IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 51, NO.

5, OCTOBER 2004

981

Neural-Network-Based Waveform Processing and

Delayless Filtering in Power Electronics
and AC Drives
Jin Zhao and Bimal K. Bose, Life Fellow, IEEE

AbstractThis paper systematically explores the static nonlinear mapping property of feedforward neural networks for
various waveform processing and delayless filtering that are
applicable to power electronics and ac drives area. Neural-network-based processing of waves gives considerable simplification
of hardware and/or software that are traditionally used for such
applications. Two general cases have been investigated: The
voltage or current waveforms which have constant frequency but
variable magnitudes, and the other case is variable-frequency
variable-magnitude voltage or current waves. The former case is
mainly important for power electronics that operate on a utility
system and general-purpose constant-frequency converter power
supplies, and the latter is important for the adjustable-speed
ac drives area. In both cases, the performance of neural-network-based waveform processing and delayless filtering with
offline training was found to be excellent. The results of this study
are also applicable to other areas of electrical engineering.
Index TermsAC drives, delayless filtering, neural network,
power electronics, waveform processing.

I. INTRODUCTION

OWER electronics and variable-frequency drive systems

often deal with complex voltage and current waves that are
rich in harmonics. These waveforms often require complex processing for control, monitoring, diagnostics, and protection of
the system. Normally, analog/digital hardware, software, or a
combination of both is required for processing these waves. One
of the important processing functions is predictive or delayless
filtering in order to retrieve the fundamental (sine wave) component of the wave. For example, a diode or thyristor phase-controlled bridge converter, operating on a 60-Hz utility line, can
generate square or six-stepped line current wave, and this waveform becomes multistepped (more than six steps) with multiple
phase-shifted bridge converters on a three-phase line [1]. Similar waveforms are also generated, respectively, in the output
voltage of a square-wave voltage-fed inverter with single bridge
or phase-shifted multibridge configuration. The harmonic-rich
line current and output voltage waves can again cause distortion in the line voltage and load current waves, respectively.
Manuscript received January 12, 2004; revised June 22, 2004. Abstract published on the Internet July 15, 2004.
J. Zhao was with the Department of Electrical Engineering, The University
of Tennessee, Knoxville, TN 37996-2100 USA, on leave from the Department
of Automatic Control Engineering, Huazhong University of Science and Technology, Wuhan 430074, China (e-mail: zhao2000617@yahoo.com.cn).
B. K. Bose is with the Department of Electrical Engineering, The University
of Tennessee, Knoxville, TN 37996-2100 USA (e-mail: bbose@utk.edu).
Digital Object Identifier 10.1109/TIE.2004.834949

It is often necessary to retrieve the fundamental component of

these waves in order to calculate, for example, the displacement
power factor (DPF), fundamental frequency active (P) and reactive power (Q), and energy measured by a kilowatthour meter. In
photovoltaic and wind generation systems coupled to the grid,
the distorted line voltage (due to converter harmonics) waves
require delayless filtering in order to generate inverter sine reference voltage waves for controlling the line DPF to unity [2],
[3]. The distorted line voltage waves also create problems in
the comparator (or zero-crossing detector) which is often essential for control of the converter (e.g., cosine-wave-crossing
control of a phase-controlled converter). Generally, an activeor passive-type low-pass filter (LPF) with narrow bandwidth is
used to filter out the harmonic components. However, an LPF
causes phase lag and amplitude attenuation that vary with fundamental frequency. For a utility system, the fundamental frequency is essentially constant and, therefore, these phase and
amplitude errors can be compensated without much difficulty
[2]. However, for variable-frequency drive applications, the inverter usually operates in pulsewidth-modulation (PWM) mode
with wide frequency variation generating machine voltage and
current waves that are complex with harmonics. If a simple LPF
with narrow bandwidth is used in these applications, the variable phase delay and amplitude attenuation for the fundamental
may not be acceptable, particularly at higher fundamental frequency. The phase error is particularly harmful in a vector-controlled drive where it creates the coupling problem and, thus,
deteriorates the drive performance. In the past, complex digital adaptive filters, such a finite-impulse response (FIR), infinite-impulse response (IIR), or a combination of both have been
proposed [3][5] to obtain delayless filtering of the fundamental
component.
In this paper, we propose the neural network solution for
waveform processing and delayless filtering problems. The artificial neural network (ANN), or neural network, a generic form
of artificial intelligence (AI), is recently offering a new frontier
in solving many control, estimation, and diagnostic problems in
power electronics and motor drives. Between the two classes of
ANN, i.e., the feedforward and feedback or recurrent types, the
former provides static nonlinear inputoutput mapping or pattern recognition property with precision interpolation capability.
With appropriate training, this property permits a feedforward
ANN to recognize a waveshape and retrieve the desired component of the wave. Since the shape or pattern of the wave remains
constant or goes through deterministic variation, simple offline
training of the network has been used in the project. The advan-

0278-0046/04$20.00 2004 IEEE

982

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 51, NO. 5, OCTOBER 2004

j j = 0 9 pu, 0.2 pu. MSE = 1 9997e 0 005). (a) Input waves. (b) Network with I/O signals.

Fig. 1. Square-wave delayless filtering by neural network ( v

(c) Output waves.

tages of ANN processing of a wave compared to that of a digital

signal processor (DSP) are obvious: it is simple and fast with a
dedicated ASIC chip due to parallel processing, and the ANN
has the properties of noise immunity and fault tolerance where
the former is particularly important in the distorted waveform
processing application. This noise immunity property also remains valid if the ANN operation is emulated by DSP. It should
be mentioned here that application of an ANN for waveform
processing and delayless filtering is not entirely new. The ANN,
as well as fuzzy logic, have been used for estimation of rms
and fundamental rms values, DPF, and PF of distorted 60-Hz
line current waves [6], [7]. It has also been used successfully
in a variable-frequency vector-controlled drive [8] for feedback
, unit vector (
,
),
signal estimation of rotor flux
from the machine voltage and current wave sigand torque
nals. Recently, the ANN has also been applied in delayless filtering to generate a 60-Hz reference current wave for active filters [9] and zero-crossing detection of distorted line voltage [10]
and current waves in a phase-controlled cycloconverter [11].
In this paper, ANN-based waveform processing and delayless
filtering have been studied systematically for two general cases.
First, the voltage or current waves that have constant frequency
but variable magnitudes have been studied. Next, the variablefrequency variable-magnitude voltage or current waves that are
relevant to an ac drive system are investigated. The performance
of the ANN in all these cases was found to be excellent.
II. CONSTANT-FREQUENCY VARIABLE-MAGNITUDE WAVES
The study in this section includes constant-frequency variable-magnitude single-phase, two-phase, and three-phase
waves. Both the square and six-step waves have been studied
thoroughly, but only the square-wave cases will be discussed.
A. Single-Phase Square Wave
In the beginning, let us consider a simple constant frequency
(60 Hz) square wave with variable magnitude, and the problem
here is retrieving its fundamental component without any phase
delay (delayless) with the help of the ANN. The square or sixstep waveform is encountered in utility system line current for a

diode or thyristor rectifier with highly inductive load, or output

voltage wave of a square-wave voltage-fed inverter (single or
three phase) in a general-purpose 60Hz power supply or uninterruptible power supply (UPS) system. A simple square-wave
signal may also be processed to generate in-phase sinusoidal
command voltage wave for a single-phase PWM inverter. Fig. 1
shows the ANN-based generation of in-phase fundamental component where the output magnitude varies linearly with the magnitude of input square-wave . Since the square-wave amplitude is constant in the half-cycle, the ANN cannot generate a
continuously variable sine wave directly from the square wave.
is generated from
For this reason, an auxiliary input wave
the square wave
through a first-order low-pass filter (LPF)
shown in the figure. The LPF time constant should be large
enough (but arbitrary) so that amplitude varies continuously
in the half-cycle. Since the ANN requires the target or desired
fundamental wave embedded in the input signal, it was retrieved
by off-line FFT analysis of the input square wave. The training
and
waves in the maginput data were generated from
nitude range of 01.0 pu with the step size of 0.1 pu and 500
data points per cycle of fundamental frequency (0.72 interval),
where
where 1.0 pu corresponds to 0.5 (actually, 0.5
V) shown in Fig. 1(a). The feedforward ANN was then
trained with the example data sets so that the estimated sine
wave is identical with the desired sine wave generated from
the square wave with the phase angle remaining locked at 0
as shown in the figure. The MATLAB-based Neural Network
Toolbox was used with LevenburgMarquardt (LM)-based fast
backpropagation algorithm for the training. There is no fixed
rule to determine the number of neurons in the hidden layer.
In the beginning of the training, a small number of neurons is
used. It is then gradually increased until satisfactory training
with the desired MSE is obtained. The network has 2, 12, and
1 neurons in the input layer, hidden layer and output layer, respectively [see Fig. 3(a)]. The training was followed by testing
to verify the ANN
cycles with intermediate magnitudes of
training performance. The trained neural network gives excellent interpolation of magnitudes and angles. Fig. 2 summarizes
the training steps of the neural network. For simplicity, only two
values of (0.9 and 0.2 pu) and the corresponding outputs are

ZHAO AND BOSE: NEURAL-NETWORK-BASED WAVEFORM PROCESSING AND DELAYLESS FILTERING

983

Fig. 2. Flowchart for training neural network using MATLAB/Simulink/

Neural Network Toolbox.

shown in Fig. 1. Note that the signals may be voltage or current

waves. Again, the input signals considered in the present study
are square or six-stepped waves. Fig. 3(a) shows the ANN configuration after training with the square-wave input. Both the
input signals and are normalized for ANN processing. The
processed output is then converted to actual signal after denormalization. The network uses bipolar linear activation function
in the input and output layers, whereas nonlinear tan-sig (or hyperbolic tan) function is used in the hidden layer. Fig. 3(b) shows
the training MSE (mean-square error) curve which indicates that
it converges to 1.9997e-005 at the end of 600 epochs which is
close to the desired goal of 2e-005. The corresponding training
time is about 1 h with a Pentium 4 (1.7 GHz)-based PC.
In the practical case, the frequency of the square wave may
drift slightly and the amplitude may have some ripple. It was
observed during testing that the slight variation of frequency
(within 0.6 Hz) and addition of commutation rise and fall
times (for current wave) in the square wave have minimal effect on the accuracy of the output. Fig. 4 compares the desired
and estimated waves after injection of a noise signal with 1%
amplitude at 2.0-kHz frequency with the square wave. With
more complexity to the network (2-15-15-1), the filtering effect was improved, as indicated in Fig. 4(b). This means that
for the same goal error of 2e-005, the MSE in Fig. 4(b) is less,

Fig. 3. (a) Neural network configuration (2-12-1) for Fig. 1. (b) Network
training MSE curve.

i.e., the estimated wave matches better with the desired wave.
In principle, one hidden layer is sufficient for universal approximation. However, in some complex applications, it is difficult
to get good performance with one hidden layer even with a large
number of neurons. Besides, the training time becomes very
long. Therefore, it is wiser to select two hidden layers in selected applications.

984

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 51, NO. 5, OCTOBER 2004

a six-step wave was also used in Figs. 1 and 5, and the correspondingly similar output waves were generated with high precision. However, with the six-step input wave, the ANN topologies were found to be slightly simpler because of the absence of
triplen harmonics.
As a generalization of the study in this section, it can be
concluded that any arbitrary input wave pattern at constant frequency can be transformed into arbitrary output wave pattern at
single or multiphase with the help of the ANN, where the output
magnitude variation follows the corresponding input variation.
The magnitude tracking can be programmed to be linear or nonlinear with any functional relation. This general waveform mapping property of an ANN can be extremely important in many
other applications in electrical engineering. From this study,
it can also be concluded that an ANN can be used as a zerocrossing detector for a distorted wave [11]. A small frequency
drift or waveshape deviation will have minimal affect on the
ANN performance. If the waveshape varies unpredictably in real
time and is unfamiliar to the ANN, the network has to be adaptive, i.e., the weights are to be updated by online training.
B. Two-Phase Square Wave
In the next phase of the study, let us consider similar waveform processing with two-phase square-wave input where the
phase shift angle is 120 . Fig. 6 shows the ANN-based delayless
filtering and polyphase waveform generation using two-phase
square-wave input ( and ). Fig. 6(a) shows the three-phase
output, and Fig. 6(b) shows the six-phase output where
is aligned with the
wave. Since the amplitudes of
and
waves are constant during the half-cycle, both of them require
filtering by a large LPF so that and vary continuously. The
phase lag and gain (attenuation) of the first-order LPF are
given, respectively, by the following equations [1]:
Fig. 4.
(v

Desired and estimated waves showing noise filtering of neural network

j j = 0 4 pu, noise = 61% amplitude at 2.0 kHz). (a) 2-12-1 ANN (MSE =
1 9997e 0 005). (b) 2-15-15-1 ANN (MSE = 1 1739e 0 005).
:

(1)
(2)

A single-phase square-wave input can be converted to balanced polyphase sine-wave output as shown in Fig. 5. Fig. 5(a)
wave remains aligned in
shows three-phase output where
phase with the input square wave. If the square-wave magnitude
,
,
varies in the range 01.0 pu, the output components
track the input signal linearly with the locked phase anand
gles. For simplicity, Fig. 5(a) is shown with
1.0 and 0.3 pu
only. The matching of the desired and estimated waves was
e
after training the
found to be excellent MSE
ANN with the inputoutput example data as discussed before
and resulting in the configuration 2-16-3. This type of waveform processing is important for three-phase 60-Hz (or any constant frequency) power supply with sinusoidal PWM inverter or
cycloconverter, where the command signal waves can be generated from a simple square-wave signal. Fig. 5(b) shows the
balanced six-phase sine-wave generation from a single-phase
square-wave input at 60 Hz with the help of the ANN. Again,
0.8 and 0.4 pu) are shown although
only two magnitudes (
the performance was found to be excellent in the whole range
(01.0 pu). This type of waveform generation is important for
multiphase sinusoidal PWM inverter. Instead of a square wave,

where is the fundamental frequency. With

s
and
corresponding to 60 Hz, the calculated values are
and
. This means that the ANN is required
to have phase and amplitude compensation in both the channels, besides filtering the harmonics of the input waves. As discussed before, the ANN output varies linearly with the squarewave magnitude. The topology of the trained ANNs and error
(MSE) are indicated in the figures. The signals and may be
two-phase (square or six-step) line currents, phase, or line voltages of a balanced three-phase system where
. The general conclusion about the ANN mapping property,
frequency, and waveform deviation made in the previous section are also valid here. An example application may be retrieving the three-phase fundamental components from distorted
two-phase voltage or current waves of a UPS system for control and feedback signals estimation. Another example is threephase synchronized sine voltage wave generation from twophase distorted line voltage waves for controlling a phase-controlled thyristor converter or cycloconverter using cosine-wavecrossing control technique.

ZHAO AND BOSE: NEURAL-NETWORK-BASED WAVEFORM PROCESSING AND DELAYLESS FILTERING

985

j j = 1.0 pu and 0.3 pu. MSE = 2 33881e 0

Fig. 5.

Single-phase square-wave delayless filtering and multiphasing by neural network. (a) Three-phase output ( v
= 0:8 pu and 0.4 pu. MSE = 2:80336e 005).

Fig. 6.

Two-phase square-wave delayless filtering and multiphasing by neural network. (a) Three-phase output ( v
= v = 0.6 pu and 0.42 pu. MSE = 7:00748e 005).

005). (b) Six-phase output (jv

4:5499e 0 005). (b) Six-phase output (jv

j j j

C. Three-Phase Square Wave

Continuing the studies further, let us now consider
three-phase square-wave inputs to the ANN and genera-

j j = j j = 0.96 pu and 0.48 pu. MSE =

v

tion of polyphase output with delayless filtering as shown in

Fig. 7. Again, square or six-stepped waves are encountered in
the line currents of three-phase diode or thyristor bridge con-

986

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 51, NO. 5, OCTOBER 2004

Fig. 7. Three-phase square-wave delayless filtering and multiphasing by neural network. (a) Three-phase output (
MSE = 1:99931e 005). (b) Six-phase output ( = v = v = 0.66 pu, 0.38 pu. MSE = 2:8732e 005).

jj j j j j

verters, or the output voltage or current waves of a three-phase

UPS system. For waveform processing and delayless filtering,
it is better to consider all the three-phase signals instead of
two-phase signals, as discussed before. In Fig. 7, all the input
s.
signals are filtered by identical LPFs
The matching of estimated output sine waves with the corresponding desired waves were excellent in the whole range
(01.0 pu) was excellent after 1000 epochs of training. The
corresponding ANN topologies and MSE are indicated in the
figures for square-wave input, although training with six-step
waves was equally successful. All the above studies were
made at steady-state condition. However, it was found that
with offline training, all the ANN output waves track very well
varies with slow
with the input waves when dc-link voltage
transient.
As a conclusion of this section, we can generalize that arbitrary -phase input signals of any wave shape can be processed
into -phase sine wave or any arbitrary waveform at the same
frequency with desired phase angle and amplitude variation
with the help of an ANN. This follows from the generalized
inputoutput static nonlinear mapping property of a feedforward neural network. A small amount of fundamental frequency
drift and noise in the input waves are permissible, as mentioned
before. Complex polyphase waveform processing is important
in multiphase rectifiers and inverters where multistepped waves
[1] are encountered. High-power multiphase diode and thyristor
converters are widely used in drives, and the latter is particulrly
important in high voltage dc (HVDC) systems. High-power
multistepped voltage-fed gate-turn-off thyristor (GTO)/integrated gate-commutated thyristor (IGCT)/insulated gate bipolar

j j = j j = j j = 0.84 pu and 0.5 pu.

v

transistor (IGBT) inverters are important, for example, in utility

system battery storage and static var generator (SVG) systems.
III. VARIABLE-FREQUENCY VARIABLE-MAGNITUDE WAVES
In this section, we will consider ANN-based waveform
processing and delayless filtering for variable-frequency variable-magnitude PWM waves which are mainly important for
variable-speed ac motor drives. These types of waveforms are
generated in voltage-fed and current-fed PWM converters.
Fixed- frequency (60 Hz) (single or three-phase) PWM waves
are also important for a general-purpose power supply or UPS
system.
A. Three-Phase PWM Voltage Waves at 60 Hz
In the beginning, let us consider three-phase PWM voltage
waves at 60 Hz only to demonstrate the effectiveness of
ANN-based waveform processing. Fig. 8(a) shows delayless filtering of three-phase PWM waves by an ANN where the
output sine waves are locked in phase with the respective input
PWM waves. The input waves, in this case, are line voltage
) generated by simulation of a sinusoidal
waves ( , , and
(with sine-triangle comparison) PWM inverter. For simplicity,
wave is shown in the figure. Since the input waves
only the
are discontinuous with constant discrete levels, these are filtered
s to convert into continuously
by LPFs
varying amplitude waves for the ANN input. For training the
ANN, 100 data points/cycle (3.6 interval) with 0.1-pu step
size were generated in the range 01.0 pu from the input waves.
Fig. 8(b) shows the ANN input waves after low-pass filtering

ZHAO AND BOSE: NEURAL-NETWORK-BASED WAVEFORM PROCESSING AND DELAYLESS FILTERING

987

Fig. 8. Three-phase PWM wave delayless filtering at 60 Hz showing harmonic filtering effect due to low carrier frequency. (a) Network showing only v input
wave. (b) ANN input and output waves at f = 10 kHz (3-5-3, MSE = 1:99796e 005). (c) ANN input and output waves at f = 1:0 kHz (3-15-3, MSE =
290519e 005).

and the corresponding desired and estimated waves at 10-kHz

switching frequency which match very closely. Here, the ANN
performs the functions of: 1) compensation of phase delay;
2) compensation for amplitude attenuation; and 3) additional
harmonic filtering. These functions were also demonstrated
with square/six-step input waves discussed before. The estimated sine waves align perfectly with the desired wave. For

simplicity, only one magnitude (1.0 pu) is shown in the figure.

The ANN structure in this case is simple (3-5-3) and the error
(MSE) converges to 1.99796e-005 after only 90 epochs. Fig. 8(c)
shows the ANN input and output waves at 1.0-kHz switching
frequency along with the desired waves using the same LPFs.
Due to larger harmonic content in the raw PWM waves, the
filtering effect of the LPF is poor, thus giving more filtering

988

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 51, NO. 5, OCTOBER 2004

Fig. 9. Three-phase PWM waves delayless filtering at variable frequency and variable voltage (f = 5 kHz, 3-23-23-3, MSE = 2:29187e 005). (a) Output
waves at 60 Hz and m = 0:96. (b) Output waves at 30 Hz and m = 0:6. (c) Output waves at 15 Hz and m = 0:4. (d) Output waves at 5 Hz and m = 0:2.

burden to the ANN. The ANN topology in this case is more

complex (3-15-3), the training iterations are larger (500 epochs)
and the MSE fails to converge below 29.0519e-005. However,
looking at the waveshapes, the harmonic filtering capability of
the ANN has been clearly demonstrated in Fig. 8(c). Using a
more complex ANN structure (3-13-13-3) and longer training
time, the harmonic content in the output of Fig. 8(c) could be
e
. In this
improved to some extent MSE
s
case, it was observed that a larger time constant
of the LPF smoothens the ANN input waves, and a simpler ANN
structure (3-10-10-3) gives improved matching of desired and
e
. However, it was also
estimated waves MSE
observed that too high a value of causes large phase delay and
high attenuation that adversely affects the ANN size. The study
that makes the input
also found that unequal
waves asymmetrical with different phase shift and attenuation,
reduces the ANN size for the same filtering performance. Instead
of line voltage waves, the phase voltage waves with respect to
,
) and isolated load neutral ( ,
dc-link center tap ( ,
,
) were also used, and the performance was found to
be equally good. In fact, any PWM algorithm can be used for
generation of these waves. The study also included three-phase
and single-phase sine-wave generation from single-phase or
two-phase (120 phase shifted) PWM waves, but these will not
be discussed.

B. Three-Phase PWM Voltage Waves at Variable Frequency

The study in the previous section will now be extended
to consider three-phase variable-frequency variable-voltage
PWM waves that are commonly encountered in variable-speed
ac drives. The PWM waves at variable frequency and variable voltage were generated by simulation of an open-loop
volts/Hz-controlled induction motor drive in the frequency
range of 560 Hz with the corresponding modulation factor
variation at fixed switching frequency
kHz. The
network configuration similar to Fig. 8(a) is also valid in this
case, but the LPF time constants are set to different values
,
, and
). The estimation
(
was also satisfactory with equal values of , but it was observed
that unequal values help ANN size reduction and reduced
training time. Fig. 9(a)(d) shows the excellent matching of the
desired and estimated waves at different frequencies (including
the 60-Hz wave) with the corresponding modulation factor
variation shown in the figure. In all the cases, training
data (1000/cycle) were generated from the ANN input waves
at 0.1-pu step size, and the ANN was trained for the desired
output waves with phase and amplitude correction. Note that
although the phase lag and magnitude attenuation are frequency
sensitive and different in each phase (because of unequal
values), these could be compensated by offline training of the

ZHAO AND BOSE: NEURAL-NETWORK-BASED WAVEFORM PROCESSING AND DELAYLESS FILTERING

network. The network structure is somewhat large (3-23-23-3)

and it took 11 h training time (1000 epochs) with a Pentium
4 PC at 2.5-GHz frequency, and the resulting MSE converged
to 2.29187e-005. After adequate training, the estimated waves
shown in Fig. 9, were generated by test for intermediate values.
The ripple in the estimated waves slightly deteriorates at
lower frequency, as indicated in Fig. 9(d), but the results are
acceptable. The network could be trained satisfactorily below
5.0 Hz and at higher than 60 Hz with appropriate modulation
factor. A small amount of dc offset at the input PWM waves
(introduced by the sensors), could be filtered by the ANN, but
this performance has not been adequately investigated. Instead
of generating three-phase delayless sine waves, additional
processing required for feedback signal processing of the drive
can be incorporated into the same ANN. For example, simple
3- to 2- transformation by the following equations [1]:
(3)

989

to 0.98 for the fundamental component, and the corresponding

channel are 1.8 20.6 and 0.99950.936,
values for the
respectively. The ANN is responsible for compensation of these
frequency-sensitive phase delays and attenuations. The data
generated at different magnitudes of current and at different
frequency were used for the ANN training. The designed ANN
topology was 2-25-25-2 after 17.5 h (600 epochs) of training
e
.
with a 1.7-GHz PC, and the resulting MSE
The estimation performance can be slightly improved with
larger training data and more training time. Figs. (10(b)(d)
shows comparison of the desired and estimated waves at different frequencies and magnitudes which indicate excellent
performance. Additional harmonics in the current waves due
to lower PWM frequency can also be filtered by ANN. Again,
dc offset in the current sensors can be eliminated, although this
performance has not been studied. Additional input signals can
be added to the same network and feedback signals, such as
, torque
, and unit vector signals (
,
rotor flux
) can be estimated for vector-controlled drives [8].

(4)
was incorporated into the same ANN. In fact, for two outputs
(
and
), the ANN structure was found to be slightly simpler (3-21-21-2), i.e., 47 neurons instead of 49, and the training
time was lower (500 epochs, 4.5 h).

C. Two-Phase Current Waves at Variable Frequency

In the last phase of the study, let us consider machine current waves at variable frequency from a SPWM voltage-fed inverter which is used commonly in industrial drives. These current waves were generated by simulation of a vector-controlled
orientation, where
induction motor drive with rotor flux
the inverter switching frequency was set to 5.0 kHz. Fig. 10(a)
shows the network configuration with two phase currents and
for an isolated neutral machine
. The netand
waves by additional
work illustrates estimation of
processing of the following equations [1]:
(5)
(6)
As indicated in the figure, the input current waves are filtered by the inherent filtering effect of the machine. Although
ANN-based direct filtering of these waves is possible [9], it was
and
found that additional LPFs of unequal size (
), as indicated in the figure, decrease the ANN size.
The ANN training is also possible with LPFs of equal , but the
size of ANN tends to be bigger. The frequency range of current
considered was between 560 Hz with current magnitude variation dictated by the load torque on the machine. For this range
of frequency variation, (1) and (2) give the channel phase lag
(due to LPF) from 0.9 to 10.7 and attenuation from 0.9998

IV. CONCLUSION
Waveform processing and delayless filtering capabilities of a
feedforward neural network have been systematically investigated in this paper using offline training, and performance was
found to be excellent. Although the voltage and current waveforms that are relevant to power electronics and the ac motor
drives area have been considered, this general waveshaping
property of a neural network is also important in other areas
of electrical engineering. In the beginning, constant-frequency
(60 Hz) square and six-step waves that are single or multiphase
were considered, and it was demonstrated that neural network
can convert them into filtered sine waves in single or multiphase
at the synchronized phase angles, and the output magnitude
linearly tracks the input magnitude. A small drift in frequency
and deviation in the waveform have only negligible effect on
the network performance. In the second part of the project,
variable-frequency variable-magnitude voltage and current
waves that are relevant to ac motor drives with PWM inverters
were considered and similar waveform processing and delayless filtering capabilities with inputoutput magnitude tracking
capability of neural network were demonstrated. In either case,
besides the harmonic filtering capability, the neural network
provides compensation of phase and magnitude deviations
introduced by the auxiliary filter (LPF) elements. In addition to
waveform processing of voltage or current signals individually,
these signals can be fed to a single network and additional
signal processing calculations can be embedded in it. As a
general conclusion, it can be stated that neural network has
the capability of converting -phase waves of arbitrary shape
into -phase waves of arbitrary shape at the same frequency
with magnitude tracking (linear or nonlinear) and locked phase
angles. These properties are also valid for a variable-frequency
wave if the wave pattern at the network input remains the same.

990

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 51, NO. 5, OCTOBER 2004

Fig. 10. Delayless harmonic filtering and processing of variable-frequency variable-magnitude current waves (f = 5 kHz, MSE = 17:047e 005). (a) Network
showing input (i and i ) and output (i and i ) current waves. (b) Output waves at 60 Hz and T = 2:0 N m. (c) Output waves at 20 Hz and T = 3:5 N m.
(d) Output waves at 5 Hz and T = 0 N m.

ZHAO AND BOSE: NEURAL-NETWORK-BASED WAVEFORM PROCESSING AND DELAYLESS FILTERING

REFERENCES
[1] B. K. Bose, Modern Power Electronics and AC Drives. Upper Saddle
River, NJ: Prentice-Hall, 2002.
[2] B. K. Bose, P. Szczesny, and R. L. Steigerwald, Microcomputer control
of a residential photovoltaic power conditioning system, IEEE Trans.
Ind. Applicat., vol. 21, pp. 11821191, Sept./Oct. 1985.
[3] S. Valiviita, Predictive Filtering Methods for Motor Drive Instrumentation, Ph.D. dissertation, Dept. Elect. Eng., Helsinki Univ. Technol.,
Espoo, Fiinland, 1998.
[4] O. Vainio and S. J. Ovaska, Multistage adaptive filters for in-phase processing of line-frequency signals, IEEE Trans. Ind. Electron., vol. 44,
pp. 258264, Apr. 1997.
[5] S. Valiviita and S. J. Ovaska, Delayless method to generate current
reference for active filters, IEEE Trans. Ind. Electron., vol. 45, pp.
559567, Aug. 1998.
[6] M. H. Kim, M. G. Simoes, and B. K. Bose, Neural network-based estimation of power electronic waveforms, IEEE Trans. Power Electron.,
vol. 11, pp. 383389, Mar. 1996.
[7] M. G. Simoes and B. K. Bose, Application of fuzzy logic in the estimation of power electronic waveforms, IEEE Trans. Power Electron., vol.
13, pp. 401409, May 1998.
, Neural network based estimation of feedback signals for a vector
[8]
controlled induction motor drive, IEEE Trans. Ind. Applicat., vol. 31,
pp. 620629, May/June 1995.
[9] M. Rukonuzzaman and M. Nakaoka, Adaptive neural network based
harmonic current compensation in active power filter, in Conf. Rec.
IEEE IECON01, 2001, pp. 22812286.
[10] S. Valiviita, Neural network for zero-crossing detection of distorted line
voltages in weak ac systems, in Conf. Rec. IEEE Instrumentation and
Measurement Technol. Conf. (IMTC98), St. Paul, MN, 1998.
[11] S. Valiviita, S. J. Ovaska, and J. Kyyra, Adaptive signal processing
system for accurate zero-crossing detection of cycloconverter phase currents, in Conf. Rec. PCC-Nagaoka, 1997, pp. 467472.

Jin Zhao was born in Hubei Province, China,

in 1967. He received the B.E. and Ph.D. degrees
from the Department of Control Science and Engineering, Huazhong University of Science and
Technology (HUST), Wuhan, China, in 1989 and
1994, respectively.
He has been with the Department of Control
Science and Engineering, HUST, since 1994. He
became an Assistant Professor in 1994 and an Associate Professor in 1997. During 20012002, he was
a Visiting Scholar in the Power Electronics Research
Laboratory, The University of Tennessee, Knoxville. He is engaged in research
and applications of power electronics, electrical drives, and intelligent control.
He is the author or coauthor of more than 70 technical papers.

991

Bimal K. Bose (S59M60SM78F89LF96)

received the B.E. degree from Bengal Engineering
College, Calcutta University, Calcutta, India, in
1956, the M.S. degree from the University of Wisconsin, Madison, in 1960, and the Ph.D. degree from
Calcutta University in 1966.
He has held the Condra Chair of Excellence
in Power Electronics in the Department of Electrical Engineering, The University of Tennessee,
Knoxville, for the last 16 years. Prior to this, he was a
Research Engineer in the General Electric Corporate
R&D Center, Schenectady, NY, for 11 years (19761987), an Associate Professor of Electrical Engineering, Rensselaer Polytechnic Institute, Troy, NY, for
five years (19711976), and a faculty member at Bengal Engineering College
for 11 years (19601971). He is specialized in power electronics and motor
drives, specifically including power converters, ac drives, microcomputer/DSP
control, EV/HV drives, and artificial intelligence applications in power electronic systems. He has authored more than 170 papers and is the holder of 21
U.S. patents. He has authored/edited six books: Modern Power Electronics and
AC Drives (Upper Saddle River, NJ: Prentice-Hall, 2002), Power Electronics
and AC Drives (Englewood Cliffs, NJ: Prentice-Hall, 1986), Power Electronics
and Variable Frequency Drives (New York: IEEE Press, 1997), Modern Power
Electronics (New York: IEEE Press, 1992), Microcomputer Control of Power
Electronics and Drives (New York: IEEE Press, 1997), and Adjustable Speed
AC Drive Systems (New York: IEEE Press, 1981).
Dr. Bose has served the IEEE in various capacities, including Chairman of
the IEEE Industrial Electronics Society (IES) Power Electronics Council, Associate Editor of the IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, IEEE
IECON Power Electronics Chairman, Chairman of the IEEE Industry Applications Society (IAS) Industrial Power Converter Committee, and IAS member
of the Neural Network Council. He has been a Member of the Editorial Board
of the PROCEEDINGS OF THE IEEE since 1995 and the Journal of Intelligent and
Fuzzy Systems since 2001. He was the Guest Editor of the PROCEEDINGS OF
THE IEEE Special Issue on Power Electronics and Motion Control (August
1994). He has served as a Distinguished Lecturer of both the IAS and IES, and
is now the Vice-Chairman of the IAS Distinguished Lecturer Program. He is a
recipient of a number of awards, including the IEEE Millennium Medal (2000),
IEEE Continuing Education Award (1997), IEEE Lamme Gold Medal (1996),
IEEE Region 3 Outstanding Engineer Award (1994), IEEE-IES Eugene Mittelmann Award (for lifetime achievement) (1994), IAS Outstanding Achievement
Award (1993), Calcutta University Mouat Gold Medal (1970), GE Silver Patent
Medal (1986), GE Publication Award (1985), and a number of IEEE prize paper
awards.