Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 53-2 (2020) 8193–8198
Signal
Signal Generation
Generation for
for Switched
Switched Reluctance
Reluctance
Signal
Motors Generation
using for
Parallel Switched
Genetic Reluctance
Algorithms
Signal Generation
MotorsGeneration
Signal for
using Parallel Switched Reluctance
Genetic Algorithms
for Switched Reluctance
Motors
Motors using Parallel Genetic Algorithms
Motors using
Mike using∗ Parallel
Parallel Genetic
Genetic
∗∗ Algorithms
Algorithms
∗
Mike Eichhorn
Eichhorn ∗∗∗ Sandro Sandro Purfürst
Purfürst ∗∗ ∗∗
Yuri
Yuri A.W.A.W. Shardt Shardt ∗∗∗
Mike
Mike Eichhorn
Eichhorn ∗
Sandro
Sandro Purfürst
Purfürst ∗∗
∗∗
Yuri
Yuri A.W.
A.W. Shardt
Shardt ∗
∗ Mike
Mike Eichhorn
Eichhorn ∗ Sandro Purfürst
Purfürst ∗∗ Yuri
Sandro Engineering, Yuri A.W.
A.W. Shardt
Shardt ∗of
∗ Department of Automation Engineering, Technical
Technical University
University of
Department
∗∗ Department
of Automation
Department
Ilmenau,
Ilmenau,
∗ Department
of Helmholtzplatz
of Automation Engineering,
Automation
Helmholtzplatz Engineering,
5,
5, 98693
98693 Technical
Ilmenau,
Technical
Ilmenau, University of
Germany
University
Germany of
∗ Department
Ilmenau, of Automation
Helmholtzplatz
(e-mail:(mike.eichhorn,
of Automation Engineering,
5, 98693
Engineering, Technical
Ilmenau,
yuri.shardt)@tu-ilmenau.de)
Technical University
Germany
University of
of
Ilmenau,
Ilmenau, Helmholtzplatz
(e-mail:(mike.eichhorn,
Helmholtzplatz 5, 98693 Ilmenau,
yuri.shardt)@tu-ilmenau.de)
5, 98693 Ilmenau, Germany
Germany
∗∗ NIDEC (e-mail:(mike.eichhorn,
driveXpert GmbH, yuri.shardt)@tu-ilmenau.de)
Ehrenbergstraße 11, 98693 Ilmenau
∗∗ NIDEC Ilmenau, Helmholtzplatz
(e-mail:(mike.eichhorn,
driveXpert GmbH, 5, 98693 Ilmenau,
yuri.shardt)@tu-ilmenau.de)
Ehrenbergstraße 11, Germany
98693 Ilmenau
∗∗
∗∗ NIDEC (e-mail:(mike.eichhorn,
driveXpert GmbH, yuri.shardt)@tu-ilmenau.de)
Ehrenbergstraße 11,
∗∗ NIDEC Germany
(e-mail:(mike.eichhorn,
driveXpert
Germany (e-mail:spu@drivexpert.de)
GmbH, Ehrenbergstraße
(e-mail:spu@drivexpert.de) 11, 98693
yuri.shardt)@tu-ilmenau.de) 98693 Ilmenau Ilmenau
∗∗ NIDEC driveXpert
NIDEC driveXpert Germany GmbH, Ehrenbergstraße
Ehrenbergstraße 11,
(e-mail:spu@drivexpert.de)
GmbH, 11, 98693
98693 Ilmenau Ilmenau
Germany
Germany (e-mail:spu@drivexpert.de)
Germany (e-mail:spu@drivexpert.de)
(e-mail:spu@drivexpert.de)
Abstract: Switched
Abstract: Switched reluctance
reluctance motors motors (SRM)(SRM) are are an
an inherent
inherent part part in in robotics
robotics and and automation
automation
Abstract:
systems
Abstract: whereSwitched
Switchedenergy reluctance
and
reluctance cost motors
efficiency
motors (SRM)
(SRM)is are an
an inherent
required.
are inherent
This part in
motor
part in
type robotics
robotics has noand
and automation
windings
automation and
systems
Abstract: whereSwitchedenergyreluctance
and cost motors efficiency (SRM)is required.
are an This motor
inherent part type
in robotics has no and windings
automation and
systems
permanent
Abstract:
systems where
where magnets
Switchedenergyreluctance
energy on and
the
and cost
rotor
cost efficiency
which
motors
efficiency (SRM)is required.
results
is required.
are in
an a This motor
simple
inherent
This motor
and
part type
robust
in
type robotics has no
structure.
has noand windings
However,
automation
windings and
and
permanent
systems where magnetsenergy onandthecostrotorefficiency
which results
is in a simple
required. This and robust structure. However,
permanent
SRMs require
systems
permanent
SRMs require
where magnets
aaenergy
magnets complex
complex
on
on the
and
the rotor
electronic
cost
rotor
electronic
which
control
efficiency
which
control
results
system
issystem
results in
required.
in to
to This motor
aa simple
generate
simple
generate
and
motor
and
type
type has
robust
aa specified
specified
robust has no
no windings
structure.
number
structure.
number ofHowever,
windings voltage
However,
of voltage
and
and
permanent
SRMs
pulses
permanent require
for magnets
each a
magnets complex
motor on
on the
phase.
the rotor
electronic
This
rotor which
control
paper
which results
system
presents
results in
in toa
the
a simple
generate
signal
simple and
and a robust
specified
generation
robust structure.
number
of multiple
structure. However,
of voltage
phases
However,
SRMs
pulses
SRMs require
for eacha
require complex
motor electronic
phase. This control system
paper presents tothegenerate aa specified
signal generation number
of multiple of voltage
phases
pulses
using only
SRMs
pulses for one
only
require
for eacha
each
complex
motor
current
acurrent
complex
motor
electronic
phase.
sensor This
inThis
electronic
phase.
control
control
paper
system
paper presents
an asymmetric
asymmetric presents
system tothegenerate
half bridge
bridge
tothe signal
(AHB).
generate
signal a specified
generation
In addition
addition
specified
generation
number
of multiple
to maintain
number
of multiple
of
of voltage
maintain phases
voltage
phases the
using
pulses for one
each motor sensor
phase. in an
This paper presents half the (AHB).
signal In
generation of to
multiple phasesthe
using
pulses
using only
predeterminedfor
only one
each
one current
phase
motor
current sensor
voltages,
phase.
sensor insufficient
in an asymmetric
This
an asymmetric
papercurrent
presents half
half bridge
measurement
the
bridge (AHB).
signal windows
(AHB). In addition
generation
In addition and ofa to maintain
minimal
multiple
to maintaincurrent
phasesthe
the
predetermined
using only one phase voltages,
current sensor in sufficient
an currenthalf
asymmetric measurement
bridge windows and a minimal current
predetermined
ripple
using
ripple only for
for the
predetermined the phase
one individual
current
phase
individual voltages,
sensor
voltages, insufficient
phases
phases anare
are current
further
asymmetric
sufficient current
further half bridge (AHB).
measurement
optimization
measurement
optimization windows
criteria
(AHB).
windows
criteria
In
for
In addition
foraddition and
signal
and
signal
to
to maintain
aa minimal
generation.
minimal
generation.maintaincurrent
current
the
The
the
The
predetermined
generation
predetermined
ripple for of
the a phase
state
phase
individual voltages,
vector
voltages, which
phases sufficient
controls
sufficient
are current
the
current
further measurement
individual
measurement
optimization windows
semiconductor
windows
criteria for and
and for
signal a
a minimal
each
minimal
generation. motor current
phase
current The
generation
ripple for of aindividual
state vector whichare controls theoptimization
individual semiconductor for each motor phase
generation
to achieve
achieve
ripple
generation
to for the ofrequired
aof
the
a
aindividual
a state vector
state
required
vector
phase
phase
phases
whichare
voltage
phases
which
voltage andfurther
controls
further
controls
and
theoptimization
simultaneously
the
simultaneously
individual
individual
criteria
semiconductor
fulfill for
for signal
the multi-objective
criteria
semiconductor
fulfill the multi-objective for each
signal
for
generation.
each
generation. motor phase
optimization
motor
optimization
The
phase
The
generation
to achieve
criteria
generation is aof a state
required
challenging.
of a state vector
phase
Due
vector which
voltage
to the
which controls
and
vast number
controls the
simultaneously
the individual
of possible
individual semiconductor
fulfill the
solutions,
semiconductor multi-objective
a genetic for
for each
eachalgorithm motor
optimization
motor phase
(GA)
phase
to
to achieve
criteria
achieve a
a required
is challenging.
required phase
Due to
phase voltage
the vast
voltage and
and simultaneously
number
simultaneouslyof possible fulfill the
solutions,
fulfill the multi-objective
a genetic algorithm optimization (GA)
criteria
was
to usedis
used
wasachieve
criteria is challenging.
to
toa find
required
challenging.
find state
statephaseDue
Due to
combinationsthe
voltage
to
combinationsthe vastthat
and
vastthatnumber
are
are suitable
simultaneously
number of
of possible
for the
forfulfill
possible
suitable the multi-objective
solutions,
formulated
solutions,
the aa genetic
multi-objective
formulated genetic algorithm
optimization
algorithm
optimization
optimization
optimization (GA)
criteria.
(GA)
criteria.
criteria
was
The
criteriaused is
results
is challenging.
to find
were
challenging.state Due
discussed
Due to
combinations
and
to the
the vastthatnumber
recommendations
vast are
number of
of possible
suitable
about for
possible solutions,
the
the formulated
genotype
solutions, a
a genetic
genetic algorithm
optimization
representation
algorithm (GA)
criteria.
and the
(GA)
was
The
was used
results
used to
to find
were
find state
state combinations
discussed and
combinations that
recommendations
that are
are suitable
about
suitable for
for the
the
the formulated
genotype
formulated optimization
representation
optimization criteria.
and
criteria.the
The
used
was
The results
genetic
used
results to were
operators
find
were discussed
state were and
given.
combinations
discussed and recommendations
Interested
that
recommendationsare about
readers
suitable
about will
for the
find
the
the genotype
detailed
formulated
genotype representation
information
optimization
representation and
about
criteria.
and the
the
used
The geneticwere
results operatorsdiscussedwereand given. Interested readers
recommendations about willthefindgenotype
detailed representation
information about and the
used
software
The
used genetic
results
genetic operators
technical
were
operatorsdiscussedwere
implementation
were given.
and
given. Interested
using the Global
recommendations
Interested readers
Global about
readers will
Optimization
will find
the
find detailed
genotypeToolbox
detailed information
from MATLAB.
representation
information MATLAB.about
and the
about the
software
used technical
genetic operators implementation
were given. using the
Interested readers Optimization
will find Toolbox
detailed from
information about the
used
software genetic
Copyrighttechnical© operators
technical
2020 The were
implementation
Authors. given.
This is Interested
using
an the
open access readers
Global will
Optimization
article under find
the detailed
CC Toolbox
BY-NC-ND information
from license MATLAB.about the
software
software technical implementation
implementation using
using the
the Global
Global Optimization
Optimization Toolbox
Toolbox from
from MATLAB.
MATLAB.
(http://creativecommons.org/
Keywords:
Keywords: Genetic algorithms,
Genetic algorithms, licenses/by-nc-nd/4.0)
Parallel programs,
Parallel programs, Switched Switched reluctance
reluctance motors, motors, Measuring Measuring
Keywords:
span,
Keywords: Genetic
Multi-objective
Genetic algorithms,
optimization
algorithms, Parallel
Parallel programs,
programs, Switched
Switched reluctance
reluctance motors,
motors, Measuring
Measuring
span,
Keywords: Multi-objective
Genetic algorithms,optimization
algorithms, Parallel programs,
programs, Switched Switched reluctance
reluctance motors, motors, Measuring Measuring
span,
Keywords:
span, Multi-objective
Genetic
Multi-objective optimization
optimization Parallel
span, Multi-objective
span, Multi-objective optimization
optimization
1. INTRODUCTION
1. INTRODUCTION the
the sensor
sensor for for an
an unambiguous
unambiguous measurement measurement of of each
each phase phase
1.
1. INTRODUCTION
INTRODUCTION the
the sensor
current.
sensor
current. for
Thus,
for
Thus, an
an unambiguous
the objectives
unambiguous
the objectives measurement measurement
of
measurement this paper
of this paperof are: of
of each
are:
each phase
(i)
phase
(i) to
to
1.
1. INTRODUCTION
INTRODUCTION the
the sensor
present
sensor
current. a for
for
Thus, an
circuit
an unambiguous
state-based
unambiguous
the objectives way
measurement
of of
thislooking
paper of each
at
each
are: phase
switched
phase
(i) to
The working principle of switched
The working principle of switched reluctance motors current. reluctance motors present a circuit
Thus, state-based
the objectives way of of
thislooking
paper at
are: switched
(i) to
The
(SRMs) working
has been
beenprinciple
known for of
for switched
more than 100 reluctance
100 years. However,motors
However, present
reluctance
current.
present
reluctance a
a circuit
motor
Thus,
circuit state-based
control
the objectives
state-based
motor state-based
control and way and way
(ii)
way
(ii) toof of
to
of looking
propose
this paper
looking
propose at
methods
are:
at
methods switched
(i)
switched to
to
The
(SRMs)
The working
has
working principle
known
principle of
of switched
more than
switched reluctance
years.
reluctance motors
motors present
reluctance
generate
present a
aa circuit
motor
control
circuit control
sequence
state-based andusing(ii)
way of
to
of looking
propose
genetic
looking at
methods
algorithms
at switched
(GA)
switched to
(SRMs)
research
The
(SRMs) has
on
working
has been
the
been known
topic
principle
known hasfor
of
for more
increased thanonly
switched
more than 100
100 years. However,
during
reluctance
years. However,
the past
motors reluctance
generate
reluctance motor
a control
motor control
sequenceand
control andusing(ii)
(ii) to
to propose
genetic
propose methods
algorithms
methods (GA)to
to
research
(SRMs) on been
has the topicknown hasfor increased
more than only
100 during
years. the past reluctance
However, generate
with
generate aa control
multi-objective
motor
control sequence
fitness
control
sequence andusing
functions.
using(ii) genetic
to propose
genetic algorithms
methods
algorithms (GA)
(GA)to
research
three decades.
(SRMs)
research on
decades.
has
on the
beenDuetopic
known has
to its
its increased
simple
for moreand and only
low
thanlow during
cost
100cost the
construction
years. past
However, with multi-objective fitness using functions.
three
research
three
without
research on the
decades. theDue
permanent
on the Due
topic
topic
topic
to
to
has
its
magnets,
has
increased
hassimple
increased
simple SRMs
increased and
only
only
low
are
only
during
during
cost
highly
during
the
the past
construction
construction
reliable
the past
and
past
generate
generate aa control
control sequence
with multi-objective sequence
fitness using functions. genetic
genetic algorithms
algorithms (GA) (GA)
three
without decades.
permanentDue to its
magnets, simple SRMs and low
are cost
highly construction
reliable and with multi-objective fitness functions.
three
can reach
three
without decades.
reach
decades.
permanentDue
mechanical
Due to to its simple
velocities
its simple
magnets, SRMs and
ofand low
upare
to cost
more
lowmore construction
than
costthan
highly 100×10
construction
reliable and3 3 with A
multi-objective
A
fitness functions.
can
without
can
min -1
reach
without .
mechanical
permanent
mechanical
This motor
permanent
velocities
magnets,
velocities
type
magnets, can SRMs
SRMs
of
of
also
up
up
to
are
to
be
are highly
more
used
highly than
for
100×10
reliable
100×10and33
position
reliable and a)
a) s A
A ss0_1
can
min reach
-1
. mechanical
This motor velocities
type can of
alsoup to
be more
used than
for 100×10
position 3 a)
a) s A
0_0
A
0_0 vv0 vv1
can
min reach
-1
control
can .
reach
-1 andmechanical
This motor
since
mechanical it velocities
type
does can
not
velocities of
alsoup
require
of up to
be
to more
used
brushes,
more than
for
than this100×10
position
motor
100×10 3
a) s
s 0_0 0 ss0_1 0_1
min
control
min -1 .
. This
This motor
and since
motor type
it type
does not can
can also
require
also be
be used
brushes,
used for
for position
thisposition
motor a) s 0_0
s0_0 A v
vv00 s
s0_1
0_1
A vv111
can -1
control
min be
control and
. used
This
and since
in
motor
since it
hazardous
it does
type
does not
can require
and
not also bebrushes,
explosive
require this
environments.
used
brushes, forthis motor
position
motor 0_0 A v00 0_1 A vv1
can
can
Many
be
control
be
control
used
and
used
in
and since
in
researchers
since
hazardous
it
hazardous does
ithave
doestried
and
not
and
not
explosive
require
explosive
to find
require
environments.
brushes, this
environments.
optimal
brushes, design
this motor
pa-
motor A
A A
A 1
can
Many be used in
researchers hazardous
have and
tried explosive
to find environments.
optimal design pa- A
A s A
A s
can
Many
can
Many
be
rameters used
with in
researchers hazardous
in the
be researchers
used have
help
help of
hazardous
have
and
tried
and
tried
explosive
to
techniques
explosive
to find
environments.
find inspired
optimal design
by
environments.
optimal design pa-
natural
pa- s1_0 1_0 s1_1 1_1
rameters with the of techniques inspired by natural ss1_0 A ss1_1
ss1_1 A
Many researchers have tried to find optimal design pa- 1_0 1_1
rameters
evolution
Many
rameters
evolution with
such
researchers
with the
as
the
such the help
Shaked
have
help
as Shaked of
tried
of techniques
and to Rabinovici
techniques find
and Rabinovici inspired
optimal
inspired by
(2006);
design
by
(2006); natural
Sihem
pa-
natural
Sihem s
s 1_0 A A
rameters
et al.
rameters
evolution with
(2012);
with
such Prabhu
the
as help
help
Shakedet of
ofal. techniques
(2015).
techniques
and inspired
These
Rabinoviciinspiredresultsby
by
(2006); natural
are the
natural
Sihem A 1_0 A
A 1_1 A
A
et al. (2012); Prabhu et al. (2015). These results are the A
A A A
evolution
et al.
motivation
evolution
et al.
such
(2012);for
such
(2012);
as
Prabhu
this
as
Prabhu
Shaked
paper
Shakedet
et al.
to
al.
and
andfocusRabinovici
(2015).
(2015). onThese
Rabinovicithe
These
(2006);
results
optimization
(2006);
results
Sihem
are the
Sihem
are theof A A A
motivation
et al.
motivation
switching
et al. (2012);forPrabhu
for
signals
(2012);
this paper
this
Prabhu forpaperet
et
to (2015).
al.
to
al.
focus on
focus
semiconductors.
(2015). on
the optimization
These
the
For
These results
optimization
cost
results are
reduction,
are the
the
of
of A
A
motivation
switching signals
motivation for
for this
this paper
forpaper to focus
semiconductors.
to
to focus
on
on the optimization
For optimization
the cost reduction, of
of b) A
A
switching
integration
motivation
switching signals
for
integrationsignals thisfor
density for semiconductors.
improvement
paper focusand
semiconductors. For
the
For cost
on redundancy, reduction,
optimization
cost several
reduction, of b) ss0_0 A
A ss0_1
switching density
signals forimprovement
semiconductors. and redundancy,
For cost several
reduction, b) s0_0
b) A vv0 0_1 vv1
integration
position
switching
integration
position
integration
density
sensorless
signals
density
sensorless
density
forimprovement
algorithms
semiconductors.
improvement
algorithms
improvement
have
have and
and
and
been
beenredundancy,
Fordeveloped.
redundancy,
developed.
redundancy,
several
cost reduction, Using
several
Using
several b)
b) ss0_0s 0_0 i
i00 vv000 sss0_1
0_1 ii1 vv111
position
one
integration
position
one sensorless
current sensor
density
sensorless
current sensor algorithms
for multiple
improvement
algorithms
for multiple have
have been
measurements
andbeen
measurementsdeveloped.
redundancy,
developed. nearly
nearlyUsing
several
Usingsi-
si- 0_0
0_0 i
ii0 v s 0_1
0_1 i
ii1
1
vv1
position
one current sensorless
current
multaneously sensor
is algorithms
for multiple
multiple have been
been developed.
measurements nearlyUsingsi- 0 v00 1
1
position
one
multaneously
one current is another
sensorless
sensor
sensor
algorithms
anotherfor
for
concept
concept
multiple
have discussed
measurements
discussed
measurements
in
in this
developed. this paper.
nearlyUsing
paper.
nearly si-
si- i0
0 ss1_0 i 1
1 ss1_1
multaneously
The
one focus
current
multaneously here is
sensor
is
The focus hereisisanotheranother
is on
another an
for concept
asymmetric
multiple
concept
on an asymmetric discussed
half
measurements
discussed
half bridgein
bridge
in this paper.
converter
nearly
this paper.
converter si- 1_0 1_1
multaneously
The focus
topology
multaneously
The focus
topology
here
with
here
with is is
one
one
on
another
is on an
current
an
current
concept
asymmetric
sensor
sensor for
concept
asymmetric
discussed
half
two
discussed
for half
in
in this
bridge
two phases,
bridge
phases, this paper.
converter
based
converter
based on
paper.
on
sss1_0 1_0 sss1_1 1_1
The
Kjaer
The focus
topology and
focus
topology
here
with is
one on
Gallegos-Lopez
here
with is
one on an
current
an
current
asymmetric
sensor
(1997)
asymmetric
sensor for
and
for
half
two
Chen
half
two
bridge
phases,
and
bridge
phases,
converter
Lubased on
(2013).
converter
based on s1_0 1_0 s1_1 1_1
Kjaer
topology and Gallegos-Lopez
with one current (1997)
sensor and
for Chen
two and
phases, Lu (2013).
based on
Kjaer
Fig.
Kjaer
Fig. 1
topology and
a)
and
1 and Gallegos-Lopez
shows
with the
one
Gallegos-Lopez
a) shows asymmetric
current
the asymmetric (1997)
sensor
(1997) and
half
halffor
and Chen
bridge
two
Chen
bridge and
for
phases,
and
for a Lu
two
Lu
a two (2013).
phases
based on
(2013).
phases Fig. 1. AHB with a) different current sensor placements
Kjaer
Fig. Gallegos-Lopez
11possible
a) shows the asymmetric (1997) and
half Chen
bridge and
for Lu (2013).
with
Kjaer
Fig.
with
Fig. 1
and
a)
possible
a) shows
current
Gallegos-Lopez
shows the
current
the
sensor
asymmetric
sensor
asymmetric
(1997) locations.
and
half
half
Chen
bridge
locations.
bridge
Fig.
for
Fig.
for 1a
1
and a
a
two
b)
two
b)
two
phases
Lushows
(2013).
phases
shows
phases
aa Fig. 1. AHB with a) different current sensor placements
with
variant,
Fig.
with 1
variant,possible
a) where
shows
possible current
only
the
wherecurrent
only one one sensor
current
asymmetric
sensor
current locations.
sensor
half bridge
locations. isFig.
sensor isFig. used
for
used 1
1 a b)
to
two
b)
to shows
sense
shows
sense all
phases
all aa Fig. Fig.
Fig.
1.
and
1.
and
1.
AHB
AHB
AHBb) aa with
b) singlea)
single
with
with a)
a)
different
sensor.
different
sensor.
different Thecurrent
The lines
lines show
current
current
sensor
sensor
show
sensor
placements
current
placements
current
placements
paths
paths
with
variant,
phase
with possible current
wherecurrent
currents.
possible only
This one sensor
current
requires
sensor alocations.
sensor
single
locations. Fig.
isFig.
currentused 1
path
1 b)
to
b) shows
sense
through
shows all a
a Fig. and
for
1.
and a
AHBb)
b) a
valid
a single
with(green)
single a) sensor.
different
sensor. and The
an
The invalid lines
current
lines show
(red)
sensor
show current
measurement.
placements
current paths
paths
variant,
phase where
currents. only
This one current
requires a sensor
single is
currentused pathto sense
through all for
and a valid
b) a (green)
single sensor. and an
The invalid lines (red)
show measurement.
current paths
variant,
phase
variant, where
currents. only
where onlyThis one current
requires
one current a sensor
single
sensor is
currentused
is used pathto sense
through
to sense all
all for
and
for a
a valid
b) a
valid (green)
single
(green) sensor. and
and an
The
an invalid
invalid lines (red)
show
(red) measurement.
current
measurement. paths
phase
phase currents.
currents. This
This requires
requires a
a single
single current
current path
path through
through for a valid (green) and an invalid (red) measurement.
2405-8963 Copyright © 2020 The Authors. This is an open access article under the CC BY-NC-ND license.
phase currents. This requires a single current path through for a valid (green) and an invalid (red) measurement.
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2020.12.2328
8194 Mike Eichhorn et al. / IFAC PapersOnLine 53-2 (2020) 8193–8198
2. GENETIC ALGORITHM
2.1 Representation
Fig. 1 shows an asymmetric half bridge with two switches
ns = 2 for each of the two phases n p = 2. This results in 16
possible states zi in one time step Ti . A control sequence
for the SRM with nst time steps can be described easily
with a state vector Z̄ consisting of nst states. This state
vector, which can be called phenotype in this form, has
to be transformed into a suitable genotype representation
to be operated by the genetic algorithm. First of all, one
individual can be understood as an array containing the
switching states ss p of each semiconductor. This way
a binary string represents the genetic information for a
specific individual which is shown in Fig. 2.
The decimal value hst of an asymmetric half bridge with
ns switches for each of the n p phases is defined using
n
p −1 n
s −1
ns n p −jns −i−1
hst = (si j 2
j=0 i=0
which is used as the state number for the practical imple-
mentation.
The switching state of every semiconductor can be consid-
ered as a single gene with two alleles. The first allele stands
for a conducting switch (=1) and the second allele stands
for non-conducting switch (=0). More detailed information
about the state vector concept can be found in Purfürst
(2019).
The following three different types of genotype represen-
tation are discussed in this paper:
Binary-representation
Every bit of the whole stream can be manipulated indepen-
dently. The genetic algorithm only operates at the binary
level.
Grouped-binary-representation
The genetic algorithm is not allowed to apply the crossover
mechanism to each bit of the stream. Only specific points
between binary groups of ns n p bits are possible breakages.
Instead, the mutation operator can be used on every bit.
Integer-representation
The whole bit stream is interpreted as a number of binary
groups represented by their integer value hst . Genetic
operators only work at this level.
z(t)
Population P(G) 0 T (nst-1)T
Chromosome I 0 z0 = 0101 z1 = 1100 ... znst −1 = 0100 = 01011100...0100
Chromosome I1 z0 = 0100 z1 = 1110 ... znst −1 = 0000 = 01001110...0000
...
...
...
Chromosome I nc −1 z0 = 0101 z1 = 0100 ... znst −1 = 1110 = 01010100...1110
Switching state s0 _ 0 s1_ 0 s0 _1 s1_1 State z1 State vector Z
Fig. 2. Population of bit string individuals for a SRM with
ns = 2 and n p = 2 according Fig. 1
2.2 Fitness function formulation
The use of a state vector for control of an asymmetric
half bridge will have some effects on the electrical, and
furthermore, the magnetic and mechanical behavior of the
motor. The resulting CPU requirements also have to be
taken into account. The following list contains possible
properties:
• average phase voltages
• current ripples
• current measurement windows
• capacitors balancing
• switching frequency
• switching equality between high-side and low-side
switches
• necessary memory space
• efficiency
In this paper, the first three requirements will be consid-
ered.
) (1)
To define a fitness function that evaluates the fitness
of each individual corresponding to the phase voltage
achievement, an error needs to be defined that repre-
sents the deviation between a number of specified refer-
ence phase voltages v pr e f and the actual average voltages
v pa c t ( Z̄) forced by the state vector for each phase. The fit-
ness value f itvolt age can be calculated using the Euclidean
distance from the error vector |e| as follows:

(v0r e f − v0 a c t ( Z̄))2 + ...+
f itvolt age = |e| = (2)
(vn p −1r e f − vn p −1 a c t ( Z̄))2
For the provision of current measurement windows, the ge-
netic algorithm is used to minimize the difference between
the desired number of measurement windows n pm d e s and
the actual number of measurement windows n pm a c t for a
specified phase current i p in phase p. The corresponding
fitness function f itwindow p is therefore
f itwindow p = max(n pm d e s − n pm a c t , 0) (3)
The max operator allows the user to set the lower bound of
the fitness value to zero, because additional measurement
windows are not necessary after achieving the desired
number n pm d e s .
The ripple which represents the difference between the
maximum current max(i p ( Z̄)) and the minimum current
min(i p ( Z̄)) of a phase controlled by the state vector is the
third optimization criterion for this paper which can be
t calculated as
f itripple p = max(i p ( Z̄)) − min(i p ( Z̄)) (4)
To merge all fitness functions described above, a weighting
function needs to be defined. This over-all fitness function
allows to weight each optimization criterion using the
...
weighting factors wvolt age , wripple p and wwindow p , that is,
f itΣ =wvolt age f itvolt age +
wripple0 f itripple0 + ..+
Gene Genome wripple n p −1 f itripple n p −1 + (5)
wwindow 0 f itwindow 0 + ..+
wwindow n p −1 f itwindow n p −1
 Mike Eichhorn et al. / IFAC PapersOnLine 53-2 (2020) 8193–8198 8195

3. PRACTICAL IMPLEMENTATION 1
I_AB Out1
2
[T,U] PLECS
PWM Out2
fcn Circuit
For the practical implementation a set of properties for StateVector
de2bi
U_AB
Terminator
all possible states needs to be defined. The system to be 2 Asymmetric Half-Bridges

implemented is represented by 2ns n p = 16 states which Lf

is shown in Table 1. Each of these states forces a phase Rshunt_HS

voltage v p which results in a current i p . To be able to PWMA1 A1 PWMB1 A3

measure the phase currents with a reduced number of ESR

I_A A I_A I_B A I_B
phase current sensors, every state includes information R_A R_B

about the measurability of the necessary phase currents 12 Vdc1 L_A DA2 DA1 Vm1 V
U_A
L_B DA3 DA4 Vm2 V
Constant U_B
(green marked fields in Table 1). Each state is accessed C_ZK

by a decimal number index hst from (1) represented by a PWMA1 PWMA2 A2 PWMB2 A4 I_A
1
PWMA2 I_B
binary code that includes information about the switching ‹1›
PWM PWMB1
PWMB2
U_A
I_AB

2
state ss p of each semiconductor. U_B U_AB

Table 1. State properties Fig. 3. Simulink system of the two simultaneously excited
asymmetric half bridges using the PLECS Blockset
State Number Binary Voltage Voltage Meas. Meas.
hs t Code v0 v1 p=0 p=1 Listing 1. MATLAB code example
% D e f i n e t h e GA p a r a m e t e r s
0 0000 -12 V -12 V 0 0 options = gaoptimset (
1 0001 -12 V 0 V 0 0 ’ Populati onSize ’ , 2 0 0 , . . .
2 0010 -12 V 0 V 0 1 ’ Generations ’ , 1 5 0 , . . .
3 0011 -12 V +12 V 0 1 ’ SelectionFcn ’ , { @ s e l e c t i o n r o u l e t t e } , . . .
4 0100 0 V -12 V 0 0 ’ CrossoverFraction ’ , 0 . 7 , . . .
5 0101 0 V 0 V 0 0 ’ FitnessScalingFcn ’ ,{ @ f i t s c a l i n g r a n k } , . . .
6 0110 0 V 0 V 0 1 ’ EliteCount ’ , 4 , . . .
7 0111 0 V +12 V 0 1 ’ StallGenLimit ’ , 1 0 0 , . . .
8 1000 0 V -12 V 1 0 ’ PlotFcns ’ , { @ g a p l o t b e s t f } , . . .
9 1001 0 V 0 V 1 0 ’ Display ’ , ’ iter ’ , . . .
10 1010 0 V 0 V 0 0 ’ CreationFcn ’ , { @ g a c r e a t i o n u n i f o r m } , . . .
11 1011 0 V +12 V 0 0 ’ Populati onType ’ , ’ bitstring ’ , . . .
12 1100 +12 V -12 V 1 0 ’ MutationFcn ’ , { @mutationuniform , 0 . 0 2 } , . . .
13 1101 +12 V 0 V 1 0 ’ CrossoverFcn ’ , { @ c r o s s o v e r t w o p o i n t B y t e } , . . .
14 1110 +12 V 0 V 0 0 ’ UseParallel ’ , t r u e ) ;
15 1111 +12 V +12 V 0 0 % Define the q u a l i t y requirements
weightVec =
3.1 Program details [ wVoltage wWindow0 wWindow1 wRipple0 wRipple1 ] ;
refVoltageVec = [10 1 0 ] ;
measurementWindowVec = [ 5 5 ] ;
For the implementation of the genetic algorithms, MATLAB % Define the f i t n e s s f u n c t i o n handle
and Simulink were used. The Global Optimization Toolbox f i t n e s s F c n = { @fitnessFcnSRM , r e f V o l t a g e V e c ,
(MathWorks (2020a)) from MathWorks provides methods measurementWindowVec , weightVec } ;
to solve problems with multiple optima. It includes genetic % S t a r t t h e GA
algorithms which have a wide range of standard mecha- [ bestChromosome , bestQ , ˜ , ˜ , p o p u l a t i o n , s c o r e s ] =
nisms. The toolbox allows the creation of a custom genetic ga ( f i t n e s s F c n , nVars , o p t i o n s ) ;
algorithm by modifying the functions for population ini-
tialization, fitness scaling, parent selection, crossover, and
mutation. The multi-objective genetic algorithm can also 4. RESULTS
solve multiple-objective optimization problems by identi-
fying the Pareto front. The possible simultaneous calcu- The results presented in this section are based on the
lation of the fitness function for the several individuals GA parameter settings in Table 2. The basics section in
enables parallelization using of the Parallel Computing the upper part of the table includes parameters that are
Toolbox (MathWorks (2020b)). This gives a rapid speedup used in each representation (see section 2.1). A weighted
during the optimization of the GA in case of a computa- fitness function according to (5) was used for the require-
tionally intensive fitness function. For a realistic and fast ments evaluation of the several individuals. All weighting
simulation of the electrical behaviour of the asymmetric factors for the used requirements (wvolt age , wripple p and
half bridge within a Simulink environment, the PLECS wwindow p ) were set to 1.0. This pragmatic weighting factor
Blockset (Plexim (2020)) was used. Fig. 3 shows the result- combination leads to desired results. The settings in the
ing Simulink-system with the integrated PLECS Circuit following table are dependent on the individual represen-
block. tations. The parameter mutation fraction was adapted to
The GA in MATLAB uses the standard interface for op- achieve a maximum convergence rate for the several rep-
timization routines, and therefore, allows easy access. A resentations. For most functions in the GA’s, the standard
good example for the usage of a binary genetic algorithm in MATLAB procedures are used. However, for the crossover
MATLAB is described in (Babatunde (2020); Babatunde function in the grouped-binary-representation and the cre-
et al. (2014)). Listing 1 shows a possible implementation ation and mutation function of the integer-representation,
using a grouped-binary-representation. custom functions had to be designed.
8196 Mike Eichhorn et al. / IFAC PapersOnLine 53-2 (2020) 8193–8198

Table 2. GA parameters a) b) c)
20
Best fitness
10 Mean fitness

fit
GA parameter value
Population size 200 0
Number of generations 150 10

fitvoltage
Crossover fraction 0.7 5
Fitness function Rank based fitness scaling
Elite count 4 0
1
Mutation fraction 0.005 (Bin.), 0.02 (Group.), 0.04 (Int.)

fitwindow
Binary and grouped-binary-representation 0.5
Creation function Uniform
Population type Bitstring 0
Mutation function Uniform 5

fitripple
Binary and integer-representation
Crossover Two point crossover
Grouped-binary-representation 0
0 50 100 150 0 50 100 150 0 50 100 150
Crossover Two point crossover (custom function)
Integer-representation Generation Generation Generation
Creation function Uniform integer (custom function)
Population type Double vector Fig. 4. Convergence of the GA with a) Binary- b) Grouped-
Mutation function Uniform integer (custom function) binary- c) Integer-representation

15
14
20
4.1 Convergence behavior 13
12
40 11
The three representations presented in Section 2.1 were 10

State Number
Generation

60 9
used for convergence analysis. Each phase should have an
8
average voltage value of about 10 V. As can be seen in 80 7
Fig. 4 b), the best individual is not changed anymore 6
after generation 54. Fig. 4 a) and 4 c) will need more 100 5
generations to get the solution. The reason for the worse 4
convergence of the binary-representation in comparison to 120 3
2
the grouped-binary-representation is due to “additional 1
140
mutations” when a crossover is carried out as a result 0
of the arbitrary choice of breakages. This often leads to 10 20 30 40 50
new states if the break position is not an integer multi- State z
ply of ns × n p . The reason for the poorer convergence of
the integer-representation in comparison to the grouped- Fig. 5. Best chromosome in generation with grouped-
binary-representation needs to be searched for in their mu- binary-representation (color corresponds to state)
tation algorithm. In the first step, the mutation algorithm
selects several states in the state vector of an individual. 4.2 Optimized PWM pattern
Secondly, the algorithm replaces these state elements with
a uniformly distributed random number between 0 and 15 The optimization result has to be translated back into a
(2ns n p −1). For comparison, the mutation algorithm used in usable PWM pattern to be applied to each of the four
the binary- or grouped-binary-representation changes only switches. Fig. 6 shows the generated pattern within a time
by single bits in the bit string. This leads to mutated states interval of 50 µs for a voltage of about 10 V per phase and
which only differ in one bit from the original state. This a number of phase current measurement windows of about
optimal mutation behavior cannot be reproduced by the 5 windows. Fifty states were optimized to reach an average
integer mutation algorithm. Furthermore, the mutation reference voltage, a given number of current measurement
fraction parameter of the integer-representation, where the windows and minimal current ripple for each phase. The
genotype length corresponds to the state vector length of state number is shown for every state upper the time axis
50, needs to be increased with regards to the grouped- in the upper plot of Fig. 6.
binary-representation. For these reasons, the grouped-
4.3 Multi-objective optimization
binary-representation is favored and will be used for the
following optimizations.
In previous tests, a weighted fitness function was used.
Fig. 5 shows the convergence of the GA. Each row contains This requires a suitable choice of the weighting factors to
the states of the best state vector found in each generation. achieve the qualitatively desired result. This often is a trial
The color indicates the state number which is generated and error process which results in many test runs. The
by the switching states of each semiconductor as a binary Global Optimization Toolbox from MathWorks provides
string converted into its decimal representation. It shows a genetic algorithm named gamultiobj to solve multi-
that the evolutionary convergence is very fast until gener- objective optimization problems by finding an evenly
ation 60. After that, only small changes can be observed distributed set of points on the Pareto front. This ap-
until the algorithm finds the optimum at generation 120. proach also allows the optimization of nonsmooth prob-

s0_0
s1_0
s0_1
s1_1
15 15 15 15 15 15 7 15 14 15 13 15 15 15 7 15 15 14 15 7 15 15 15 14 15 15 11 13 15 15 15 15 15 9 15 15 7 13 13 15 15 7 15 15 15 10 15 15 15 15
0 0.5 1 1.5
Measurement Window
0
0 0.5 1 1.5
75
Phase 0 Phase 1
Current in A
70
65
0 0.5 1 1.5
Fig. 6. PWM pattern after optimization
lems (MathWorks (2020a)). For the test, the grouped-
binary-representation with the main part of the previous
settings was chosen. The required phase voltages are 10
V for each phase and the number of phase current mea-
surement windows should be 5. To achieve a good conver-
gence, the number of generations had to be increased to
400. The selection function had to be changed to tour-
nament selection, which is the only supported function
in the gamultiobj algorithm. Two additional parame-
ters - the Pareto fraction (0.5) and migration fraction
(0.35) - were defined. Fig. 7 shows the Pareto front. It is
characterized by three two-dimensional tradeoff curves for
the discrete measurement window fitness values f itwindow
of 0, 1 and 2. Most individuals of the Pareto set lie
on the inner curve where f itwindow = 0. The red star
marks the chosen optimal individual with the fitness val-
ues f itvolt age =0.179, f itwindow =0 and f itripple =0.915. In
comparison to the results of the previous tests where a
weighted fitness function was used, all fitness values are
equal to or smaller than those (for comparison grouped-
binary-representation: f itvolt age =0.179, f itwindow =0 and
f itripple =1.026). This shows the efficiency of the multi-
objective optimization and its application where a large
number of requirements needs to be fulfilled (see list in
section 2.2). The disadvantage of this approach is the
increased computing time in comparison to a GA using
a weighted fitness function. A possible combination of the
two approaches is to use the found Pareto set to detect
1.5
1 Optimal individual
fitripple
0.5
0 evaluations decreases while the communication time in-
0 2
10
fitvoltage
Fig. 7. Pareto front
Mike Eichhorn et al. / IFAC PapersOnLine 53-2 (2020) 8193–8198 8197
suitable weighting factors for a classical GA. The design
idea is to find a weighting factor set which leads to the
smallest fitness value f itΣ for the chosen optimal individual
2 2.5 3 3.5 4 4.5 5
in comparison to all other individuals of the Pareto set.
t in s 10-5
4.4 Parallelization
2 2.5 3 3.5 4 4.5 5
To analyze the benefit of parallelization, the GA algorithm
t in s 10-5 was run on a Windows 64-bit operating system using
MATLAB 2016b on a Dell Precision T5600 Workstation
with two Intel Xeon E5-2665 processors. This allows the
usage of 16 processor cores. For the tests, the grouped-
2 2.5 3 3.5 4 4.5 5 binary-representation with the Simulink system in Fig. 3
t in s 10-5
was used. All tests ran three times with a high priority set
for the MATLAB process. Fig. 8 shows the computing time
with different worker threads nw . The ideal and the real
speedup as well as modeled speedup curves S(nw ) using
Amdahl’s law (adapted for GA)
Ts + n p Te
S(nw ) = n (6)
Ts + Te nwp
and a developed speedup equation for GA
Ts + n p Te
S(nw ) =   (7)
n
Ts + nw Tc + Tep nwp
which combines the speedup equations of Cantú-Paz
(2000) and Trobec (2009) are shown in Fig. 9. This equa-
tion considered all the observed effects by the parallel
GA computation and allows a better modeling of the real
speedup in comparison to Amdahl’s Law. The usage of
two individual values for the single fitness calculation time
(serial execution: Te , parallel execution: Tep ) allows the
modeling of the superlinear speedup for a low number
of workers (nw < 6) as a result of a more efficient us-
age of resources using the Parallel Computing Toolbox
(Te > Tep ). These two values also acknowledge the fact that
the fitness calculation time is dependent on the simulation
time of the Simulink system which can vary. According
to Trobec (2009), the value of Tep describes the time for
the concurrent execution of nw fitness evaluations in a
group and, hence, it corresponds to the longest period
of the several fitness evaluations in the group. The serial
start of all worker processes as well as the communication
time between the GA (master) and the fitness evaluations
(slaves) are considered by the product of nw Tc according
fitwindow = 0 to Cantú-Paz (2000). The ceiling  function
 () in the
n
fitwindow = 1 calculation of the split groups nwp considers the fact
fitwindow = 2 that the last group with fewer fitness evaluation calls as
available worker threads needs the same time as the other
groups. The parameters in (6) and (7) are calculated from
regression using the measured computing times. Table 3
includes the parameter descriptions and the calculated
values. The analysis of (7) shows that with the increasing
number of workers, the computation time for all fitness
creases. The number of workers needed to approach the
w)
1 minimal execution time can be solved using ∂S(n ∂nw = 0.
20 0
For these tests, the optimal number nw is 29, where a
fitwindow maximum speedup S(nw ) of 14.75 will be achieved. An
estimation of the necessary workers for an efficient work
can be determined using the parallel efficiency
8198 Mike Eichhorn et al. / IFAC PapersOnLine 53-2 (2020) 8193–8198

Table 3. Speedup equation parameters 5. CONCLUSION

This paper presented a concept to generate control se-

Parameter value
quences for a SRM with a reduced amount of current
Population size n p 200 sensors using GA. Different genotype representations were
Number of generations 150 tested for increasing the convergence rate. To fulfill all
Amdahl’s Law (6) requirements, a weighted multi-objective fitness function
Single generation time without fitness calculations Ts 0.487 s as well as a Pareto front solution were used successfully.
Single fitness evaluation time Te 0.230 s
Equation for GA (7)
An approach to determine the number of cores to solve
Single generation time without fitness calculations Ts 0.132 s GA algorithms efficiently was presented. The software
Single fitness evaluation time (serial execution) Te 0.236 s implementation based on MATLAB is discussed in detail.
Communication time Tc 0.053 s
Single fitness evaluation times (parallel execution) Te p 0.222 s REFERENCES
Babatunde, O. (2020). Bi-
S(nw ) nary genetic algorithm hezy 2013. URL https://
E(nw ) =
(8)
nw www.mathworks.com/matlabcentral/fileexchange/
Using an efficiency E of 0.8, which corresponds to an 46961-binary_genetic_algorithm_hezy_2013.
80% performance of the parallel GA algorithm of its ideal Babatunde, O., Armstrong, L., Leng, J., and Dean, D.
potential, results in a number of workers nw = 15.6 or (2014). A genetic algorithm-based feature selection. In-
rather 15. This means the usage of more than 15 workers ternational Journal of Electronics Communication and
does not lead to any significant computing speedup effects. Computer Engineering, 5(4), 889–905.
The equations presented above and the relationships allow Cantú-Paz, E. (2000). Efficient and Accurate Parallel
a rough estimation of the necessary hardware to solve GA Genetic Algorithms, volume 1. Kluwer Academic Publ,
algorithms efficiency in terms of time and cost. Boston, Mass.
Chen, H. and Lu, S. (2013). Fault diagnosis digital
method for power transistors in power converters of
7000 switched reluctance motors. IEEE Transactions on
Industrial Electronics, 60(2), 749–763. doi:10.1109/TIE.
6000
2012.2207661.
Computing Time T in s

5000 Kjaer, P.C. and Gallegos-Lopez, G. (1997). Single-sensor

current regulation in switched reluctance motor drives.
4000 In IAS ’97. Conference Record of the 1997 IEEE Indus-
try Applications Conference Thirty-Second IAS Annual
3000 Meeting, volume 1, 645–652 vol.1. doi:10.1109/IAS.
2000
1997.643138.
MathWorks (2020a). Global Optimization Tool-
1000 box. URL https://www.mathworks.com/products/
global-optimization/.
0 MathWorks (2020b). Parallel Computing Toolbox.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
URL https://www.mathworks.com/products/
Number of Workers n w
parallel-computing/.
Plexim (2020). Plecs Blockset - The Power Electronics
Fig. 8. Computing time using several worker threads Toolbox for Simulink. URL https://www.plexim.com/
16 products/plecs_blockset.
Ideal Speedup Prabhu, V.V., Rajini, V., Balaji, M., and Prabhu, V.
14 Real Speedup (2015). A comparative study of operating angle op-
Amdahl's Law (6) timization of switched reluctance motor with robust
12 Equation for GA (7) speed controller using pso and ga. Journal of Electrical
Engineering and Technology, 10(2), 551–559. doi:10.
Speedup S

10
5370/JEET.2015.10.2.551.
8 Purfürst, S. (2019). Beitrag zur schaltungszustands-
basierten Ansteuerung geschalteter Reluktanzmaschi-
6 nen unter Berücksichtigung der Strommessung bei re-
duzierter Sensorik. Ph.D. thesis, TU Ilmenau.
4 Shaked, N.T. and Rabinovici, R. (2006). Structural
switched reluctance motor optimization using the ge-
2 netic algorithm. Journal of Electrical Engineering, 6(4).
Sihem, M., Amar, B., and Hocine, B. (2012). Teeth
1 2 4 6 8 10 12 14 16
shape design of a switched reluctance motor for high
Number of Workers n w
torque using genetic algorithms. International Journal
of Scientific & Engineering Research, 3(10), 1128–1132.
Fig. 9. Ideal, real and modeled speedup curves Trobec, R. (2009). Parallel Computing - Numerics, Appli-
cations, and Trends. Springer, Dordrecht.