

(7)

(8)
u c = f (x)
would lead to the closed loop system expressed as in
eqn.(4), i.e., the known modeling uncertainties could be
well compensated.
Unfortunately, the non-linear function f(x) is unknown a
priori in practice. Therefore the above modified controller
could not be implemented. However, this controller
suggests indeed that a well estimated function f ( x) of the
non-linear function f(x) could be used to improve the
stepper motor positioning performance. It is noted that
there exist another nonlinear function h(x~ ) such that
f ( x ) = h( ~
x ) by taking notice of the definition of the state
variable ~
x . With the same control law (6), (7) and a new
compensating control law, which will be discussed
subsequently, the closed loop stepper motor control system
can be expressed as,
~
(9)
x = Ac ~
x + B[u c + h( ~
x )]
Due to their great approximation capability, artificial
neural networks will be used in this paper to identify this
non-linear function [8-10]. For this we make the following
assumptions.
Assumption 1: The closed loop stepper motor control
system, whose controller is designed based on the nominal
stepper motor model, is stable, and the tracking error
vector ~
x belongs to a compact set.
Assumption 2:
(i) Given a positive constant

0 and a continuous

function h : C R , where C R mr is a
compact set, there exists a weight vector = *
such that the output h( ~
x , ) of the neural
network architecture with n* nodes satisfies
x , * ) h( ~
x ) | ,
max | h( ~
~
xC

where n* may depend on precision parameter

0 and the function h.
x , ) of the neural network
(ii) The output h( ~
architecture is continuous with respect to its
x , ).
arguments for all finite ( ~
Remark 1: This paper is mainly concerned with the
compensating
controller
design
for
positioning
performance, thus it is reasonable to assume the control
system based on the nominal model is stable in the
assumption 1. Assumption 2 is also reasonable due to the
universal function approximation capability of the neural
networks.
3. A new control scheme
x ) be
Suppose the unknown nonlinear function h( ~
parameterized by a static RBF neural network with output
*
h( ~
x , ) , where R n is the adjustable weight, and n
denotes the number of weights in the neural network
approximation. Then the eqn.(2) can be rewritten as
~
x = A ~
x + B[u + h( ~
x , * ) + (h( ~
x ) h( ~
x , * ))]
(10)
c

where * denotes the optimal weight values in the

x belonging to a compact set
approximation for ~
2n
C ( M x ) R , that is C ( M x ) := {~
x : || ~
x || M x } . In
general, the "optimal" weight * in eqn.(10) could take
arbitrarily large values. However, in order to avoid any
numerical problems that may arise due to too large weights
and to prevent the weights from drifting to infinity, we are
only concerned with weights that belong to a large
compact set B( M ) , where M is a design constant, and
B( M ) := { : || || M } denotes a ball of radius M . In
the design of adaptive law, we also restrict the estimate of
* to the compact set B( M ) through the use of a
projection approach. In this way, the optimal weight * is
defined as the element in B( M ) that minimizes the
x ) h( ~
x , ) | | for ~
x C , i.e.
function || h( ~
x ) h( ~
x , ) ||}
* := arg min { sup || h( ~
B ( M ) ~x C ( M )
x

Now eqn.(10) can be expressed as

~
x = A ~
x + B[u + h( ~
x , * ) + ( ~
x )]
c

of functions defined element-wise as

x ci | 2
|~
x ) = exp(
i = 1, 2, .., n*.
),
(~
i2
For the sake of tractable analysis, ci and i are chosen a
priori and kept fixed. The local training techniques
presented in [11] could be used for appropriately choosing
the centres and widths of the RBF neural network. Or the
centres can be chosen as the mesh points equally
distributed inside the compact set. In such case, the only
adjustable parameters appear linearly with respect to the
known nonlinearity (x).
Now, eqn.(11) can be rewritten as
T
~
x = A ~
x + B[u + * ( ~
x ) + (~
x )]
(14)
c

Next, we will show that an adaptive compensating

controller can be added to achieve improved positioning
performance. The new compensating control law is as
follows:
u = T ( ~
x)
(15)
c

where is the estimated parameter for *.

This control law leads to the closed loop system expressed
as
T
~
x = Ac ~
x + B[ * ( ~
x ) + (~
x ) T ( ~
x )]
~
T
x + B[ ( ~
x ) + (~
x )]
= Ac ~
(16)
~
*
where := denotes the parameter estimation error.
Choosing a Lyapunov function candidate
1 T ~ 1 ~ 2
V = ~
x Px +
|| || F
>0
2
2

(17)

where || || 2F denotes the Frobenius matrix norm, defined

as || R || 2F := ij | rij | 2 , and the matrix P is positive definite
and satisfies the following Lyapunov equation
PAc + AcT P = Q

(18)

with Q 0.
(11)

where denotes the error due to the use of the neural

network,
x ) := h( ~
x ) h( ~
x , * )
(12)
(~
The error is bounded by a finite constant vector 0 ,
where
x ) h( ~
x , * ) ||
:= sup || h( ~
0

T
h( ~
x , * ) = * ( ~
x)
(13)
where * is a matrix representing the optimal weight
values subject to the constraints ||*|| M, the vector field
*
x ) R n which is refereed to regressor, is Gaussian type
(~

~
xC

According to the properties of the RBF neural networks,

x , * ) can be expressed in the form
the function h( ~

Taking the time derivative of V along the trajectories of

eqn.(16), we have
1 T ~ ~ T ~ 1 ~ T ~
V = [ ~
x Px + x Px ] + tr ( )

2
=

~
 ~
1 ~T
1
x ] + [ T ( ~
x ) + ]T BT P~
x + tr (T )
[ x ( PAc + AcT P) ~

~
 ~
1 ~T ~
1
[ x Qx ] ( ~
x ) T B T P~
x + T B T P~
x + tr ( T )

2
(19)

Noting that

~
~
(~
x ) T B T P~
x = tr ( B T P~
x (~
x )T )
if we choose the parameter update law as
~

x PB T ( ~
x)
= ( ~
,
x )~
x T PB c0
M 2
1 if || T || = M and ~
x T PB T > 0
c0 =
0
otherwise

2 2 max ( P)
.
min (Q )
This implies that the positioning error will also belong to
the residue.
to a residue of radius R0 = 0 with :=

(20)

4.
(21)

It can be easily verified that if the initial parameters are

chosen to be inside the ball, i.e., || (0) ||:=
{tr[ (0) T (0) ]}1/2 M, then we have || (t ) || M for
all t 0. Then we have
~  ~
1 T ~ 1
V = ~
x Qx + tr (B T P~
x (~
x ) T + T ) + T B T P~
x

2
~

1 T ~ 1
= ~
x Qx + tr[ ( ~
x )~
x T PB + ) T ] + T B T P~
x

2
~
1 T ~
x T PB T ~ T
(22)
)
= ~
x Qx + T B T P~
x c 0 tr (
2
M 2
Now let's have a look at the last term in the above
equation,
~
x T PB T ~ T
)
c 0 tr (
M 2
~
~
x T PB T
= c0
tr ( T )
2
M
~
x T PB T
= c0
tr ( T *T )
2
M
0
Actually, it is noted that when || || < M or || || = M and
~
x T PB T 0 , then c0 = 0, the above inequality is trivial;
when || || = M and ~
x T PB T > 0 , then, due to the fact
that ||*|| M, we can have ( T *T ) 0 , therefore,
the above inequality is also true. In other words, the
projection will not make the derivative of the Lyapunov
type function more positive. Therefore we have
1 T ~
V ~
x Qx + T B T P~
x
2
1
min (Q) || ~
x || 2 + 0 max ( P) || B || || ~
x ||
2
(23)
1
= || ~
x || [ min (Q ) || ~
x || 2 2 0 max ( P)]
2
where max(P) and mix(Q) denote the maximum
eigenvalue of matrix P and the minimum eigenvalue of
matrix Q respectively.
Consequently, it can be concluded from the eqn.(23) that
the system is convergent and tracking error ~
x will belong

Conclusions

A new stepper motor control scheme is developed in this

paper. The proposed scheme consists of a conventional
feedback linearization controller, which is based on the
known nominal stepper motor dynamics model, and a
compensating controller, which is based on the RBF neural
network. The compensating controller is used to improve
the stepper motor positioning performance. The neural
network is trained on-line based on Lyapunov theory and
learning convergence is thus guaranteed.
References
[1] M. Zribi and J. Chiasson, "Position control of a PM
stepper motor by exact linearization", IEEE Trans.
Automat. Contr., vol.AC-36, pp.620-625, 1991.
[2] M. Ilic-Spong, R. Marino, S. Peresada, and D.G.
Taylor, Feedback linearizing control of switched
reluctance motors, IEEE Trans. Automat. Contr.,
vol.AC-32, no.5, 1987.
[3] D.G. Taylor, M. Ilic-Spong, and S. Peresada,
Nonlinear composite control of switched reluctance
motors, Proc. 12th IEEE Conf. Industrial
Electronics, Milwaukee, WI, Oct. 1986.
[4] J.B.D. Cabrera and K.S. Narendra, Issues in the
application of neural networks for tracking based on
inverse control, IEEE Trans. Automat. Contr.,
vol.44, pp.2007-2027, 1999.
[5] S. Jagannathan and F.L. Lewis, Discrete-time neural
net controller for a class of nonlinear dynamical
systems, IEEE Trans. Automat. Contr., vol.41,
pp.1693-1699, 1996.
[6] R.M. Sanner and J.J.E. Slotine, Gaussian networks
for direct adaptive control, IEEE Trans. Fuzzy
Systems, vol.3, pp.837-863, 1992.
[7] T. Kenjo, Stepping motors and their microprocessor
controls. Oxford, U.K., Clarendon, 1984.
[8] E.J. Hartman, J.D. Keeler and J.M. Kowalski,
"Layered neural networks with gaussian hidden units
as universal approximations", Neural Computation,
vol.2, pp. 210-215, 1990.
[9] J. Park and I.W. Sandberg, "Universal approximation
using radial-basis-function networks", Neural
Computation, vol.3, pp.246-257, 1990.
[10] R.M. Sanner and J.J.E. Slotine, "Gaussian networks
for direct adaptive control", Proc. American Control
Conf. pp.2153-2159, 1991.
[11] T. Holcomb and M. Morari, "Local training of radial
basis function networks: towards solving the hidden
unit problem", Proc. American Control Conf.,
pp.2331-2336, 1991.