See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/251232769

Control of a Permanent Magnet Synchronous Machine Using Neural Network

Backstepping Approach

Article  in  Advanced Materials Research · December 2012

DOI: 10.4028/www.scientific.net/AMR.628.410

CITATIONS READS

0 164

2 authors:

Messaoud Mokhtari Noureddine Golea

University of Batna 2 Université Larbi Ben Mhidi
6 PUBLICATIONS   6 CITATIONS    77 PUBLICATIONS   615 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Adaptive Neural Network Control for Inverted Pendulum Using Backstepping with Uncertainties View project

differents Control Implementation of Permanent Magnet Synchronous Machine using a dSPACE View project

All content following this page was uploaded by Messaoud Mokhtari on 23 April 2017.

The user has requested enhancement of the downloaded file.

Advanced Materials Research Vol. 628 (2013) pp 410-419
Online available since 2012/Dec/27 at www.scientific.net
© (2013) Trans Tech Publications, Switzerland
doi:10.4028/www.scientific.net/AMR.628.410

Control of a Permanent Magnet Synchronous Machine using Neural

Network Backstepping Approach

Messaoud Mokhtari1,a and Noureddine Golea2,b

1
Department of Electronics, University of Batna, Algeria
2
Electrical Engineering Institute, University of Oum El-Bouaghi Algeria
a b
[email protected], [email protected]

Keywords: Control, Lyapunov, Neural network backstepping technique, PMSM.

Abstract. This paper aims to control a nonlinear uncertain system using the neural network
backstepping technique by way of the following steps: the first one consists in localizing and
modelling the unknown nonlinearities depending upon non-measurable states which in their turn
subjected to uncertainties as well as external constraints, the second step is to design a suitable control
law to overcome either the uncertainties and the external constraints. In short, this work comes down
to solve a non linear and uncertain problem with some external constraints, so the suitable solution to
this problem is the use of an adaptive backstepping control capable of eliminating the system
uncertainty, attenuate remarkably the effect of the external constraints and reduce the uncertainty
approximation errors. All these things are performed using the neural-network technique associated
with the backstepping one, i.e., the use of the neural network backstepping approach to control the
permanent magnet synchronous motor.

Introduction
For some time, much of progress has been made in the field of nonlinear control systems [3]. The
backstepping technique belongs to these new rises in this field [5,7], it offers a systematic method to
design a controller for nonlinear systems. The main advantage of this method is to guarantee the
stability of the controller-process relationship. The backstepping method allows getting an adaptive
control law capable of updating the system parameters, and ensuring the stability of the
controller-process relationship.[1,3,6,8]
The backstepping algorithm is a recursive technique for nonlinear control systems which is based
upon the Lyapunov function whose drawbacks are listed below:
• The parameterization: it is assumed to be linear between the non-linear system and the
unknown or uncertain parameters.
• The non-linear functions are assumed to be known.
• In addition to that, the system should be controlled taking into account the following
constraints:
• Beyond the parameter uncertainty and the nonlinearity of the system, we are faced with some
external constraints which are unpredictable in time and amplitude.
• The existence of a nonlinear parameterization between the nonlinear system and the unknown
parameters.
In order to avoid all these constraints an extension of the backstepping algorithm is used. It is based
on neural network technique called Radial Basis Function (RBF). This extension leads to a new type
of control, Known as artificial neural network nonlinear adaptive control (ANNNAC).[4,7,9]

All rights reserved. No part of contents of this paper may be reproduced or transmitted in any form or by any means without the written permission of TTP,
www.ttp.net. (ID: 128.173.127.127, Virginia Tech University, University Libraries, Blacksburg, United States of America-22/07/13,16:35:03)
 Advanced Materials Research Vol. 628 411

Network
RBF

Input
signal Output
Control Process

Fig.1 General diagram of the backstepping-RBF control

Permanent Magnet Synchronous Motor (PMSM)

In order to design a control system we need to know all about the system structure namely its
mathematical model. In our case, the system to be controlled is the synchronous permanent magnet
motor whose model is given below. The permanent magnet synchronous motors have experienced an
increased attention especially in training applications thanks to their raised power-factor, higher
power density, higher torque in comparison to inertia, big energy compared to its weight, big
effectiveness, and robustness. All these advantages are due to the technological permanent magnets
improvement.
After the establishment of the synchronous motor model and the application of the backstepping
technique, the simulation of the system is carried out [7].
The stator currents create a magnetic rotary field whose speed is given by:
ω f
ωr = or nr = (1)
p p
with:
ωr : speed of the rotary field pattern in rad/s (mechanical speed);
ω : speed of alternating currents in rad/s, ω =2 π.f (electrical speed);
nr : speed of the rotary field pattern in rps;
f : frequency of alternating currents in Hz;
p : number of pole pairs.

Modeling of the Synchronous Machine

Consider the following transformed voltage system of equations. [10]
ϕ d = − Rs .id + p.ω r .ϕ q + vd (2)

ϕ q = − Rs .iq − p.ω r .ϕ d + vq (3)

For a permanent magnet machine we have:

φ = L .I + φ (4)
d d d m

φ = L .I (5)
q q q
with:
φ : Inductor flux generated by permanent magnets.
m
412 Manufacturing Engineering and Technology for Manufacturing Growth

The electromechanical equation of the machine is given by:

dω r
Te = Tl + J . + f .ωr (6)
dt
For a sinusoidal electromotive force with the flux imposed by permanent magnets, the
electromagnetic torque is given by:

Te =
3. p
2
( ( )
. φm I q + Ld − Lq I d I q ) (7)

Without considering saliency and damping, i.e. Ld = Lq we get:

3. p
Te = .φm I q (8)
2
The torque depends only on the component along the q axis of stator current. The power input for a
given torque is optimal for the value of I d = 0 , provided that the torque is regulated by I q . If the motor
is either shaped with a direct salience, Ld > Lq or an inverse one Lq > Ld , then the torque depends
respectively upon I q or I d . In the case the permanent magnet motors only I d is taken into account for
the sake of decreasing the flux within the motor.
According to the Park model, the stator currents and the mechanical speed are used as state
variables in the mathematical representation of the permanent magnet synchronous motor, the stator
voltages are used as control inputs.
dI d
Vd = Rs .I d + Ld . − p.ω r .Lq .I q (9)
dt

diq
Vq = Rs .I q + Lq . + p.ωr .Ld .I d + p.ωr .φm (10)
dt
with :
Vd , Vq : stator and rotor voltages along d and q axes;
I d , I q : stator and rotor currents along the d and q axes;
Rs : per-phase stator resistance;
Ld , Lq : stator inductances along the d and q axes;
Tl , Tl : electromagnetic and load torques;
J : moment of inertia of the motor-load set;
f : friction coefficient of the motor;
p : pole pair of the motor;
ωr : rotor angular velocity;
φm : permanent magnet flux.
Notice that the electromagnetic torque is proportional to the current along the q axis of smooth pole
machine ( Ld = Lq ).
It is easy to see that the speed control can be carried out by controlling the Vq component of the
supply voltage by setting the value of I d to zero. Taking into account this assumption, the system of
equations can be reduced as follows: [2]
Vd = − p.ωr .Lq .I q (11)
 Advanced Materials Research Vol. 628 413

dI q
Vq = R.I q + Lq . + p.ωr .φm (12)
dt

dω r
Te = Tl + J . + f .ωr (13)
dt

3. p
Te = φm .I q (14)
2
Notice that this assumption can be neither reasonable nor practical, what makes us obliged to look
for a practical solution to cancel the current I d when using the complete model of the machine.[2]
According to Park, the PMSM model can be described by the following equations:
 dω
 r =  (φm I q + (Ld − Lq )I d I q − Tl − fωr )
1 3 p 
 dt J 2 
 dI 1

d
= (Vd − Rs I d + pωr Lq I q ) (15)
 dt L d
 dI q 1
 = (Vq − Rs I q − pωr Ld I d − pωrφm )
 dt Lq

Let:
f 1 T 3
a1 = , a2 = , a3 = l , µ = p 2 (16)
J J J 2
Provided that: ω = p.ωr

with:
ω : electric speed;
ωr : rotor angular velocity.
Hence the following system of equations is obtained:
 dω
 = −a1ω + a2 µ (φm + (Ld − Lq )I d ) I q − p.a3
 dt
 dI q R L φ 1
 = − I q − d I d .ω − m ω + Vq (17)
 dt Lq Lq Lq Lq
 dI R L 1
 d = − I d + q I q .ω + Vd
 dt Ld Ld Ld

The choice of the nonlinear functions rests on the principle of avoiding the various couplings
between the state variables. Consequently, the following expressions are adopted:

ϕ1 = a2 µ .(Ld − Lq )I d I q ; ϕ 2 = −
Ld L
I d .ω ; ϕ3 = q I q .ω (18)
Lq Ld
414 Manufacturing Engineering and Technology for Manufacturing Growth

Equation (17) becomes:

 dω
 = −a1ω + ϕ1 − p.a3 + a2 µφm I q
 dt
 dI q R φ 1
 = − I q + ϕ 2 − m ω + Vq (19)
 dt Lq Lq Lq
 dI R 1
 d = − I d + ϕ3 + Vd
 dt Ld Ld

The nonlinearities of the system ϕ1 , ϕ 2 , ϕ 3 are uncertain.

The approximation technique of the nonlinear functions using the RBF implies the following
definitions:
ϕ1 = ϕ1*T .W + ϕ~1 (20)

ϕ 2 = ϕ 2*T .W + ϕ~2 (21)

ϕ3 = ϕ3*T .W + ϕ~3 (22)

Such that:
ϕi*T : function to be defined by the RBF ;
W : weight ;
ϕ~i : approximation error
The new representation will be then:
 dω
 = − a1ω − p.a3 + ϕ1*TW + ϕ~1 + a2 µφm .I q
 dt
 dI q R φ 1
 = − I q − m ω + ϕ 2*TW + ϕ~2 + Vq (23)
 dt Lq Lq Lq
 dI R 1
 d = − I d + ϕ3*TW + ϕ~3 + Vd
 dt Lq Ld

Based on reference [4], we obtain:

m1
ϕ1*TW = ∑Wi ,1 . ϕ1*,i ( ω (k ) − ξ i ,1 ) (24)
i =1

m2
ϕ W = ∑Wi , 2 . ϕ 2*,i ( I q (k ) − ξi , 2 )
*
2 (25)
i =1

m3
ϕ3*W = ∑Wi ,3 . ϕ3*,i ( I d (k ) − ξ i ,3 ) (26)
i =1
with:
ϕi*, j : radial base function (in our case Gaussian’s are chosen)
ξi, j : centers;
Wi , j : weights;
mi : number of chosen neurons.
 Advanced Materials Research Vol. 628 415

In reference to [4], we get:

1  ω (k ) − ξi ,1 
ϕ1*,i = exp  (27)
2πλ  2λ
2


1  I (k ) − ξ i , 2 
ϕ 2*,i = exp q  (28)
2πλ  2λ2 

1  I (k ) − ξ i ,3 
ϕ3*,i = exp d  (29)
2πλ  2λ2 
with:
λ : RBF design parameter

Backstepping Procedure
The errors are defined by the following expressions:
 z1 = ω * − ω

 z2 = I q − I q
*
(30)

 z3 = − I d
For the flux control, the chosen reference is carried out according to I d* = 0
The error dynamics is defined by:
 z1 = ω * − ω

 z2 = I q − I q
*
(31)

 z3 = − I d
Using Eq. 23, Eq. 31 becomes:

 ~
 z1 = ω + a1ω + p.a3 − ϕ1 W − ϕ1 − a2 µφm I q
* *T

 R φ 1
 z2 = I q +
*
I q + m ω − ϕ 2*T W − ϕ~2 − Vq (32)
 Lq Lq Lq
 R 1
 z3 = I d − ϕ3*TW − ϕ~3 − Vd
 Lq Ld

First step:
To determine the system stability, the first Lyapunov function is defined:
1 2
V1 = z1 (33)
2
Using Eq. 30 and Eq. 32, the derivative of the previous function leads to:
V = z z = z (ω * + a ω + p.a − ϕ *T W − ϕ~ − a µφ I )
1 1 1 1 1 3 1 1 2 m q
(34)
= z1 (ω + a1ω + p.a3 − ϕ W − ϕ~1 − a2 µφm I q* + a2 µφm z 2 + k1 z1 − k1 z1 )
* *T
1
416 Manufacturing Engineering and Technology for Manufacturing Growth

So, the first stabilizing function is obtained:

1
I q* = (ω * + a1ω + p.a3 − ϕ1*TW + k1 z1 ) (35)
a2 µφm
and Eq. 34 becomes:
V1 = − k1 z12 + a2 µφm z1 z 2 − z1ϕ~1 (36)
Then, we can conclude:
z1 = − k1 z1 + a 2 µφm z 2 − ϕ~1 (37)
Second Step:
The second Lyapuntov function is defined as follows:
1 2
V2 = V1 + z2 (38)
2
From Eq. 32, the derivative of the previous equation is given by:
R φ 1
V2 = V1 + z 2 z 2 = V1 + z 2 ( I q* + I q + m ω − ϕ 2*TW − ϕ~2 − Vq ) (39)
Lq Lq Lq
The derivative of Eq. 35 is obtained as follows:
I q* =
1
a2 µφm
[ ( )
ω * + a1 − a1ω − p.a3 + ϕ1*T W + ϕ~1 + a2 µφm I q − ϕ1*T W + k1 (− k1 z1 + a2 µφm z 2 − ϕ~1 ) (40) ]
Then:
V = − k z 2 + a µφ z z − z ϕ~
2 1 1 2 m 1 2 1 1

 1
+ z2  [
ω * + a1 (−a1ω − p.a3 + ϕ1*T W + ϕ~1 + a2 µφm I q ) (41)
 a2 µφm
φ 1 
− ϕ1*T W + k1 (− k1 z1 + a2 µφm z 2 − ϕ~1 ) +
R
Lq
]
I q + m ω − ϕ 2*T W − ϕ~2 − Vq 
Lq Lq 

The control law Vq is obtained as follows:


Vq = Lq k 2 z 2 + a2 µφm z1 +
a
1
µφ
[
ω * + a1 (−a1ω − p.a3 + ϕ1*TW + a2 µφm I q )
 2 m
(42)
φ 
R
]
− ϕ1*TW + k1 (− k1 z1 + a2 µφm z2 ) + I q + m ω − ϕ 2*TW 
Lq Lq 

a1 k1
V2 = − k1 z12 − k2 z22 − z1ϕ~1 + ϕ~1 z2 − ϕ~2 z2 − z2ϕ~2 (43)
a2 µφm a2 µφm
So, we can deduce:
a1 k1
z 2 = −k 2 z 2 + ϕ~1 − ϕ~2 − ϕ~2 (44)
a2 µφm a 2 µφm
 Advanced Materials Research Vol. 628 417

Third Step:
The third Lyapunov function is defined by:
1 2
V3 = V2 + z3 (45)
2

R 1 
V3 = V2 + z3 z3 = V2 + z3  I d − ϕ3*TW − ϕ~3 − Vd  (46)
L Ld 
 q
The control law Vd is obtained as follows:

 R 
Vd = Ld  k3 z3 + I d − ϕ3*TW  (47)
 Lq 
 
Then, we can finally obtain:
 a1 ~ k1 ~ ~ 
V3 = − k1 z12 − k 2 z 22 − k3 z32 − z1.ϕ~1 − z 2  − ϕ1 + ϕ 2 + ϕ 2  − z3ϕ~3 (48)
a
 2 m µφ a 2 µφ m 
What guarantee that the system is stable for zi → 0 when t → ∞ .

Simulation Results
The adaptation gains are: k1=0,001; k2=100; k3=1500. For the RBF network, the design parameter is
λ =1,4. The RBF centers are ξi,1=[0;1,5;2];ξi,2=[0;1,5;2];ξi,3=[0;1,5;2].The PMSM parameters are:
R=2 [Ω], Ld=0.04244 [H], Lq= 0.07957 [H], p=1, J=0.03 [g.m2], f=0.001 [Nm/rad/s], φm =0.311 [Wb],
the load torque is 20 [N.m], and ω * = 100 [rad/sec].
By applying the classical backstepping control, the system is likely to diverge if a small
modification of the nonlinear function ϕ i appears, but the implementation of the RBF approximation
strategy allows obtaining the results displayed in the following figures.

150
140

120

100
100
80
Speed w

Error z1

60

40
50

20

0 -20
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
time(s) time(s)

Fig. 2 Rotor electrical angular velocity Fig. 3 Evolution of the error

418 Manufacturing Engineering and Technology for Manufacturing Growth

1
20

0.8

15 0.6

0.4

10 0.2

Current id
Current iq

5 -0.2

-0.4

0 -0.6

-0.8

-5 -1
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
time(s) time(s)

Fig. 4 Current distribution via q axis Fig. 5 Current distribution via d axis

5 120

110
0

100
-5
90
-10
Control input vq
Control input vd

80

-15
70

-20 60

-25 50

40
-30
30
-35
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
time(s) time(s)

Fig. 6 Control voltage waveform via d axis Fig. 7 Control voltage waveform via q axis

Conclusion
The use of neural network technique has permitted a good estimation of the nonlinear functions. The
coupling between the backstepping approach and the neural network one has given some satisfactory
results by ensuring the system convergence (stability) even if the settling time of nearly two seconds is
considered a little bit long.

References
[1] A. Feurer and A.S. Morse, Adaptive control of single-input, single-output linear system, IEEE
Transactions on Automatic Control, Vol. 23(4), pp. 557-569 , 1978.
[2] N. Goléa, A. Goléa and M. Kadjouj, backstepping Adaptive Control of Permanent Magnet
Synchronous Motor Drive. EE Institute, Oum El-Bougahi University,04000 , Oum El-Bouaghi,
Algeria, 2004.
[3] I. Kanellakopoulos, P.V. Kokotović, and R. H. Middelton, Indirect adaptive output-feedback
control of a class of nonlinear systems, Proceedings of IEEE Conference on Decision and
Control, , Honolulu, Hawaii, pp. 2714-2719, 1990.
[4] T. Knoho, and H. Unbehauen, ANNNAC-extension of adaptive backstepping algorithm with
artificial neural networks, IEE Proc., Control Theory Appl.,Vol. 147(2), pp.177–183,March
2000.
 Advanced Materials Research Vol. 628 419

[5] P.V. Kokotović, Lecture notes in control and information sciences: Foundations in adaptive
control, Springer Verlag, Berlin, 1991.
[6] P.V. Kokotović and H.J. Sussmann, A positive real condition for global stabilisation of nonlinear
systems, Systems and control letters, Vol. 13, pp. 125-133 , 1989.
[7] M. Kristić, I. Kanellakopoulos, and P.V. Kokotović, Nolinear and adaptive control design, John
Willey and Sons, New York, 1995.
[8] J. Tsinias, Sufficient Lyapunov-like conditions for stabilization, Math. Contr. Signal, pp.
343-357 , 1989.
[9] D. Wang and J. Hung, Adaptive neural network control for a class of uncertain nonlinear systems
in pure-feedback form, Automatica, Elsevier Science, pp. 1365-1372 , 2002.
[10] R. Abdessmed, Modélisation et simulation des machines électriques, ellipses edition,
pp.104-130 , 2011.
 Manufacturing Engineering and Technology for Manufacturing Growth
10.4028/www.scientific.net/AMR.628

Control of a Permanent Magnet Synchronous Machine Using Neural Network Backstepping Approach
10.4028/www.scientific.net/AMR.628.410

View publication stats