NEURAL NETWORK BASED ESTIMATION

OF FEEDBACK SIGNALS FOR A VECTOR

CONTROLLED INDUCTION MOTOR DRIVE

M. Godoy Shn8es* and Bimal K.Bose

Department of Electrical Engineering

The University of Tennessee
Knoxville, TN 37996

-
Abstract Neural network is recently showing a good promise of converters and drives, but its application in estimation, particularly
for application in power electronics and motion control systems. with time-varying input signals, is practically new. The present work
So far, it has been applied for a few cases, mainly in the control explores feedforward neural network technique for estimation of
of converters and drives, but its application in estimation is feedback signals of a direct vector-controlled (DVC) induction motor
pratically new. This paper explores the application of neural drive.
network for estimation of feedback signals in induction motor Fig. 1 shows the block diagram of a DVC drive system where
drive system. A feedforward neural network receives the the estimation of feedback signals, such as rotor flux ( ~ 3unit ,
machine terminal signals at the input and calculates flux, torque vectors ( ~ 0 ,~sinee)
9 ~ and torque (Te) are indicated by a DSP as
and unit vectors (cos@, and sine,) at the output which are then well as a neural network. The estimated torque can be used in an
used in the control of a direct vector-controlled drive system. additional feedback loop within the speed control loop, if desired. In
The threelayer network bas been trained extensively by a DVC drive system, the feedback signals are calculated fiom the
NeuralWorks Professional ryPlus program to emulate the DSP- machine terminal voltages and currents by using the following
based computational characteristics. The performance of the equations 131:
estimator is good and is comparable to that of DSP-based
estimation. The drive system has been operated in the wide
torque and speed regions independently with DSP-based
estimator and neural network-based estimator, and are shown to
have comparable performance. The neural network estimator
has the advantages of faster execution speed, harmonic ripple
immunity and fault tolerance characteristics compared to DSP-
based estimator.

L INTRODUCTION

Adjustable speed vector-controlled induction motor drives have

found wide popularity for high performance motion control
applications. Since the control and feedback signal processing for the
drive are very complex, a powerful microcomputer or digital signal
processor @SP) is invariably used for the computation. Recently,
artificial neural network (ANN) or neural network technique is (7)
showing promise as a competetive method of signal processing for
such applications. Neural network has the advantages of extremely
fast parallel computation, immunity tiom input harmonic ripple, and
fault tolerance characteristics due to distributed network intelligence.
It can perform dedicated signal processing function and can easily vl',
cose, = - (9)
co-exist as a peripheral chip with a system control DSP. Y,
A neural network consists of many neurons or processing
elements interconnected together to constitute a parallel neuro-
computing network. Each neuron, again, can be modelled as an op-
amp summer-like configuration where the output is transmitted and
squashed (or limited) through a non-linear transfer function. So far,
neural network has been applied for a few cases mainly in the control
*Prof. S i m k is currently in leave of absence from University of sa0 paulo, where: vdsS ( vqss ) = stator voltage in d-axis (q-axis),
Brazil and is being supported by National Council for Scientific and i& (iqsS ) = stator current in d-axis (q-axis),
TechnologicalDevelopment (C"@
0-7803-1993-1/94 $4.00 0 1994 IEEE 471
 AC ) INDUCTION
MOTOR
t
t voltages
t currents

1
x1
t I I I
I t
1 1
&
r!3............ 2,........... !k...........
I
J I
I
normalization^
I 1

DSP NETWORK
j ==TOR / ESTlMAToR
............ "............................... .

Fig. 1. Block diagram of direct vector controlled induction motor drive showing DSP based and neural network based
feedback signal estimation.

\ v b S (wqSs) = stator flux linkage in d-axis (q-axis),

II. NEURAL NETWORK ESTIMATION PRINCIPLE
w h s (wqmS) = airgap flux linkage in d-axis (q-axis),
\vbs(\vqrS) = rotorflux linkage in d-axis (q-axis), The structure of the proposed feedforward neural network used
R, = stator resistance, for estimation is indicated in Fig. 2. The network has three layers,
Lis = stator leakage inductance, i.e., input layer, hidden layer and the output layer. The circles in the
Llr = rotor leakage inductance network represent the neurons. The input and output layers have
L,, = magnetizing inductance, neurons equal to the respective number of signals, whereas the
L, = rotor inductance and hidden layer in the present design has 20 neurons. The topology can
P = number ofpoles be defined as four-twenty-four network.The network is fully
connected, i.e.,the output of each neuron is connected to all the
neurons in the forward layer through a weight which is not shown in
the figure (see Fig. 4). Besides, a bias signal is coupled to all the
All the signals with superscript s indicate that they are in neurons of the hidden and output layers (not shown) though a
stationary reference frame. The integrations in (1) and ( 2 ) can be weight.
merged with the low-pass filter (LPF), as indicated in Fig. 1, by using
very low comer frequency, and can easily be implemented by The input layer neurons have linear transfer characteristics, as
dedicated hardware (op-amps). The reason for merging the indicated. On the other hand, the hidden layer and output layer
integration function with the LPF hardware is that the feedforward neurons have hyperbolic-tan type nonlinear transfer function given as
neural network, used in the present application, is basically a pattern
recognition network and can not perform any dynamic function. In
Fig. 1, both the DSP and neural network receive the variable
frequency variable magnitude signal waves &, w',,:i and , and
then compute (3) - (10) to estimate rotor flux, unit vectors and
torque, as shown. The DSP-based estimator output is used for where a is the gain. The transfer function is bipolar, monotonic,
comparison of the neural network based estimator performance. differentiable, and has the largest gain (a) at zero signal. Fig. 3
shows the plot of the transfer function with adjustabie slope, and Fig.
4 shows the structure of the neural element that incorporates the
transfer function.
472
 I T I T -
TrCOSC ---- -
= % ? , =

Fig.2. Topology of the neural network for estimation

)--*- _',,.'
_..
,,.:::./,"~
/."::* ............
__..
,.....
_.......
__...
slope
adjustment by a
The weight w in series with the transfer function 0 helps
adjusting its slope, as indicated in Fig. 3. The nonlinear transfer
function associated with the neurons gives nonlinear mapping
property of the network and helps performing highly nonlinear
computations, such as multiplication, division and square-rooting,
besides addition and substraction, as indicated in (3) - (10). The input

/:::.....
....,. ....
_....I..I I:::
_..' _.."
...I

_,_..._/
/''_
J.'
...,..,.

-1
1 variable frequency variable magnitude near-sinusoidal signals are
converted to per-unit form through the normalizer gains, and then
&er computation, the output are brought back to actual values
through the denormalizer gains, as indicated.
The feedforward neural network is usually trained by back-
propagation training algorithm first proposed by Rumelhart, Hinton
Fig. 3. Hyperbohotan transfer function with adjustable slope and Williams in 1986. The distributed weights in the network
contribute to the distributed intelligence or "associative memory"
property of the network. With the network initially untrained, i.e.,
with the weights selected at random, the output signal pattern will
totally mismatch the desired output pattern for a given input pattern.
Xl The: actual output pattern is compared with the desired output
\
patt em and the weights are adjusted by the back-propagation
algorithm until the pattern matching occurs, i.e., the pattern errors
beoome acceptably small. The network training is highly automated
and is usually performed off-line through a PC-based Simulator
pro,yam, such as Neuralworks Professional IllpLUS [6]. The
pariicular Simulator program has a graphical interface where all the
hWtI set-ups are done via dialog boxes. It is possible to add some visual
instruments like rms error gauge, weight histogram and cofision
matrices for helping the user to see whether the training is
converging. The Simulator can use ASCII, binary or commented
ASCII files for training. The network topology can be edited to
change the connections, delete weights or to implement different
transfer functions.

Fig. 4.Structure of neural element with the transfer function The training procedure used in the present project can be
summarized as follows:

473
 25
Simulate the induction motor drive system, as shown in Fig.
1, by PC-SII"0N. 17
Generate the input/output data table for different operating h

conditions with the help of PC-MATLAB .The input data E 8

z
v
correspond to the machine terminal variables $&,$ ,, & and W O
j ; , and the output data correspond to the desired estimated
3
#
tr:
signals derived by solving (3) - (10). o
F
-8
Convert the data table in per unit form and feed to the
-1 7
Neuralworks Simulator program located in PC.
The Simulator assigns small but random weights initially to
the network.
-25t
0.0
'
0.2
' "
0.4
"
0.6
"
0.8
"
1.o
An inputloutput data pair is selected from the table. For the TIME (s)
given input data pattern, calculate the network output and
compare with the desired data output to derive the error Fig. 5. Torque estimator performance with four-five-one network
pattern.
From the error pattern, compute and adjust the network , , I

weights by the back-propagation algorithm so that the new 17

error is small.
Repeat the above steps with each set of inputloutput data 8
patterns until the rms error for the entire training set
0
converges below the desired threshold value.
After completion of the training, test the network -8
performance with arbitrary input pattern to ensure successhl
training. -1 7

The design and training of a neural network for satisfactory 0.0 0.2 0.4 0.6 0.8 1.o
performance requires very time consuming iterative procedure with TIME (s)
large training data table. Fortunately, the Simulator program highly
automates the training procedure. The selection of hidden layer
neurons may require several stages of iteration. If the number is Fig. 6. Torque estimator performance with four-ten-one network
small, the error will not converge to the satisfactory level. Again, if
the number is too large, the network will tend to memorize (look-up Being encouraged with the feasibility of estimation and
table function) rather than learn. After satisfactory training with the performance improvement with larger number of hidden layer
help of Simulator program, the weights are down-loaded to the neurons, it was decided to design the &U-fledged estimator with
prototype network. four-twenty-four network, as shown in Fig. 2. As mentioned before,
there is no unique way to determine the optimum number of hidden
layer neurons. To some extent, it involves trial-and-error iteration
III. ESTIMATOR PERFORMANCE procedure. Figs. 7(a) - 7(d) show, respectively, the torque, flux, cos0
e and sine, output of the estimator &er success&l training of the
In the beginning, it was decided to estimate the torque (Td only network with a very large number of data sets. As before, the drive
with a simple four-five-one network to validate the feasibility of system operates with variable frequency but with a constant PWh4
estimation. A 5 hp DVC induction motor drive system, shown in Fig. carrier frequency of 15 kHz. Therefore, the input idsS
1, with pure inertia load was simulated with the flux and speed current waves are reasonably harmonic-free. The ripple gene2tek;
control loops open but with a constant ids* (i.e., constant flux). A the network is evident on torque and flux, but the unit vector signals
triangular bi-directional iqs* profile was injected to generate the are clean. Again, the estimator ripple frequency has some
proportional torque profile, as indicated in Fig. 5. As a result, the correspondence with the motor fundamental frequency. For example,
machine speed varies correspondingly generating variable frequency at time below 0.2 s and above 0.8 s in Fig. 7, the fundamental
variable magnitude $*, $ ,, &, and ib, waves. frequency is very low showing the corresponding low ripple
frequency on the estimated torque. The overall performance of the
An inputloutput data table (with 10,000 points) was used to
neural network based estimator appears to be very encouraging. Fig.
8 shows the estimator performance for the same conditions as above
train the network. The network performance, although somewhat
crude in comparison with the actual torque, indicates the correct except the switching frequency is low (2 kHz). At low switching
trend for estimation. The rms threshold error was reasonably low frequency, the input idsS and iqsS current waves have larger harmonic
inspite of large output ripple which was found to be present even at ripple, but the estimator output is relatively immune to this input
constant torque condition ( see torque step in Fig. lO(a)). The ripple ripple.

amplitude can be attenuated to some extent with a low pass filter of This noise-immune performance of neural network is
small time constant. Fig. 6 shows improvement of the estimato. particularly important for drive system feedback signals estimation
performance when the hidden layer neurons were increased from fi. with low switching freauencv. such as GTO inverters. Otherwise, any
v .

to ten. attempt to reduce the riiple by a low pass filter, will cause
474
 20 20
I I
I
I
I
15 15
10 10
n h
E E
z
v 5 v2 5
W O U 0
3
8 -5 s
E -5
-1 0 E -1 0
-1 5 -15
-201
0.0
"
0.2
"
0.4
"
0.6
"
0.8
' I
1.o
-201
0.0
' ' '
0.2
' ' ' '
0.4
' ' ' '
0.6
' ' ' '
0.8
1 ' ' ' I
1 .o
a) a)
0.50 t 1 0.50 I

, I I
0.08
I I
I I I I
0.0 0.2 0.4 0.6 0.8 1 .o
b)
1.1 1.1

0.7 0.7

w 0.2 w 0.2
B
m
8 -0.2
w8 -0.2

-0.7 -0.7

-1.1 -1.1
0.0 0.2 0.4 0.6 0.8 1 .o 0.0 0.2 Q.4 0.6 0.8 1 .o
c) c>
1.1 r 1.1 I I

0.7 0.7

0.2 0.2
w wi
25
m S'
U3
-0.2 -0.2

-0.7 -0.7

-1.1 -1.1
0.0 0.2 0.4 0.6 0.8 1.o 0.0 0.2 0.4 0.6 0.8 1.o
TIME (s) TIME (s)
d) 4
hndamental frequency sensitive phase lag which may not ' -
acceptable. Fig. 9 shows the validity of flux estimation in the IV.SYSTEM SIMULATION STUDY AND
constant torque as well as in field-weakening regions. PERFORMANCE COMPARISON

Once the neural network based estimator performance was

0.50 I I I I
found to be satisfactory, it was decided to operate the drive system
of Fig. 2 with closed loop flux and speed control independently with
DSP-based estimator and neural network based estimator to compare
the system performances. The inverter switching frequency was
maintained at 15 WIZ. Two sets of tests were carried out: (a)
Operate the system at constant speed but at stepped load torque and
(b) then at free acceleratioddecelerationmode with pure inertia load.
The performances at both the conditions with the neural network
based estimator were found to be excellent. Fig. 10 shows the
performance comparison with stepped load torque with the set speed
of 1195 rpm where the left column indicates the DSP-based
performance whereas the right side indicates the neural network
based performance. The performances match very well, but the ripple
0.0 0.5 1 .o 1.5 2.0 2.5 with neural network based estimator was somewhat large. Evidently,
the torque and flux ripples have little effect on the drive because of
TIME (s) inertia filtering. Fig. 1 1 shows the performance comparison of all the
signals when the speed was cycled with a bi-directional profile in the
constant torque mode. The performances were found to match very
Fig. 9. Flux estimation in constant torque and field-weakening well at all the speeds. The drive also performs well in the field-
regions weakening range, although no results are shown.

NEURAL NETWORK

16
15

I
i
n
14 14
2 13
v
!

2
v
13
g 12 w 12

t
3
2
0
6
11
10
__-A11
i
I i z
41 10
9 9

0.0
I
0.5
I
1 .o 1.5
8 r "
0.0
' ,,
' '
I
0.5
I
1 .o
I
1.5

0.50 0.50 1 I
0.43 0.43
0.36
h
D
t 0.29
v

X
0.21
cr,
0.14

0.07
0.00 I I I I

b)
Fig.10. Performance comparison of vector controlled drive system for stepped load torque and constant
speed (1 195 rpm) (a) Torque, (b) Flux, (c) Cosine wave and (d) Sine wave (Left column indicates DSP
based estimation and the right column indicates neural network based estimation)
476
 DSP NEURAL NETWORK

1.1

0.6

W
z
E; 0.0
8
-0.6

-1.1 ...
1.0 1.0s 1.1 1.o 1.05 1.1

1.1

0.6

w
5 0.0
v1

-0.6

-1.1
...
1.o 1.os 1.1 1.o 1.os 1.1

TIME (s) TIME (s)

4
Fig.10. Performance comparison of vector controlled drive system for stepped load toque and constant
speed (1 195 rpm) (a) Torque, (b) Flux, (c) Cosine wave and (d) Sinewave (Ldt column Se DSP
based estimation and the right column indicates neural network based estimation)

DSP NEURAL NETWORK

w
W
& -500 - \
I

Fig.11. Performance comparison of vector controlled drive system with q d ~ dspeed profile @ure
inertia load) (a) Speed (b) Torque, (c) Flux, (d) Cosine wave and (e) Sic wave ( I column
, & indicates
DSP based estimation and the right column indicatesn d network based esthtim)
477
 NEURAL NETWORK

-15 I- I

I
-20 1 1 I i J
0.0 0.5 1.o 1.5 2.0
b)
0.50 I 0.50 I

0.42 --
I
A---
I
- 0.42

-g 0.33 -
I
I
I
i
-
n 0.33
W
0.25 -__-___F-- 4
--- I

I
W
%:
0.25
3
E2 0.11
1
+ 2 0.17
I
I I
0.08
1 I
0.08
i
0.00
1
I
1
0.00
0 0.0 0.5 1.o 1.5 2.0
c)
1.1 I 1.1

0.6 0.6

w
3
U
0.0
w
z
tj 0.0
8
-0.6 -0.6

-1.1 ' I 1 I I -1.1

0.0 0.5 1.o 1.5 2.0 0.0 0.5 1 .o 1.5 2.0

1.1 , , ...
1.1 .
0.6 0.6

-wz
v)
0.0
w
rA
0.0

-0.6 -0.6

-1.1 -1.1
0.0 0.5 1.o 1.5 2.0 0.0 0.5 1 .o 1.5 2.0
TIME (s) e) TIME (s)
Fig.11. Performance comparison of vector controlled drive system with cycled speed protile (pure
inertia load) (a) Speed @) Torque, (c) Flux, (d) Cosine wave and (e) Sine wave (Left column indicates
DSP based estimation and the right column indicates neural network based estimation)
478
 V. CONCLUSION

The paper successllly demonstrates the application of neural

network in the estimation of feedback signals for a vector-controlled
induction motor drive system. A three-layer feedforward neural
network of the structure four-twenty-four has been trained with the
Neural-Works Professional IYPLUS Simulator program using the
variable hquency variable magnitude input/output data fiom a
simulated drive system, and the performance of the estimator was
found to be excellent at all operating conditions. Although the
estimator performance was demonstrated for a direct vector-
controlled induction motor drive , it can be extended to scalar or
vector control ( direct or indirect) of any type of drive system. The
estimator can be implemented by dedicated parallel hardware chip,
such as Intel 80 170- ETA" (electrically trainable d o g neural
network) or by DSP-based software. The dedicated hardware
estimator chip can easily CO-existwith the system control DSP and
relieve its computational burden. The neural network has distinct
advantages when compared to the conventional DSP-based estimator
and promises to be the fbture choice for application in industrial
drives.

VL REFERENCES
[I] J. Lawrence, Introduction to Neural Networks, California
Scientific Software Press, 1993.
[2] E.S. Sinencio and C. Lau (Ed.), Artificial Neural Networks,
IEEE Press, NY,1992.
[3] B.K. Bose, Power Electronics and AC Drives, Prentice Hall,
1986.
[4]B.K. Bose, "Expert system, I k z y logic and neural network
applications in power electronics and motion control", Proceedings
of the IEEE, August 1994 (to be published).
[SI D.H. Nguyen and B. Widrow, "Neural networks for self-leaming
control systems", IEEE Control Sys. Mag. vol. 10, pp.18-23, April
1990.
[6] "Using Neuralworks Professional 11-Plus", Neuralware
Reference Manual, 1992
[7] S. Weerasooriya and M.A. El-Sharkawi. "Identification and
control of a dc motor using back-propagation neural networks",
IEEE Trans. Energy Conversion, vol. 6,pp. 663-669,Dec. 1991.

479