IEEE Transactions on Energy Conversion,Vol. 13, No.

4, December 1998 311

On-line Adaptive Artificial Neural Network

Based Vector Control of Permanent Magnet
Synchronous Motors ’

M.A Rahman, Fellow IEEE and M.A. Hoque

Memorial University of Newfoundland
St. John’s, Nfld., Canada A1B 3x5, Fax (709)737-4042
Abstract- This paper presents a novel approach of speed synchronous motor which is also known as the brushless dc
control for a permanent magnet synchronous motor (PMSM) motor, and (ii) the sinusoidal wave fed permanent magnet
using on-line self tuning artificial neural network (ANN).
Based on motor dynamics and non-linear load characteris- synchronous motor which is known as brushless permanent
tics, an A N N speed controller is developed and integrated magnet synchronous motor (in short, permanent magnet
with the vector control scheme of the PMSM drive. The synchronous motor).
combined use of off-line and on-line weights and biases ad-
justments offers a unique feature of on-line system iden- Fast and accurate response, quick recovery of speed from
tification and precise speed control of a high performance any disturbances and insensitivity to parameter variations
PMSM drive. The complete drive system is implemented in are some of the main criteria of high performance drive sys-
real time using a digital signal processor controller board-
DS1102 on a laboratory 1 hp PMSM. using the experimen- tems used in robotics, rolling mills, machine tools, etc. In
tal setup, the performances of the proposed drive system these applications, equivalent performance characteristics
are evaluated under various operating conditions. The test of a separately excited dc motor can be obtained from the
results validate the efflcacy of the A N N for precise speed
control of the PMSM drive. Furthermore, the use of A N N PMSM if the closed loop vector control scheme is employed
makes the drive system robust, accurate and insensitive to [a]. The PMSM drive system involving the vector control
parameter variations. scheme not only decouples the torque and flux which prc-
vides faster response but also makes the control task easy.
The speed controller used in PMSM drive system plays an
KeywoTds- Artificial Neural Network, Permanent Magnet
Synchronous Motor, On-line System Identification, Hystere- important role to meet the other required criteria of the
sis Current Controller, Speed Control HPD. It should enable the drive to follow any reference
speed taking into account the effects of load impact, satu-
ration and parameter variations. Conventional controllers
I. INTRODUCTION such as proportional integral (PI) or proportional integral
DC motors have been widely used as precision control mo- derivative (PID) have been widely used in both dc and ac
tors till mid 1970s. There are certain limitations of dc mo- motor controls. But these types of controllers are diffi-
tor drives such as lack of robustness, overload capability, cult to design if an accurate system model is not available.
narrow speed range, and frequent maintaince requirements, Moreover, unknown load dynamics and other factors such
particularly due to brushes and commulators. Need to find as noise, temperature, saturation, etc. affect the perfor-
alternatives of using dc motors in the modern drive system mance of these controllers for wide range of speed opera-
has been advocated over many years. Recent developments tions [3]. However, there existCSsome adaptive controllers
in microprocessors, magnetic materials, and semiconductor such as model reference adaptive controller (MRAC) [4],
technology have offered an excellent opportunity to use ac sliding mode controller (SMC) [5], variable st,ructure con-
motors in high performance drive [HPD] systems. Among troller [6], self tuning regulator [7,8] which have been used
the various types of ac motors, the PMSM is becoming a in the motor drive with large number of unknown param-
very popular choice in drive technology over the last few eters. In addition, these types of controllers are usually
years due to some of its inherent advantageous features [l]. based on system model parameters. Hence, unavailability
These features include high torque to current ratio, large of the accurate system model often leads to a cumbersome
power to weight ratio, higher efficiency, and robustness. design approach for these controllers.
Depending on the control strategy with the inverter sup- Now-a-days, considerable attention has been focused on
ply system, the ac permanent magnet synchronous motor the use of artificial neural network on the system modeling
(PMSM) can be classified in two categroies; (i) the rectan- and control applications [3,9-181. In high performance mo-
gular wave electronically commutated permanent magnet tor drives including the PMSM, the ANN can play a key
role on system identification and speed control, because
the ANN has many advantageous features including par-
PE-098-EC-2-06-1997 A paper recommended and approved by the allel and distributed processing, efficient non-linear map-
IEEE Electric Machinery Committee of the IEEE Power Engineering ping between inputs and outputs without an exact system
Society for publication in the IEEE Transactions on Energy Conversion.
Manuscript submitted April 30, 1996; made available for printing June model. Although there are some works reported recently
11, 1997. for the dc motor control using the ANN [9,12-151,but its
use in the ac motor control is in its initial stage [3,16-

0885-8969/98/$10.00 0 1997 IEEE

312

191. Sharkawi et al. [3] have implemented an ANN based pole pairs. The developed electric torque can be expressed
brushless dc motor control. Their work is based on the off- as
line trained ANN with an indirect model reference adaptive
control technique. Shiguo et al. [16] proposed a digital con- (3)
L
trol scheme for a brushless dc servo system with an analog
speed controller to obtain the better servo performance. and the motor dynamics can be represented by
An ANN has been used by Theocharis and Petridis [17]
for the estimation of state variables for an induction motor T'e = JpWr + Bur + TL (4)
drive. Burton e2 al. [18]proposed an on-line random train-
ing algorithm for the stator current control of an induction where TL is the load torque and B is the damping co-
motor using the ANN where the reference model is used efficient and J is the rotor inertia. The load torque in
to train the network in real time. Ben-Brahim et al. [19] (4) can have any non-linear unknown mechanical charac-
proposed a neural network based speed estimator for the teristics. The following equat'ion can be used to model a
vector control of an induction motor where reference model non-linear load [ 141
and neural network model have been used to generate the
error for the on-line back propagation training algorithm.
z= K1W,2 + I<2w, + IC3 (5)
So far no work has been reported on the on-line self-tuning where K 1 , IC2 and I(3 are constants. An efficient control
ANN based speed control of the PMSM employing the vec- strategy of the vector control technique is to make the d- -
tor control scheme. The dynamic model of an interior type axis current id zero so that the direct axis flux linkage Ad
PMSM is relatively complex due to the presence of magnet becomes dependent only on the flux linkage by the perma-
brake torque [l]. Thus there exists a dearth of works on nent magnet rotor. With this control strategy, the machine
the application of artificial neural networks in the speed model becomes simpler as can be described by the following
control of the PMSM. equations
In this work, an ANN structure is developed based on the
inverse dynamics of the vector-controlled PMSM. A novel 1
pi, 1 -(U, - Ri, - I<~w,) (6)
feature of adaptive on-line weights and biases up-dating of L,
the ANN has also been included in this work. A laboratory 1
set up is built and the performances of the proposed control pT -[Te - Ti - Bw,] (7)
J
scheme are evaluated in a 1 hp PMSM in real time using a
DSP controller board. where I<b = PA, and the developed torque T, in ( 4 ) is
proportional to the q-axis current as given by
11. INVERSE OF P M S M
DYNAMICS
T,= ICTi,
The use of the ANN in any system does not require the where KT = %A., Equation ( 8 ) resembles the torque
model of the system under study. The inputs and outputs expression of a separately excited dc motor, where i, cor-
of the ANN can be arbitrarily chosen from the system vari- responds to the armature current of a dc machine. Hence,
ables. However, it is preferred to model the dynamics of a precise torque control of the PMSM is possible by con-
the drive system so that inputs and outputs of the ANN trolling the q-axis current i,. Now to make the control task
can be selected in a more defined way [9]. This guarantees easier, the equations of a PMSM can be expressed as a sin-
that the ANN will capture the system dynamics. gle input single output system in continuous time domain
In this work, unlike the conventional approach of control- by combining (5)-(8), giving
ling the speed, the vector control technique is incorporated
with the ANN to obtain highest torque sensitivity of the
PMSM drive. The vector control strategy is formulated
in the synchronously rotating reference frame. The math-
ematical models of a PMSM and load are determined by
equations in the d-q synchronously rotating reference frame K,RW,!(t) + K s R - I < ~ v , ( t )= 0 (9)
as la01
The discrete time model of the simplified PMSM drive can
Uq Riq + PA, + PWTAd (1) be obtained by replacing all continuous quantities by their
vd = Rid + PAd - P W , A q (2) finite difference, giving

+
where A, = L,i, and Ad = L d i d A m , U, and V d are the wr(n + 1) = n w r ( n ) + p w r ( n - 1) + yw,2(n) +
d,q-axis voltages, i d and i, are the d,q-axis stator currents, 6w,!(n - 1 ) + f.,(.) + 29 (10)
Ld and L , are the d,q-axis inductances, Ad and A, are the
d,q-axis axis flux linkages, R is the stator resistance , A, is where a ,p, -y, 6,t and 8 are functions of the motor param-
the constant magnet flux linkage produced by permanent eters as well as sampling interval which are given in Ap-
magnet rotor, w, is the motor speed and P is the number of pendix A. The numerical value of the motor parameters
 313
are also given in Appendix B. Equation (10) is modified in a cost function of the error between the outputs and the
order to obtain the inverse model of the drive system as target of the ANN 1211. Depending on the applications,
the training of the ANN could be off-line or on-line. If
.p(n) = [Ur(n + 1) - a w r ( n ) - Pur(. - 1) - y ~ , 2 ( n ) the weights and biases of the ANN is determined through
-6w,2(n - 1) - d ] / € (11) off-line training only then an intensive training has to be
performed considering almost all operating conditions of
Now in discrete form, the q-axis current can be expressed the system which are almost impossible for the control of a
in terms of v e ( n )and U , by replacing the continuous terms PMSM. As for example in reference [14], the load is mod-
of equation (6) by their finite differences, giving eled by (5) which is not always true in practical situations.
Hence, the need for on-line weights and biases up-dating
iq(n) = Aiq(n - 1) + Bzvq(n) + CUT(.) (12) arises. However, the on-line training task could be eased
and the system can be made more stable if an initial set of
where
weights and biases is generated a priori through the off-line
training. A combination of the off-line and on-line train-
A=1--
RAT
, B 2 = -A- ,T c=--ATKb ings has been used for the
L, Lq L,
present work. The initial weights and biases are obtained
and A T is the sampling interval. Thus the expressions for through the off-line training. These are up-dated only
the q-axis current can further be modified as when an error limit between the actual output and the
target of the ANN exceeds a preset value.
i q ( n ) = A i q ( . - 1) + B z [ w , - (-~ )

- P U r ( n - 2) - T W ; ( ~ )-
(CY
3
+-
6 ~ , 2 (-
-
w,.(~z 1)

n 2) -
............................. . .........................
.

(13)
The right hand side of (13) is a non-linear function of the
speed w,.. The purpose of using the ANN is to map the non-
linear relationship between the q-axis current ip(n) and the
Hidden-layer 0utpu t-layer
speed w,(n) according to (13). Equation (13) reveals the
structure of the ANN for the speed control of the PMSM.

111. ANN STRUCTURE FOR PMSM Fig. 1. ANN structure for PMSM
One of the important aspects of applying an ANN to any
particular problem is to formulate the inputs and output(s) IV. OFF-LINETRAINING
of the ANN structure under study. The inverse dynamics Data fc the off-line training can be obtained eithe by
of the PMSM as described in equation (13) dictates the simulation or by experiment, such as running the motor
inputs and output of the ANN used in the control sys- on open loop basis and recording input currents and out-
tem. According to (13), the inputs of the proposed ANN put speeds at different operating conditions if the motor
are the speed of the motor at the present and previous parameters are not available. If the motor parameters are
two sample intervals in addition to the previous sample available, training data can be generated using (13) where
of the q-axis current. Therefore the input vector becomes the input pattern is [w,(n),w,(n- l ) , u r ( n - 2 ) , i q ( n - l ) ] .
[wr(n),w,.(n - l ) , u r ( n - 2 ) , i q ( n- I)]. The corresponding The corresponding target i q ( n ) can be generated by us-
output target is the present sample of q-axis current i,(n). ing this input, vector and CY, /?,y,6 , c , 29 and A , B , C in the
After the inputs and output are formulated, the next step right hand side of equation (13). For this present work,
is to incorporate the hidden layer(s). Number of hidden off-line training data are obtained by simulation on open-
layers and number of neurons in the hidden layer are cho- loop PMSM motor control scheme by considering the load
sen by trial and error, keeping in mind that the smaller as described by (5).
the numbers are, the better it is in terms of both memory
and time requirement to implement the ANN in the motor The simulation is carried out at random q-axis currents
control. For the present work, the structure of one hidden in order to obtain various speeds according to the specified
layer having three neurons gives satisfactory results. The currents. At first, the simulated speed and voltage data
ANN structure for the PMSM is shown in Fig. 1. The are down sampled by 50 times and finally two sets of data
transfer functions used in the hidden and output layers are with 800 samples are obtained. A window length of four
log-sigmoid and tan-sigmoid, respectively. Once a design , ( n- 1))U,. ( n-a), ie( n - l)]is slid over
containing [U,. ( n ) U,.
of the ANN structure is done, the next step is to deter- the entire speed and q-axis current samples which give an
mine the weights and biases of the ANN through training input matrix of size 4 x 800. The output vector is the size
to achieve the specific target with the given inputs. The of 1 x 800, containing the q-axis current data. The off-line
back-propagation training algorithm is used for this pur- training of the ANN is performed with these training data
pose which is based on the principle of minimization of using the back-propagation algorithm of the Matlab neural
 314

Fig. 2. Block diagram of the proposed ANN based PMSM drive

network tool-box [22]. The initial set of weights and biases The reference speed trajectory is selected using a second
obtained through off-line training is given in Table 1. The order reference model that makes the system asymptoti-
training data were generated using (13). The CPU time cally stable [lo]. The reference model can be described by
taken for the off-line training to generate a quick set of the following equation
initial weights and biases is one hour.
U:(.) = U ~ W ; ( T Z- 1) + u ~ w : ( . - 2) + r ( n ) (14)
IV. REAL-TIME
A N N BASEDPMSM
where a1 and a2 are constants chosen for a specific refer-
The main goal of the control system is to track reference ence trajectory which ensures that the poles and zeros of
speeds by providing the appropriate i, depending upon the the reference model are within the unit circle; r ( n ) is the
operating conditions. In real time, using the rotor position bounded input to the reference model. The value of r ( n )
information and the output, of t,he A N N which is considered can be calculated beforehand using (14) for a desired ref-
now as q-axis reference current if, are used to generate the erence speed U:(.). If the tracking error is assumed to be
three phase reference currents i:, i l and iz. This is accom- small and since the selected reference model
plished by Park’s axis transformation [20]. The three phase is asymptotically stable, the rotor speed at the nth sample
reference currents are compared with actual motor currents can be predicted from (14) as [lo]
in a hysteresis current controller which provides necessary
drive signal for the inverter. The complete control scheme W;.() = ( L ~ W , ( ~-I 1) + n 2 ~ , ( 7 i- 2) + ~(71) (15)
is shown in Fig, 2.
Hence in addition to previous sample of predicted q-axis
Table 1: Iilitid set of weights and biases current, with one sample of predicted speed and two sam-
ples of actual speed, an input sequence of [Gr(n),w,(n-
NW1 l),w,.(n - 2),i^,(n- l)] is formed and used as the input
-0.4266 0.2340 0.0410 to the ANN as shown in Fig. 2. With the initial weights
0.4397 0.6552 0.5160 and biases and this sequence of input and the output of
the A N N , i^g(n)is computed. Using actual speed samples
-0.7329 0.0680 -0.4717 and one delayed sample of q-axis current, another input
NW2 sequence containing [w,(n),w,(n - l ) , w r ( n- 2),it(n- I)]
is formed to generate the output i ; ( n ) . This ii(n)is com-
-3.7272 -0.9813 -3.7907
pared with ;q(n) in order to compute the error which is
NB1
-0.1266 0.5937 1.4648

NB2
3.4837
 315
used to update the previous set of weights and biases using Net input of the kth neuron of the output layer at time
the back-propagation algorithm. Reference speed w: ( n ) instant n is given by
can be used as the sole speed command in the proposed
M
ANN based PMSM drive system. However, to ensure over-
all stability, the reference model of the bounded input ~ ( n ) S i ( n )= W;&Z)Ojh(n) (22)
j=1
has been selected as suggested in reference [12]. This pro-
vision of the reference model always forces the system to where M is the number of neurons in the hidden layer,
follow the bounded input by providing an error between Wjo,(n)is the weight between j t h neuron at the hidden
the actual and estimated driving signals of the reference layer and kth neuron at the output layer. Therefore, output
q-axis current. It is noted here that the two ANNs shown from the lcth neuron at the output layer at time instant n
in Fig. 2. have the same set of weights and biases with two can be written as:
different sets of inputs and outputs.
The weights and biases are up-dated at each instant us- G ( n )= f"(S,"(n) + Bko(n)) (23)
ing improved back-propagation algorithm. The minimized
error function is given by where f " is a non-linear transfer function. This is acted
upon by each neuron at the output layer. These are de-
l fined in (20) and (21). B,"(n)is the bias of the lcth neuron
~ ( n=) -e2(n)
2 (16) at the output layer.
(a) Output layer weights and biases updating
where

e ( n ) = error = ;^,(?I) - i i ( n ) Weights and biases of the output layer are up-dated as:
(17)
Like the conventional feedback speed control approach w g n + 1) = Wf&, + ?/s,"(n)op(n) (24)
where difference equation of the speed acts on the digi- B,"(n+ 1) = B,"(n)+ qS,"(n) (25)
tal speed controller, a sequence of speed samples as men-
tioned earlier are used as the input to the proposed speed where 77 is the learning rate and 6;(n) is the local gradient
controller. One output of the ANN is the reference q-axis at the Output layer and is expressed
current which drives the current controlled voltage source 6e(n)
transistorized inverter. Another output is the estimated 6i(n)= e(71)- - S,O(n))Ojh(n)
sg(?t)(l (26)
q-axis current which is used to compute the error. Thus *Oi(n)
the two ANNs shown in Fig. 2 have the same weights and (b) Hidden layer weights and biases updating
biases with different input sequences to compute the refer-
ence q-axis and the estimated q-axis currents, respect,ively. The updated weight at the hidden layer can be written as:
On-line weights and Bias Updating
+
W&(n 1) = Wh(n) ?/Sjh(n)li(n)+ (27)
The equations required to update the weights and biases Bjh(n + 1) = Bt(7t) + $jh(n) (28)
can be summarized as follows.

the time instant 11 is given by

N
s,"(.) = W$('?%)li(71) (18) In real time implementation, error is calculated at each
i=l instant and when it goes beyond a predetermined level,
weights and biases updating procedure enables; If the er-
where q$is the weight between ith neuron at the input ror is within prescribed level, the previous set of weights
layer and the j t h neuron at the hidden layer, ri is the ith and biases is kept intact to compute the q-axis reference
input and N is number of inputs. The output from the j t h current.
neuron from the hidden layer at nth instant is given as:
V . LABORATORYIMPLEMENTATION
Ojh(n, = f h ( S i " ( 7 l ) + Bjh(n)) (I9) The complete digital control scheme of the A N N based in-
-
where B; is the bias of the j t h and f h is the non- terior type PMSM is implemented through software using
linear transfer function acting on each neuron at the hid- a DSP controller board DS1102. The laboratory is
den layer. This function could be tan-sigmoidal or log- shown in Fig. 3.
sigmoidal which can be defined as The controller board has the TMS320C31 main proces-
sor. The board is installed in a PC-AT with uninterrupted
communication capabilities with the latter through dual
(20) port memory. The feedback signals of the DSP controller
1 board are the three phase currents and motor shaft angular
logsig(r) = -
1 + e-" (21) positions.
316

The motor line currents are measured by the Hall-effect for implementing the proposed ANN based PMSM drive
sensors. The sensed currents are converted to voltage sig- system is 5 kHz. Since the initial set of weights and biases
nals which are then converted into digital values by the have been generated apriori, the error in (17) was found
A/D sub-system module for using in the current control mostly to lie below the prescribed level of 0.01. Thus up-
loop. The motor shaft-position is measured by an optical dating was not necessary at every instant. As a result the
incremental encoder installed at the motor shaft. The en- control of the ANN based PMSM drive was found adequate
coder generates the necessary pulses according to the rotor in test cases.
position information which are fed to the built-in encoder
port of the controller board. A 32-bit counter is used to VI. RESULTSA N D DISCUSSIONS
count the encoder pulses and is read by a calling function
in the software. The motor speed is deduced from the fairly Several tests were performed to evaluate the performances
accurate rotor position by backward difference interpola- of the proposed ANN based vector control of PMSM. The
tion. speed responses are observed under various operating con-
A digital moving average filter w a s used to remove the ditions such as change in reference speed, step change in
noise in the speed signal. The input sequence of the ANN is load, parameter variations, etc. Some of the sample results
generated from the actual speed and delayed sample of the are presented in this paper.
output of the ANN. At each sampling instant, the weights Experiments were performed to obtain the speed responses
and biases are updated when necessary. The weights and for change in reference speed with full load. Fig. 4(a)
biases of the ANN are used to generate the q-axis reference shows the reference and actual speeds. Due to the lim-
current which is converted to three phase reference currents itation of regeneration of the inverter, reference speed is
using the rotor position angle. The measured actual motor changed only from low to high. The reference speed for
currents and three phase reference currents are used in the this case was 900 rpm before the step change in speed,
software implemented hysteresis current controller. The and was 1800 rpm after the step change in speed. It is
four-channel D/A subsystem is used to output the rotor seen from the Fig. 4(a) that the proposed system quite ac-
position angle and three switching signals corresponding curately tracks the reference speed with no over-shooting.
to each of the three phases of the inverter. An external Test has also been performed to evaluate the performances
electronic circuit has also been built up to generate six base of the of the proposed control system under sudden load
drive signals from the four D/A outputs for the transistor impact. The motor was running at no load conditions and
inverter. Proper commutation signals are generatled in this after some time the rated load of 3.5 N-m is applied sud-
hardware circuit so that two transistors can not switch-on denly. Fig. 4(b) shows the speed response of the proposed
at the same time. system under this operating condition. The ANN adjusts
The complete algorithms are implemented in software its weights and biases to this changing circumstance of sud-
using a high level programming in “C”. The program is den load impact and provides appropriate q-axis reference
compiled using Texas Instrument “C” code generator and current so that the system responds according to the ref-
then down loaded to the DSP controller. The sampling rate erence speed. It is quite evident from Fig. 4(b) that unlike
the conventional controller, the on-line adaptive property
of ANN reduces the possibility of large speed drooping due
n to the sudden application of load.
The performances of the proposed scheme are also eval-
vb uated experimentally with two important parameter varia-
tions: (i) the inertia of the motor is doubled by coupling it
with another motor of higher inertia and (ii) an additional
resistance is inserted in the stator circuit. Figs. 4(c) and
4(d) illustrate the effects of changes in inertia and arma-
ture resistance on speeds, respectively. From these figures
it is seen that although there is a negligible overshooting
commutation in the case of change in inertia but the ANN based PMSM
system accurately follows the reference speed for both the
currents
cases of parameter changes. The on-line weights and biases

-i=”
PC-AT
TMS320C31 I <
up-dating feature of the A N N makes the control scheme ro-
bust against these parameter variat)ions.
It is to be noted here that t,here was a possibility of er-
ror accumulation due to the use of a delayed sample of
q-axis reference current in the input sequence of the ANN.
DSP-DS1102 This problem has been overcome by using a very fast type
DSP TMS320C31. Moreover, the on-line weights and bi-
ases were updated at each iteration to minimize the error.
Fig. 3. Laboratory impleinentation of ANN based PMSM drive
There was no biiilt,-in fault protection circuit in the lab-
 317

or at or y Inverpowes transistorized volt age source inverter

used for the experiment. The supply voltage was applied
through a variac for safe operation. This experimental lim-
itation of applying dc bus voltage affects the transient syn-
chronizing time of the proposed PMSM drive system which
in turn causes delay in the starting response of the mo-
tor. During experiment it was found few times that direct
switching of the rated dc bus voltage causes spikes of sig-
nificant time duration which may be longer than the safe
turn-on time of the transistors used in the laboratory in-
verters.
VII. CONCLUSION
2v 1.06 HOLD
In this work, a vector control scheme of an on-line self-
tuning ANN based interior type PMSM has been developed
and experimentally implemented in the laboratory. A sys-
tematic mathematical formulation is presented to develop
the inverse dynamic model of the PMSM drive system with
the vector control scheme. The structure of the proposed
ANN is derived from this inverse dynamic model of the
PMSM. The scheme has successfully identified non-linear
operating characteristics of the load applied to the motor.
In order to achieve the adaptive control over a wide operat-
ing range, an on-line weights and biases up-dating method
has been incorporated with the control system. This offers
a unique feature of adaptive properties of the system par-
ticularly in case of load disturbances and parameter vari-
ations. The laboratory test results validate the efficacy of
the ANN in the permanent magnet ac motor control.

APPENDIX
A: INVERSEDYNAMIC
CONSTANTS

(Y = [2L,J + AT(RJ + L,B + ICzL,) - AT2(RB+ KTICb

+J{zLq)]/D; P = -Lq J / D ; = -[ATKl(L, + RAT)]/D;
6 = [ATIC1Lq]/D;E = [ICTAT~]/D; 6 = -[K3RATZ]/D.
where, D = L, J + AT(RJ + L , B + ICzL,)
B: MOTORDATA
APPENDIX

1 hp, 3-4, 208 VI 60Hz, Interior type PM motor.

Number of poles = 4; L , = 0.0795 H; L d = 0.0424 H;
R=1.5 R; J= 0.003 Kg.m2; B=0.00008 (N-m)/rad/sec;
= 0.314 volts/rad/sec.

REFERENCES
G.R. Slemon, “Electric machines and drives”, Addison-Wesley
publishing, Co., Inc., 1992
F. Blaschke, “The principle of field orientation as applied to the
new transvector closed-Loop control system for rotating-field
machines”, Simens Review, 1972.
M.A. El-Sharkawi, A.A. El-Samahy, and M.L. El-Sayed, “High
performance drive of dc brushless motors using neural network”,
IEEE Trans. on Energy Conversion, Vol. 9, No. 2, 1994, pp.
Fig. 4. Experiniental speed responses of the ANN based PMSM : Y 317-322.
141 H. Naitoh and S. Tadakuma, “Microprocessorbased adjustable
axis: 750 rpm/div; (a) with change in reference speed; (b) with step speed dc motor drives using model reference adaptive control”,
change in load; (c) with change in inertia, J=2J; (d) with change in IEEE Trans. on Industry Applications, Vol. 23, No. 2, 1987, pp.
stator resistance, R=2R. 313-318.
 318

[5] Y. Dote and R.G. Hoft, “Microprocessor based sliding mode M. Azizur Rahman (S’67-M’6&SM173-F’88)
controller for dc motor drives”, Conference Record, IEEE/IAS was born in Santahar, Bangladesh on January
Annual Meeting 1980, pp. 641-645. 9, 1941. He received B.Sc.Eng. (electrical) de-
[6] M.A. El-Sharkawi and C.H. Huang, “Variable structure tradcing gree from the Bangladesh University of Engi-
of dc motor for high performanceapplications”, IEEE Trans. on neering and Technology (BUET), Dhaka, the
Energy Conversion, Vol. 4, No. 4, 1989, pp. 643-650. M.A.Sc. from the University of Toronto, and
[7] A..Brickwedde, “Microprocessor-basedadaptive speed and p* the Ph.D. degree from Carleton University, Ot-
sition control for electric drives”, IEEE Trans. on I n d u s i q Ap- tawa, Canada, in 1962, 1965 and 1968, respec-
plication, Vol. IA-21, No. 5., 1985, pp. 1154-1161. tively.
[SI M.A. El-Sharkawi and S. Weerasooriya, “Development and im- In 1962, he joined the Department of Electrical Engineering, BUET,
plementation of self-tuning tracking controller for dc motors”, as a Lecturer. He became an Assistant Professor in 1969, Associate
IEEE Trans. on Energy Conversion, Vol. 5, No. 1, 1990, pp. Professor in 1972, and Professor in 1975. From 1963 to 1964, he was a
122-128. Resident Consultant with Dhaka Electric Supply, Dhaka, Bangladesh.
[9] S. Weerasooriya, and M. El-Sharkawi, “Identification and con- In 1968, he joined the Canadian General Electric Co., Peterborough,
trol of a dc motor using back-propagation neural Iletworks”, as a Research Scientist. He was a Visiting Research Fellow at the
IEEE Trans. on Energy Conversion, Vol. 6, No. 4, 1991, pp. Technische Hogeschool Eindhoven, the Netherlands in 1973 and 1975,
663-669. a Nufiield Fellow at the Imperial College, London, U.K., from 1974 to
1975, a Research Fellow at the University of Manitoba in 1975, and a
[IO] T. F’ukuda and T. Shibata, “Theory and applications of neu- Visiting Fellow at the University of Toronto in 1975 and 1984 - 1985.
ral networks for industrial control systems”, IEEE Trans. on From 1975 to 1976, he was with Teshmont Consultants, Winnipeg.
Industrial Electronics, Vol. 39, No. 6, 1992, pp. 472-489. In September 1976, he joined the Memorial University of Newfound-
[ll] K.S. Narendra and K. Parthasarathy ,“Identification and con- land, St. John’s, Canada, where he is a University Research Pro-
trol of dynamical systems using neural networks”, IEEE Trans. fessor. From 1978 to 1979, he was a Resident Consulting Engineer
on Neural Neiworks, Vol. 1, No. 1, 1990, pp. 1-27. with the General Electric Company, Schenectady, NY. In 1977, 1980-
[I21 S. Weerasooriya and M.A. El-Sharkawi, “Laboratory implemen- 1981, and 1988, he was a Resident Consultant with Newfoundland
tation of a neural network trajectory controller for a dc motor”, and Labrador Hydro. He was a Visiting Professor at the Nanyang
IEEE Trans. On Energy Conversion, Vol. 8, No. 1, 1993, pp. Technological University, Singapore, from 1991 to 1992, and Centen-
107-113. nial Visiting Professor at the Tokyo Institute of Technology, Japan
[13] F.M. El-khouly, A S . Abdel-Gaffer, A.A. Mohammed and A.M. in 1992. His current interests include machines, power electronics,
Shard, “Artificialintelligent speed control strategies for perma- power systems, and digital protection. He has published more than
nent magnet dc motor drives”, Conference Record, IEEE/IAS 365 papers.
Annual Meeting, Vol. 1, 1994, pp. 379-384. Dr. Rahman was a recipient of the IEEE Outstanding Students
[14] M.A. Hoque, M.R. Zaman and M.A. -an, “Artificialneural Counsellor’s Award in 1980, the Notable Service Award for contri-
network based permanent magnet dc motor drives”, Conference butions to the IEEE and engineering professions in 1987, the IEEE
Record, IEEE/IAS Annual Meeting, Vol. l., 1995, pp. 9S103. Industry Applications Society Outstanding Achievement Award in
[15] M.A. Hoque, M.R. Zaman and M.A. Rahman, “Artificialneu- 1992, the Association of Professional Engineers and Geoscientist of
ral network based controller for permanent magnet dc motor Newfoundland Merit Award in 1994, and the IEEE Canada Out-
drives”, Conference Record, IEEE/IAS Annual Meeiing, Vol. standing Engineering Educator’s Medal in 1996. He is a Registered
2., 1995, pp. 1775-1780. ProfessionalEngineer in the Provinces of Newfoundlandand Ontario,
[16] C. Shiguo, D.G. Holmes and W.A. Brown, “Digital control Canada, a member of the Institute of Electrical Engineers, Japan, a
of a servo system using neurd networks”, Conference Record, Fellow of the Institution of Electrical Engineers, U.K., and a Life
IEEE/IAS Annual Meeiing, Vol. 1, 1995, pp.129-133.
Fellow of the Institution of Engineers, Bangladesh.
[17] J. Theocharis and V. Petridis, “Neural network observer for in-
duction motor Control”, IEEE Control System Magazine, 1994,
pp. 26-37.
[18] B. Burton, F. Kamran, R.G. Harley, G.T. Habetter, M. Broike,
and R. Poddar, “Identificationand control of induction Motor M. Ashraful Hoque was born in Rangpur,
stator currents using fast on-line random training of a neural Bangladesh, in 1962. He received his B.Sc.
network” Conference Record, IEEE/IAS Annual Meeting, Vol. degree in Electrical and Electronic Engineer-
2, 1995, pp.1781-1787. ing from the Bangladesh University of Engi-
[I91 L. Ben-Brahim, T. Kudor, K. Shimane and H. Naitoh, “Imple- neering and Technology, Dhaka in 1986. He
mentation of an induction motor speed estimator using neural obtained his M.Engg. and Ph.D. degrees from
network”, Conference Record, Ini. Power Electronics Confer- the Memorial University of Newfoundland,Canad
ence, IPEC, Yokohama, 1995, Vol. 1, pp. 52-57. ’e
in 1993 and 1996, respectively. After his grad-
[20] P.C. Krause, “Analysis of electric machinery”, Mc-Graw-Hill uation, he joined the Department of Electrical
Inc., 1986 I

and Electronic Engineering, Bangladesh Insti-

[21] H. Haykin, “Neural networks : A comprehensive foundation”, tute of Technology, Rajshahi, where he is now
IEEE Press, Macmillan college publishing company, Inc., 1994. an Assistant Professor.
[22] H. Demuth and M. Beale, “Neural network control tool-box Dr. Hoque has published 17 papers so far. His research interests
user’s guide”, The Math Works Inc., 1993.
are electric motor drives, power electronics and application of neural
networks to power apparatus.