Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 54-1 (2021) 1059–1064
Motor
Motor Classification
Classification with
with Machine
Machine Learning
Learning
Motor
Motor Classification
Classification
Methods for with
with
Predictive Machine
Machine
MaintenanceLearning
Learning
Motor Classification
Methods for with
Predictive Machine
MaintenanceLearning
Methods
Methods for
for Predictive
Predictive Maintenance
Maintenance
Methods for Predictive Maintenance
Christoph Kammerer ∗∗ Michael Gaust ∗∗
∗∗
Micha Küstner ∗∗
Christoph
Christoph Kammerer
Kammerer ∗ Michael Gaust ∗∗ Micha Küstner ∗
∗ Michael Gaust ∗∗ Micha Jesser Küstner
Christoph
Pascal Kammerer
Starke ∗∗
∗∗ Roman∗ Michael
∗ Radtke ∗
Gaust ∗∗ Micha
∗ Alexander
∗∗ Küstner ∗ ∗
∗ ∗
Pascal
Christoph
Christoph
Pascal Starke
Kammerer
Kammerer
Starke ∗∗
∗∗
Roman
∗ Michael
Roman Radtke
Michael
Radtke ∗ Alexander
Gaust
Gaust
∗ ∗∗ Micha
AlexanderMicha Jesser
Küstner
Küstner
Jesser ∗ ∗
∗ ∗
Christoph
Pascal
Pascal Kammerer
Starke
Starke ∗∗ Roman Michael
Radtke Gaust
∗ AlexanderMicha Küstner
Jesser ∗
∗∗ Roman Radtke ∗ Alexander Jesser
∗∗ ∗ ∗
∗ Pascal
Pascal Starke
Starke Roman
Roman Radtke
Radtke Alexander
Alexander Jesser
Jesser ∗
∗ University of Applied Sciences Heilbronn, Max-Planck-Str. 39, 74081
∗ University of Applied Sciences Heilbronn, Max-Planck-Str. 39, 74081
∗ University of Applied Sciences Heilbronn, Max-Planck-Str. 39, 74081
∗ University
∗ Heilbronn,
Heilbronn,
University of
of Applied
Germany Sciences
Applied(e-mail:
Germany Heilbronn,
Heilbronn, Max-Planck-Str.
Max-Planck-Str. 39,
[email protected])
(e-mail: [email protected]) 39, 74081
74081
∗ University
Heilbronn,
∗∗ of Applied
Germany Sciences
(e-mail: Heilbronn, Max-Planck-Str.
[email protected]) 39, 74081
University
Heilbronn,
∗∗ CeraCon
∗∗ CeraCon
Heilbronn, of Applied
Germany
GmbH,(e-mail:
GmbH,
Germany Sciences
Talstraße
Talstraße
(e-mail: Heilbronn,
2, 97990
2, Max-Planck-Str.
[email protected])
97990 Weikersheim,
Weikersheim, Germany
[email protected]) Germany39, 74081
Heilbronn,
∗∗ CeraCon
Heilbronn, Germany
GmbH,
Germany (e-mail:
Talstraße
(e-mail: [email protected])
2, 97990 Weikersheim,
[email protected]) Germany
∗∗ CeraCon GmbH, Talstraße 2, 97990
∗∗ CeraCon GmbH, Talstraße 2, 97990 Weikersheim, Germany Weikersheim, Germany
CeraCon GmbH,
∗∗ CeraCon GmbH, Talstraße
Talstraße 2, 2, 97990
97990 Weikersheim,
Weikersheim, Germany Germany
Abstract:
Abstract: Reducing
Reducing costs
costs is
is an
an important
important part
part in
in todays
todays buisness.
buisness. Therefore
Therefore manufacturers
manufacturers try
try
Abstract:
Abstract:
to reduce Reducing
Reducing
unnecessary costs
costs
work is
is an
an important
important
processes and part
part
storage in
in todays
todays
costs. buisness.
buisness.
Machine Therefore
Therefore
maintenance manufacturers
manufacturers
is a big, complex, try
try
to reduce
Abstract:
Abstract:
to reduce unnecessary
Reducing
Reducing
unnecessary work
costs
costs
work processes
is
is an
an and
important
important
processes and storage
part
part
storage in
in costs.
todays
todays
costs. Machine
buisness.
buisness.
Machine maintenance
Therefore
Therefore
maintenance is a big,
manufacturers
manufacturers
is a big, complex,
complex, try
try
Abstract:
to reduce
regular Reducing
unnecessary
process. In costs
work
addition, is an
the important
processes and
spare part
storage
parts in todays
costs.
required buisness.
Machine
for this Therefore
maintenance
must be kept manufacturers
isina big,
stock complex,
untiltrya
regular
to
to reduce
reduce
regular process.
unnecessary
unnecessary
process. In
In addition,
work
work
addition, the
processes
processes
the spare
and
and
spare parts
storage
storage
parts required
costs.
costs.
required for
Machine
Machine
for this
this must be
maintenance
maintenance
must be kept
kept is
isin
ina
a stock
big,
big,
stock until
complex,
complex,
until a
to reduce
regular process.
machine
machine
regular unnecessary
process.
fails. In
fails. In
In In
order
order work
addition, processes
to avoid
to
addition, avoid thea
the and
a production
production
spare storage
spare parts costs.
parts breakdown
required for
breakdown
required Machine
for this
in this
in maintenance
must of
the event
the event
must bean
of
be kept
an
kept is a
unexpected
unexpected
in big,
in stock until aaa
complex,
stock failure,
failure,
until
regular
machine
regular process.
fails.
process. In In
In addition,
order to
addition, avoid the
the a spare parts
production
spare required
breakdown
parts breakdown
required for
in
for this
the
this must
event
must be
of kept
an in
unexpected
bemachines.
kept stock
in stock until
failure,
until a a
machine
more
more and
and fails.
more
more In order to avoid
manufacturers a
rely production
on predictive in
maintenance theforevent
their of an unexpected This failure,
enables
machine
machine
more
machine and fails.
fails.
more
fails. Inmanufacturers
In
In
order
order to
manufacturers
order to avoidrely
to avoid
avoid
a
rely
a
on
on predictive
a production
production
predictive
production
maintenance
breakdown
breakdown in
in the
maintenance
breakdown in thefor
the for their
event
event
their
event ofmachines.
of
of
an
an unexpected
an unexpected
machines.
unexpected
This
This enables
failure,
failure,
enables
failure,
more
more and
precise
precise
and more
more manufacturers
planning
planning of
of
manufacturers necessary
necessaryrely
rely on on predictive
maintenance
maintenance
on predictive and maintenance
and repair
repair
maintenance for
work,
work, their
as well
as well
for their
their wellmachines.
as a precise
as aa precise
machines. preciseThis enables
ordering
ordering
Thisordering
enables
more
more and
precise
and more
more manufacturers
planning of
manufacturers necessaryrely
rely on predictive
maintenance
predictive maintenance
and repair
maintenance for
work, as
forwell
their machines.
as
machines. This enables
Thisordering
enables
of
of
morethe
the precise
spare
spare
precise planning
parts
parts
planning of
required
required
of necessary
for
for
necessary this.
this. maintenance
A
A large
large
maintenance and
amount
amount
and repair
of
of past
past
repair work,
as
as
work, as
well
as well
as
as
well as
current
current
as aa precise
information
information
precise ordering is
is
more
of
morethe precise
spare
precise planning
parts
planning of
required
of necessary
for
necessary this. maintenance
A large
maintenance and
amount
and repair
of past
repair work,
as
work, as
well
as well
as
well as
current
as a
a precise ordering
information
precise ordering is
of the
required
required
of the spare
to
spare parts
to create
create
parts required
such
such a
required for
a predictive
predictive
for this.
this. A large
forecast
forecast
A large amount
about
about
amount of past
machines.
machines.
of past as
With
With
as well
well the
theas
as current
classification
classification
current information
of motors
of
information motors is
is
of
of the
required
the spare
to
spare parts
create
parts required
such a
required for
predictive
for this.
this. A
A large
forecast
large amount
about
amount of past
machines.
of past as
With
as well
well theas
as current
classification
current information
of
information motors is
is
required
based
based
required on
on to
to create
vibration,
vibration,
create such
this
this
such a
a predictive
paper
paper
predictive deals
deals forecast
with
with
forecast about
the
the
about machines.
implementation
implementation
machines. With
Withof
of the classification
predictive
predictive
the maintenance
maintenance
classification of
of motors
motorsfor
for
required
based
required on tovibration,
to create such
create such
this a paper
a predictive
predictive deals forecast
with
forecast about
the
about machines. With
implementation
machines. Withof the classification
classification
predictive
the maintenance of motors
of motorsfor
based
thermal
thermal on vibration,
systems.
systems. this
There
There paper
is an
ispaper
an deals
overview
overview with
of
of the implementation
suitable
suitable sensors
sensors and
and of
datapredictive
processing maintenance
methods, for
as
based
based
thermal
based
on
on
on
vibration,
vibration,
systems.
vibration,
this
this
There
this is paper
an deals
deals with
deals with
overview of the
the implementation
implementation
suitable sensors and ofdata
of
data processing
predictive
predictive
processing methods,
maintenance
maintenance
methods, as
for
for
as
thermal
well
well
thermal
thermal
well
as
as
as
systems.
various
various
systems.
systems.
various
There
classification
classification
There
There
classification ispaper
is
is
an
an
an
overview
algorithms.
algorithms.
overview
overview
algorithms.
with
of
In
of
In
the
of the
In the implementation
suitable
end,
end,
suitable
suitable
the end,
sensors
the
the best
best
sensors
sensors
the best and ofdata
and predictive
sensor-algorithm
sensor-algorithm
and data
data
sensor-algorithm
processing
processing
processing
maintenance
methods,
combinations
methods, are
combinations
methods,
combinations
for
as
are
as
as
are
thermal
well
shown.
shown.
well as
as systems.
various
various There
classification
classification is an overview
algorithms.
algorithms. of
In
In suitable
the
the end,
end, sensors
the
the best
best and data
sensor-algorithm
sensor-algorithm processing methods,
combinations
combinations as
are
are
well
shown.
well as
as various
various classification
classification algorithms.
algorithms. In
In the
the end,
end, the
the best
best sensor-algorithm
sensor-algorithm combinations
combinations are
are
shown.
Copyright © 2021 The Authors. This is an open access article under the CC BY-NC-ND license
shown.
shown.
shown.
(http://creativecommons.org/
Keywords:
Keywords:
Keywords:
Predictive
Predictive
Predictive licenses/by-nc-nd/4.0)
Maintenance;
Maintenance;
Maintenance;
Industry
Industry
Industry
4.0;
4.0;
4.0;
Internet
Internet
Internet
of
of
of
Things;
Things;
Things;
Big
Big
Big
Data;
Data;
Data;
Industrial
Industrial
Industrial
Keywords:
Internet.
Internet.
Keywords: Predictive
Predictive Maintenance;
Maintenance; Industry
Industry 4.0;
4.0; Internet
Internet of
of Things;
Things; Big
Big Data;
Data; Industrial
Industrial
Keywords:
Internet.
Keywords: Predictive
Predictive Maintenance;
Maintenance; Industry
Industry 4.0;
4.0; Internet
Internet of
of Things;
Things; Big
Big Data;
Data; Industrial
Industrial
Internet.
Internet.
Internet.
Internet.1. INTRODUCTION
1. INTRODUCTION cal
cal risk
risk analysis
analysis methods
methods (FMEA,
(FMEA, risk
risk graph,
graph, fault
fault tree
tree
1.
1. INTRODUCTION
INTRODUCTION cal
cal risk
risk
analysis) analysis
analysis
(3). CDC methods
methods
is the (FMEA,
(FMEA,
assignment risk
risk
of a graph,
graph,
wear fault
fault
component tree
tree
1.
1. INTRODUCTION
INTRODUCTION analysis)
cal
cal risk
risk
analysis) (3).
analysis
analysis
(3). CDC
CDC is
methods
methods
is the
the assignment
(FMEA,
(FMEA,
assignment of
risk
risk
of a
a wear
graph,
graph,
wear component
fault
fault
component tree
tree
The topic
topic of 1.
of predictive INTRODUCTION
predictive maintenance
maintenance (PMA) (PMA) is is becoming
becoming of cal
of risk
analysis)
the analysis
(3).
industrial CDC methods
is
system the to (FMEA,
assignment
a potentially risk
of a graph,
wear
occurring fault
component tree
defect.
The the
analysis)
analysis)
of the industrial
(3).
(3).
industrial CDC
CDC system
is
is
system the
the to a potentially
assignment
assignment
to a potentially of
of a
a occurring
wear
wear
occurring defect.
component
component
defect.
The
The
more topic
topic
and of
of
morepredictive
predictive
important maintenance
maintenance
for industrial (PMA)
(PMA)
plants is
is becoming
becoming
and is the analysis)
of the
Depending (3).
industrial
on CDC
the is
system
numberthe assignment
toof a potentially
possible of a
defects,wear
occurring
a component
defect.
component
more
The and more
topic
The topic
more topic
and more
of important
predictive
of predictive
predictive
important for industrial
industrial
maintenance
maintenance
for plantsis
(PMA)
(PMA)
plants and
is and
becoming
is the Depending
is the
becoming of
of the
the
Depending
of the
on
industrial
industrial
on
industrial
the
the number
system
system
number
system
to
toofaapossible
toof
potentially
potentially
apossible
potentially
defects, aa component
occurring
occurring
defects,
occurring
defect.
defect.
component
defect.
The
more
key and
topic of
more
in important
mechanical maintenance
for industrial
engineering (PMA)
from plants
the is becoming
and
Industry is the
4.0 Depending
can
can therefore
therefore
Depending on
on the
have
have
the number
several
several
number of possible
CDCs.
CDCs.
of possible Each
Each defects,
CDC
CDC
defects, aawas
wascomponent
assigned
assigned
component
key
more
more
key topic
and
and
topic in mechanical
more
more
in important
important
mechanical engineering
for
for from
industrial
industrial
engineering from the
plants
plants
the Industry
and
and
Industry is
is 4.0
the
the
4.0 Depending
can therefore on the
have number
several of possible
CDCs. Each defects,
CDC awascomponent
assigned
more
key and more important for industrial plants and is the Depending
can on
therefore have
an potential
potential the
have number
several
detection of possible
CDCs. Each
measure, Each defects,
CDC was
e.g. physical
physical a component
wasvibration
assigned
vibration
key topic
aspect
aspect
topic(1).
(1).in mechanical
PMA
PMA
in PMA
mechanicalis engineering
defined
is defined
defined as from
from the
condition-based
as condition-based
engineering condition-basedthe Industry mainte-
mainte-
Industry 4.0
4.0 an can
can therefore
therefore detection
several measure,
CDCs. e.g. CDC assigned
key
aspect
key topic
topic
aspect
nance (1).
(1).
which
in
in mechanical
mechanical
PMA
is carried
carriedis engineering
is defined
defined
out on theas
engineering from
from
as condition-based
the condition-based
basis of a
the
athe
wear
Industry
ormainte-
Industry
mainte-
4.0
4.0
service
an
can
an therefore have
potential
potential
measurement
measurement
an potential have several
detection
several
detection
or
or electrical
electrical
detection
CDCs.
measure,
CDCs.
measure,
current
current
measure,
Each
e.g.
Each
e.g.
CDC
CDC was
physical
physical
measurement.
measurement.
e.g. physical
assigned
wasvibration
assigned
vibration
Suitable
Suitable
vibration
nance
aspect
aspect
nance which
(1).
(1).
which is
PMA
PMA
is carriedis
is out
defined
out on
on as
as
the basis of
condition-based
basis of a wear
wear or
or service
mainte-
mainte-
service an potential
measurement detection
or electrical measure,
current e.g. physical
measurement. vibration
Suitable
aspect
nance
life (1). (2).
which
forecast PMA
is carriedis defined
out on as condition-based
the basis of a wear ormainte-
service an potential
measurement
sensors
sensors were
were
measurement detection
or electrical
selected
selected
or for
for
electrical measure,
current
the
the analyzed
analyzed
current e.g. physical
measurement.
detection
detection
measurement. vibration
Suitable
measures
measures
Suitable
life
nance
nance
life forecast
which
which
forecast (2).
is
is
(2).carried
carried out
out on
on the
the basis
basis of
of a
a wear
wear or
or service
service measurement
sensors were
measurement or electrical
selected
or for
electrical current
the analyzed
current measurement.
detection
measurement. Suitable
measures
Suitable
nance
life which (2).
forecast is carried out on the basis of a wear or service sensors and
and
sensors were
analyzed
analyzed
were selected
with
with
selected for
regard
regard
for the
the to
to analyzed
the
the
analyzed PMA
PMA detection
strategy.
strategy.
detection measures
CDC’s
CDC’s
measures
life
life forecast
forecast (2).
(2). sensors
and were
analyzed selected
with for
regard the to analyzed
the PMA detection
strategy. measures
life forecast (2). sensors
with
with
and were
and analyzed
analyzed
the same
the selected
with regard
same detection
detection
with for
regard the
methods
methods
to analyzed
to thethewerePMA
were
PMA detection
strategy.
combined
combined
strategy. andCDC’s
measures
and CDC’s
mea-
mea-
CDC’s
PMA uses methods that allow for individual maintenance and
with analyzed
the same with
detectionregard to
methods the PMA
were strategy.
combined and CDC’s
mea-
PMA usesuses methods
methods that that allow
allow forfor individual maintenance and
individual maintenance with analyzed
the
surement
surement
with the same
data
data
same with
detection
recorded
recorded
detectionregard with
with tothe
methods
methods thewere
the PMA
respective
respective
were strategy.
combined
combined sensors.
sensors. and
and CDC’s
mea-
PMA
PMA uses
intervals ofmethods
an that
industrial allow
plantfor individual
to be maintenance
determined and with the
with the same
surement same
data detection
recorded
detection methods
with
methods the were
were combined
respective
combined sensors. and mea-
and mea-
mea-
intervals
PMA
PMA
intervals uses
uses of an
methods
methods industrial
that
that plant
allow
allow for
for to be
individual
individual determined
maintenance
maintenance and surement
surement data
data recorded
recorded with
with the
the respective
respective sensors.
sensors.
PMA usesof
intervals
the of
maintenance an
methods
an industrial
that allow
industrial
process plant
toplant
be to
to be
for initiated
individual
be determined
maintenance
determined
automatically. and
and surement
The
The core
core
surement data
of
of
data this
thisrecorded
work
work
recorded is
is with
the
the
with the respective
evaluation
evaluation
the respective of
of a
asensors.
combination
combination
sensors.
the maintenance
intervals
intervals
the of an
of
maintenance process to
an industrial
industrial
process toplant
be initiated
plant
be initiated
to be
to automatically.
be determined
determined
automatically. and The core of this work is the evaluation of a combination
and
intervals
the part
As ofofa
a an
maintenance
part R&D industrial
process
cooperation plant
to to
beproject bebetween
initiated
project determined
automatically.
CeraConand The of
of core
detectionof this work
measures isforthedataevaluation
processing of amethods
combination and
AsAs
the
the
the part of
maintenance
maintenance
of a
maintenance
R&D
R&D cooperation
process
process
cooperation
process
to
to be
be initiated
initiated
project
to beproject
initiated
between CeraCon
automatically.
automatically.
between CeraCon
automatically. Thedetection
The
of
The
core
core of
detection
core of this
of
measures
this work
work is
measures
this work isfor
for
is
the
the
the
data
data processing
evaluation
evaluation
processing
evaluation of aaamethods
of
of
combination
combination
methods
combination
and
and
As
GmbH
GmbH
As part
part of
and
and
of a
a R&D
the
the
R&D cooperation
Heilbronn
Heilbronn
cooperationUniversity
University
project of
of between
Applied
Applied
between CeraCon
Sciences,
Sciences,
CeraCon of
ML
ML
of detection
algorithms.
algorithms.
detection measures
The
The
measures for
optimal
optimal
for data
data processing
combination
combination
processing methods
of
of
methods these
these and
is
is
anda
a
As
GmbH
As part of
and
partand a R&D
the
ofsystem
athe
R&D cooperation
Heilbronn
cooperation project
University
project of between
Applied
between CeraCon
Sciences,
CeraCon of
ML
of detection
algorithms.
detection measures
The
measures for
optimal
for data
data processing
combination
processing methods
of
methods these and
is
anda
GmbH
a
a thermal
thermal
GmbH and system
the Heilbronn
is
is to
to
Heilbronn be
be University
set
set up
up
Universityunder
under of
of Applied
automation
automation
Applied Sciences,
and
and
Sciences, a
a ML
ML algorithms.
prerequisite
prerequisite
algorithms. for
for an
anThe
The optimal
efficient
efficient
optimal PMA
PMA combination
strategy
strategy
combination that
that of
of these
can
can be
be
these is
used
used
is a
a
GmbH
aGmbHthermaland
and the
system
the Heilbronn
is to
Heilbronn be University
set up
Universityunder of
of Applied
automation
Applied Sciences,
and
Sciences, a ML
ML algorithms.
prerequisite
algorithms. for anThe
The optimal
efficient
optimal PMA combination
strategy
combination that of
of these
can be
these is
used
is a
a
a
PMA
PMA
a thermal system
strategy
strategy
thermal system is
is is
to
to
is to
be
be
to be
be set up
implemented,
implemented,
set up under
under automation
which
which
automationshould
should and
then
then
and a
a prerequisite
for
for the
the for
respective
respective
prerequisite for an
an efficient
industrial
industrial
efficient PMA
PMAplant.
plant. strategy
strategy that
that can
can be
be used
used
a
PMA
abe thermal
thermal system
strategy
system is is
to
is to
be
to be
be set up
implemented,
set up under
under automation
which
automationshould and
then
and a
a prerequisite
for the
prerequisite for
respective
for an
an efficient
industrial
efficient PMA
PMAplant. strategy
strategy that
that can
can be
be used
used
PMA strategy to
adaptable is other
tois to be
other be industrial
implemented,
industrial plants which 1 should
1 .. should
1 Due to then
tothen
the for for the respective industrial plant.
bebe
PMA
PMA
PMA
adaptable
strategy
strategy tois
adaptable
strategy
to
is other
to be
to
implemented,
be industrial
implemented,
implemented,
plants
plants which
which
which 1
Due
should
. should
Due to the
then
the
then for the
for the respective
the respective industrial
respective industrial plant.
industrial plant.
plant.
be adaptable
complexity
complexity
be adaptable of
of to other
industrial
industrial
to other industrial
plants,
plants,
industrial an
an plants
intelligent
intelligent
plants 1 .. Due
1 to
solution
solution
Due to the
is
is
the 1.1
1.1 State
State of
of the
the Art
Art
be adaptable
be adaptable
complexity of to
to other industrial
industrial
other industrial
plants, an plants
intelligent
plants 1 .. Due
Due to the
solution
to the
is 1.1 State of the Art
complexity
required
required
complexity in
in of
order
order
of industrial
to
to be
be
industrial able
ableplants,
to
to
plants, an
offer
offer
an intelligent
individual
individual
intelligent solution
maintenance
maintenance
solution is
is 1.1
1.1 State
State of
of the
the Art
Art
complexity
required
complexity in of
order
of industrial
to be
industrial ableplants,
to
plants, an
offer
an intelligent
individual
intelligent solution
maintenance
solution is
is 1.1
1.1 State
State of
of the
the Art
Art
required
strategies
strategies
required in
in order
depending
depending
order to
to be
be able
on
on
ablethe
theto
to offer
state
state
offer individual
of
of the
the
individual maintenance
plant.
plant. For
For
maintenance this
this A
A study
study by
by Bearingpoint
Bearingpoint (4)
(4) shows
shows that
that PMA
PMA implemen-
implemen-
required
strategies
required in
in order
depending
order to
to be
be able
on
abletheto
to offer
state
offer individual
of the
individual maintenance
plant. For
maintenance this A study by Bearingpoint (4) shows that PMA implemen-
strategies
reason, the
reason,
strategies depending
thedepending
project uses
project useson
on the
machine
machine
the state
state of
learning
learning
of the
the plant.
(ML)
(ML)
plant. For
methods.
methods.
For this
this A study
tations
tations
A study by
capture
capture
by Bearingpoint
76%
76%
Bearingpoint of
of (4)
the
the
(4) shows
relevant
relevant
shows that
thatdata
dataPMA
PMA using
using implemen-
suitable
suitable
implemen-
strategies
reason,
strategies the depending
project
depending useson
on the
machine
the state
state of
learning
of the
the plant.
(ML)
plant. For
methods.
For this
this A study
tations by
capture Bearingpoint
76% of (4)
the shows
relevant thatdataPMA using implemen-
suitable
reason,
reason, the
the project
project uses
uses machine
machine learning
learning (ML)
(ML) methods.
methods. A study
tations
sensors,
sensors,
tations by
capture Bearingpoint
although
although
capture 76%
76%only
only of
of 59%
59% (4)
the
the shows
relevant
of
of the
the
relevant thatdata
process,
process,
dataPMA using implemen-
measurement
measurement
using suitable
suitable
reason, the
reason, the project
project usesuses machine
machine learning
learning (ML) (ML) methods.
methods. tations
sensors, capture
although 76%only of the
59% relevant
of the data
process, using
measurementsuitable
tations
sensors,
and
and capture
although
machine
machine
sensors, although 76%
data
data only
are
are
only of the
59%
evaluated
evaluated
59% relevant
of
of the
the in
in a
a data
process,
process, using
measurement
targeted
targeted
measurementsuitable
manner.
manner.
The essential
The essential steps
steps of of an
an intelligent
intelligent PMA PMA strategystrategy are are and sensors, although
machine data only
are 59% of
evaluated the inprocess,
a measurement
targeted manner.
The essential steps of an intelligent PMA strategy are sensors,
and
There
There
and are
arealthough
machine
machinethree
three data only
basic
basic
data are
are 59%
approaches
approaches of
evaluated
evaluated the in
to
to
inprocess,
a
implementing
aa targeted measurement
targeted
implementing manner.
a
a PMA
PMA
manner.
The
the essential
digital steps
acquisition of of
an intelligent
(sensor) data, PMA their strategy
evaluation,are
the
The digital
The
the
The
essential
essential
digital
essential
acquisition
steps
steps of
acquisition
steps of of
of an(sensor)
an
of
an
intelligent
intelligent
(sensor)
intelligent
data,
data, PMA
PMA
PMA
their
their evaluation,
strategy
strategy
evaluation,
strategy are and
are
are
There
and
There
machine
strategy
strategy
There
are
machine
are
are
three
three
(5).
(5).
three A
A
data
basic
data
basic
basic
are evaluated
areapproaches
evaluated
approaches
approach
approach
approaches is
is
in
to
in
to
toto
to
ausetargeted
implementing
targeted
implementing
use the
the
implementing
manner.
already
already
a
a
a
PMA
manner.
PMA
PMAim-
im-
the
the digital
analysis
analysis
digital acquisition
of
of the
the
acquisition of
acquired
acquired
of (sensor)
data
data
(sensor) data,
and
and
data, their
the
the
their evaluation,
prediction
prediction
evaluation, of
of There
strategy are three
(5). A basic approaches
approach istotoimplementing
use the alreadya PMAim-
the
the digital
analysis
digital acquisition
of the
acquisition of
acquired
of (sensor)
data
(sensor) data,
and
data, their
the
their evaluation,
prediction
evaluation, of There
strategy
plemented
plemented
strategy are three
(5).
(5). A basic
sensors
sensors
A basic of
of approaches
approach
the
the
approach plant
plant is
istoto
for
for
toimplementing
use
process
process
use the
the alreadya PMA
monitoring.
monitoring.
already im-
im-
analysis
probable
probable
the of the
events.
events.
analysis of the acquired
acquired data data andand the the prediction
prediction of of plemented
strategy (5). A
sensorsbasic of approach
the plant is to
for use
process the already
monitoring. im-
the analysis
probable
the of
events.
analysis of the
the acquired
acquired data
data and
and the
the prediction
prediction of
of strategy
plemented
This
This (5).
passive
passive
plemented A
sensorsbasic
method
method
sensors of
of approach
the
is
is
the plant
particularly
particularly
plant is to
for
for use
process
suitable
suitable
process the already
monitoring.
for
for im-
systems
systems
monitoring.
probable
probable events.
events. plemented
This passive sensors
method of the
is plant
particularly for process
suitable monitoring.
for systems
probable events.
probable events. plemented
This
that
that
This passive
are
are sensors
already
already
passive method
methodin
in of the
is
operation.
operation.
is plant
particularly
particularly for
Another
Another process
suitable
passive
passive
suitable monitoring.
for
for systems
approach
approach
systems
This
that passive
are already methodin is
is particularly
operation. Another suitable
passive for systems
forapproach
First, possible
First, possible component
component defect defect combinations
combinations (CDC) (CDC) of of Thisthat
is
is
thatto
to
passive
are already
introduce
introduce
are already
methodin
dedicated
dedicated
in
particularly
operation.
operation. sensors
sensors Another
Another
suitable
into
into passive
the
the
passive system.
system.
systems
approach
The
The
approach
First, possible component defect combinations (CDC) of that
is
thatto are already
introduce
are already in operation.
dedicated
in operation. sensors Another
Anotherinto passive
the
passive approach
system. The
approach
First,
the
the possible
industrial
industrial
First, possible component
plant
plant were
were
component defect
analyzed
analyzed
defect combinations
using
using
combinationsstandard
standard (CDC)
techni-
techni-
(CDC) of
of is to introduce
additional
additional
is to sensors
sensors
introduce dedicated
are
are
dedicated sensors
introduced
introduced
sensors to
to into
monitor
monitor
into the
the system.
defined
defined
system. The
wear
wear
The
First,
the
First, possible
industrial
possible component
plant were
component defect
analyzed
defect combinations
using
combinationsstandard (CDC)
techni-
(CDC) of
of is to introduce
additional sensors dedicated
are sensors
introduced to into
monitorthe system.
defined The
wear
the industrial
the industrial
industrial plantplant were
plant were analyzed
were analyzed
analyzed usingusing standard
using standard techni-
standard techni- is
techni- components to introduce
additional
components
additional sensors
and
and
sensors dedicated
to
toare
aredetect sensors
introduced
detect potential
potential
introduced to
to into
monitorthe
defects.
defects.
monitor system.
defined
In
In the
the
defined The
wear
third
third
wear
the
the industrial plant were analyzed using standard techni- additional sensors
components
additional sensors
and toare
are introduced
detect potential
introduced to monitor
to monitor
defects. defined
In the
defined wear
third
wear
1
1 Sponsored
1 Sponsored
by
by ZIM ZF4644801.
ZIM ZF4644801. components
approach,
approach,
components a and
test
aa test
and to
signal
signal
to detect
is
is
detect potential
actively
actively
potential fed
fed defects.
into
into
defects.the
the In the
system.
system.
In the third
The
The
third
components
approach, and to
to detect potential defects. In the
the third
1 Sponsored
1 Sponsored
by
by ZIM
ZIM ZF4644801.
ZF4644801. components
approach,
approach, aa test
and signal
test
test signal
signal
is
detect
is
is
actively
potential
actively
actively
fed
fed
fed
into
defects.
into
into
the
the
the
system.
In
system.
system.
The
third
The
The
1 Sponsored
Sponsored by by ZIM
ZIM ZF4644801. approach, a test signal is actively fed into the system. The
1 ZF4644801.
2405-8963 Copyright © 2021 The Authors. This is an open access article under the CC BY-NC-ND license.
Sponsored by ZIM ZF4644801. approach, a test signal is actively fed into the system. The
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2021.08.126
1060 Christoph Kammerer et al. / IFAC PapersOnLine 54-1 (2021) 1059–1064

degree of wear of the components to be monitored can be · MPU 60.50

deduced from the feedback. An example of this is Time • Three-axis magnetic field sensor (magnetometer)
Domain Reflectometry (TDR) (5). In this paper it was · MLX 90393
evaluated if vibration sensors can be used to classify the • Multi sensors with three-axis acceleration, three-axis
rotation of a fan motor for a later implementation in a yaw rate and three-axis magnetic field measurement
PMA strategy. · MPU 92.65, BNO 055, GY 250, GY 521

2. DATA COLLECTION These recorded the acceleration, the rotation rate and the
surrounding magnetic field of the fan motor
2.1 Sensor Resolution R3G180-AJ11-XF from ebm-papst Mulfingen GmbH &
Co. KG used in the thermal system. Figure 2 shows the
When buying industrial sensors, you often have to commit measurement set-up with the selected three-axis accelera-
to a sensor resolution. This requires that you have a tion sensors (6).
basic understanding of what accelerations occur on the The fan motor was operated at fixed speeds, which were
component. For this purpose, the effects were previously divided into 7 classes. This classification was based on the
considered in an experiment when an accelerometer with specific values 0%, 50%, 60%, 70%, 80%, 90% and 100%
an insufficient resolution is used. In this case, the sensor of the maximum engine speed. During the operation of
generates vibrations that exceed the sensor resolution. the fan motor the vibration of the crankcase was sensed
A CDC of the fan motor is that the fan wheel has an and recorded by the sensors. More than 980,000 structured
imbalance. This fault situation was simulated by attaching sensor data sets per measurement series and sensor type
an unbalance to the fan blade. were recorded. A total of more than 2.6 million data sets
The result of this simulation is shown in figure 1 (a). There have thus been recorded for all sensor types. A section of
are shown the measured accelleration values in x- and y- a full data set is shown in Table 1.
axis of an accelleration sensor with a maximum resolution An example of a recorded data set is shown in Figure
of ±2G. The red values show the vibrations of the motor 3. It shows the measured acceleration from the housing
without an imbalance and the blue values show the vibra- vibration in the spatial x- and z- orientation. The individ-
tions which occurs with an imbalance. It can be clearly ual classes are highlighted in color to make a distinction
seen that the vibrations on the motor increased due to possible. Due to the highest spatial coverage, it can be
the imbalance. It can also be seen that vibrations that go seen that the measurement results for class 80% can be
beyond the set sensor resolution of ±2G were not recorded assigned to the resonance range of the fan motor, since
correctly. They are in line with the maximum accelera- the acceleration values in the x- and z- alignment are at
tion of ±2G. The measured values that did not exceed their maximum values here.
the maximum resolution were not affected by this. The
experiment shows that the CDC ”imbalance” can cause
very strong vibrations. The vibrations are so strong that 3. DATA CONDITIONING
they exceed a sensor resolution of ±2G. If a sensor is used
that can only record values up to a resolution/acceleration In order to be able to better differentiate the individual
of ±2G, these are recorded incorrectly. The values that classes, it is in some cases advantageous if the data records
exceed the maximum resolution are then incorrectly saved are processed before classification. The methods used for
in the data record (6). To prevent such problems, it is data conditioning are presented here:
important to see how large the vibrations can be. The
One possibility to process the data sets consists of the
sensor resolution should have at least this value with a
differencing and absolute value formation of subsequent
safety buffer. In figure 1 (b), instead of the resolution
values according to equation 1.
of ±2G, the double resolution of ±4G was chosen for
the same motor level. In the picture you can see that
no ”lines” have formed and therefore the vibrations were Xi = |Xi − Xi+1 | (1)
not greater than the sensor resolution. The resolution of
±4G is therefore much more suitable than the resolution of Xi and Xi+1 are the successive sensor values. Another
±2G. The experiment has shown that a correctly selected processing method is the integration of the data according
sensor resolution is a prerequisite for obtaining meaningful to equation 2. Here the area under two successive values
results. If the vibrations are greater than the resolution of Xi and Xi+1 is calculated.
the sensor, the incorrectly stored measured values cannot
be classified correctly (6). 
Xi + 0.5 · (Xi+1 − Xi ) if Xi < Xi+1
2.2 Sensors and Test Set-Up Xi = Xi − 0.5 · (Xi − Xi+1 ) if Xi > Xi+1 (2)

Xi if Xi = Xi+1
The requirement for a condition-based PMA is a struc-
tured data collection of sensor values. The following sen- In both the processing methods, an additional smoothing
sors were used to obtain status data: can be carried out by calculating the moving average
• Three-axis acceleration sensors (Accelerometer): according to equation 3.
· LIS 3DH, MMA 8451, ADXL 343, ADXL 345 i+g
1 
• Three-axis acceleration sensors with three-axis yaw Xi = · Xi (3)
rate sensor (gyroskope) G i−g
 Christoph Kammerer et al. / IFAC PapersOnLine 54-1 (2021) 1059–1064 1061

Fig. 1. Comparison of an unbalanced fan with a resolution of ±2G (a) and a resolution of ±4G (b).
Table 1. A Section of the Data measured with the MPU 60.50 Gyroskope.
60.50 AccelX 60.50 AccelY 60.50 AccelZ 60.50 GyroX 60.50 GyroY 60.50 GyroZ Target
416 6088 60146 2709 62478 65187 Motor 000
426 6026 60148 2733 62469 65185 Motor 000
404 6110 60146 2727 62478 65192 Motor 000
470 6046 60140 2720 62486 65188 Motor 000

Fig. 2. Measurement setup with the three-axis acceleration Fig. 3. Measurement results (raw data) from the accelera-
sensors LIS3DH, MMA 8451, ADXL 345 and ADXL tion sensor MMA 8451.
343.
not unique and can in principle be assigned to any class.
The parameter G specifies the degree of smoothing. The In Figure 4 (b) the classes are more delimited after the
parameter g is the difference between the indices between differencing and smoothing and thus a class assignment is
the instantaneous value Xi and the maximum value Xg±i clearer. For example, the value 0.7 can be clearly assigned
specified by the degree of smoothing. Thus g depends to the class shown in gray.
on the degree of smoothing G and can be determined
according to equation 4. 4. EVALUATION
G−1
g= (4) The evaluation of the ML algorithms with regard to the
2
respective sensors and the data processing was divided into
a training and a test phase. In the training phase, the
With the degree of smoothing G, first optimizations re-
data records were divided evenly by feeding every tenth
garding the classification of the measured sensor data can
data value of the respective training method to the ML
be carried out (7). Figure 4 shows the effect of processing
algorithm. As a result, the respective ML algorithm was
by means of differencing compared to the unprocessed raw
trained with 10% of the data. The complete data set was
data. The coloring in the pictures illustrates the different
then evaluated in the test phase. Several ML algorithms
class assignments.
were considered for the recorded data sets. These included
Figure 4 (a) shows the acquired raw data of an acceleration decision trees (8) (13), the gradient boost method (8) (14),
sensor in the x-orientation. It can be seen that a delimi- a focus cluster algorithm (9) and artificial neural networks
tation regarding the classes is not clear. For example, the (ANN) (10). The investigations revealed that ANNs are
acceleration in the direction of the x-axis at −8m/s2 is less suitable for data sets with low attribute numbers due
1062 Christoph Kammerer et al. / IFAC PapersOnLine 54-1 (2021) 1059–1064

Fig. 4. Comparison of raw data (a) and data prepared by differencing (b) (ADXL 345) .

to the long duration in the training phase. Therefore, only tables of the sensors which have achieved the best results
the decision trees, the gradient boost method and the focus of the sensor types examined are listed in Tables 3 to 5.
cluster algorithm were used for the further experiments These were the ADXL 345 (accelerometer), the MPU 60.50
(7). (gyroscope) and the GY 521 (accelerometer, gyroscope and
magnetic field). The processing method with the highest
Figure 5 shows two confusion matrices (11) for the focus
performance for the respective algorithm is shown for each
cluster algorithm, which show the distribution between the
of the sensors.
actual class and the class determined by the algorithm.
The numbers on the axes correspond to the seven defined The result of the examinations according to Table 3
classes in which the data records have been categorized. was that all acceleration sensors achieved the greatest
The darker an area, the more often the ML algorithm has performance with smoothing (G = 99) and data prepared
assigned data records to a class. A correct assignment is by differencing. The focus cluster algorithm achieved the
obtained if the assigned class corresponds to the actual highest performance.
class. Ideally, you would get a black diagonal from top left
to bottom right. The result according to Table 4 is that the highest per-
formance was achieved with unprocessed and unsmoothed
Figure 5 (b) shows the result for the data sets prepared data with the gyroscopes. The gradient boost process
after differencing and smoothing. It can be seen that the achieved the highest performance.
majority of the data records were assigned to the actual
classes, the hit rate here was over 98%. On the other hand, In the case of the multifunction sensors with acceleration,
it can be seen in Figure 5 (a) that a significantly lower hit magnetic field sensors and gyroscope, it can be seen from
rate has been achieved for the unprepared data sets. Table 5 that the highest performance was achieved with
smoothing (G = 99) and differenced data using the cluster
cluster algorithm.
5. RESULTS
6. CONCLUSION
To compare the results of the ML algorithms for the
respective sensors, a matrix with the relevant properties It has been found that the processing of the raw data in
was created for each combination of ML algorithm and the form of smoothing and differencing in combination
sensor: with the focus cluster algorithm gave the best results
• Classification accuracy (performance) of the algo- for acceleration sensors. The gyroscopes examined showed
rithms that the unprocessed raw data without smoothing in
• Computing time for the training and testing phase of combination with the gradient boost method achieved
the algorithms the highest classifiability. The multisensors examined gave
• Smoothing factor G the best results when using the focus cluster algorithm
in combination with smoothed and differenced data. The
Table 2 shows the comparison of the processing methods overall best results were achieved with the multisensor GY
examined using the combination of the multifunction sen- 521 and using the focus cluster algorithm and differenced
sor GY 521 and the gradient boost method. data (99.89%).
It can be seen that the highest performance is achieved
when using the raw data. In the training and test phases 7. OUTLOOK
the integrated and the differenced data are slightly faster.
A comparison was made for each sensor and ML algorithm Based on these results, the combination of detection mea-
combination. The best performing data processing method sure, data processing method and ML algorithm can in
was then selected for each combination. The comparison the next step be used for a PMA strategy. For a complete
 Christoph Kammerer et al. / IFAC PapersOnLine 54-1 (2021) 1059–1064 1063

Fig. 5. Comparison of the prediction results of the focus cluster algorithm with raw data (a) and data prepared by
differencing (b) (ADXL 345).
Table 2. Comparison of the data preparation methods with the combination of the multifunction
sensor GY521 and the gradient boost method.
Sensor GY 521 with Performance (%) Computing time training (s) Computing time testing (s)
Gradient Boost method G=0 G = 99 G=0 G = 99 G=0 G = 99
Raw data 98.51 83.97 4.85 6.88 1.11 0.9
Integration 82.84 85.15 5.45 3.82 1.07 0.66
Differencing 79.76 85.1 5.04 3.83 1.5 0.61

Table 3. Comparison of the ML algorithms on the combination ADXL 345 (accelerometer) and
the best performing data processing method.
ADXL 345 Decision tree Gradient boost method Focus cluster algorithm
Conditioning method Differencing, G = 99 Differencing, G = 99 Differencing, G = 99
Performance 93.79% 96.6% 98.98%
Computing time training 1.78s 5.88s 0.32s
Computing time testing 31.99s 1.44s 34.88s

Table 4. Comparison of the ML algorithms on the combination MPU 60.50 (gyroskope) and the
best performing data processing method.
MPU 60.50 Decision tree Gradient boost method Focus cluster algorithm
Conditioning method Raw data, G = 0 Raw data, G = 0 Differencing, G = 99
Performance 91.46% 95.24% 90.89%
Computing time training 0.09s 5.54s 0.91s
Computing time testing 0.28s 1.46s 31.99s

Table 5. Comparison of the ML algorithms on the combination GY 521 and the best performing
data processing method.
GY 521 Decision tree Gradient boost method Focus cluster algorithm
Conditioning method Raw data, G = 0 Raw data, G = 0 Differencing, G = 99
Performance 95.7% 98.51% 99.89%
Computing time training 0.09s 4.85s 1.27s
Computing time testing 0.28s 1.11s 37.63s

PMA, further detection measures have to be examined. probabilities of the individual assemblies. Such models
For that purpose, this procedure is continued with further could be based on Dynamic Bayesian Networks (DBN)
sensor types in order to find an optimal combination for all (12), auto regresssion and moving average (ARMA) (15)
necessary detection measures. In the future, a prediction or, as the focus cluster algorithm has yielded such an high
model is to be developed on the basis of these results, with performance, a multi dimensional focus trajectory.
which predictions can be made about the degree of wear of
system components of a thermal system under automation.
Formal aging and error models of the respective system REFERENCES
components must also be created in order to map the
aging process of components. These models can then be [1] S. Feldmann, O. Herweg, H. Rauen and P.-M. Synek.
used to make probabilistic statements about the failure (2017) ”Predictive Maintenance - Service der Zukunft.”
Roland Berger GmbH.
1064 Christoph Kammerer et al. / IFAC PapersOnLine 54-1 (2021) 1059–1064

[2] ”DIN EN 13306: Instandhaltung – Begriffe der In-

standhaltung; Dreisprachige Fassung EN 13306:2017
(Deutsche Norm),” Beuth Verlag, Berlin, 2018.
[3] B. Ebert, Prozessoptimierung bei Industrie 4.0 durch
Risikoanalysen, Berlin, Heidelberg: Springer-Verlag,
2018.
[4] F. Duscheck, R. Blameuser und S. Gehrmann,
”https://www.bearingpoint.com,” BearingPoint
GmbH, 2018. [Online].
[5] H. M. Hashemian, ”State-of-the-Art Predictive Main-
tenance Techniques,” in IEEE Transactions On Instru-
mentation And Measurement Vol. 60, 2011.
[6] M. Küstner, ”Auswahl und Bewertung eines Sen-
sorkonzepts für die Implementierung einer Predictive
Maintenance zur Temperaturbehandlung unter Au-
tomation,” Bachelorthesis, Weikersheim, Künzelsau,
2019.
[7] M. Gaust, ”Entwurf und Implementierung von
maschinellen Lernalgorithmen zur Klassifikation von
Maschinendaten für eine vorausschauende Wartung
von Industrie-Thermosystemen,” Bachelorthesis, Weik-
ersheim, Künzelsau, 2019.
[8] Pedregosa, ”JMLR 12 Scikit-learn: Machine Learning
in Python,” Journal of Machine Learning Research,
2011.
[9] A. Géron, Hands-On Machine Learning with Scikit-
Learn and TensorFlow: Concepts, Tools, and Techniques
for Building Intelligent Systems, Boston: O’Reilly UK
Ltd., 2017.
[10] T. Rashid, Make Your Own Neural Network, North
Charleston, SC, USA, CreateSpace Independent Pub-
lishing Platform, 2016.
[11] M. Füllsack, ”Systems sciences at ISIS,” Institute
for Systems Sciences, Innovation and Sustainability Re-
search at the Karl-Franzens University Graz, Austria,
2013. [Online]. Available: http://systems-sciences.uni-
graz.at/etextbook/bigdata/confusionmatrix.html. [Zu-
griff am 28 11 2019].
[12] P. Dagum, A. Galper und E. Horvitz, ”Dynamic
Network Models for Forecasting,” Proceedings of the
Eighth Conference on Uncertainty in Artificial Intelli-
gence. AUAI Press., p. 41–48, 1992.
[13] Thomas T., P. Vijayaraghavan A., Emmanuel S.
(2020) Applications of Decision Trees. In: Machine
Learning Approaches in Cyber Security Analytics.
Springer, Singapore
[14] A. Binding, N. Dykeman and S. Pang, ”Machine
Learning Predictive Maintenance on Data in the Wild,”
2019 IEEE 5th World Forum on Internet of Things
(WF-IoT), Limerick, Ireland, 2019, pp. 507-512.
[15] Jinjiang Wang, Yuanyuan Liang, Yinghao Zheng,
Robert X.Gao and Fengli Zhang, ”An integrated fault
diagnosis and prognosis approach for predictive mainte-
nance of wind turbine bearing with limited samples” in
Renewable Energy Vol. 145 p. 642 - 650, 2020