Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Available
Available onlineonline
Available
Available online
online
at
at www.sciencedirect.com
at www.sciencedirect.com
at www.sciencedirect.com
www.sciencedirect.com

ScienceDirect
Procedia Manufacturing 00 (2019) 000–000
Procedia Manufacturing 00 (2019) 000–000
Procedia Manufacturing
Procedia Manufacturing
Procedia 51 00
Procedia Manufacturing
Manufacturing 00
(2019)
(2020) 000–000
330–336
00 (2019)
(2019) 000–000
000–000
www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
30th International Conference on Flexible Automation and Intelligent Manufacturing (FAIM2021)
30th International Conference on15-18
Flexible
JuneAutomation andGreece.
2021, Athens, Intelligent Manufacturing (FAIM2021)
30th
30th International
International Conference
Conference on Flexible
on15-18 JuneAutomation
Flexible Automation and Intelligent
Intelligent Manufacturing
andGreece.
2021, Athens, Manufacturing (FAIM2021)
(FAIM2021)
15-18 June 2021, Athens, Greece.
15-18 June 2021, Athens, Greece.
Deep
Deep learning
learning enabling
enabling quality
quality improvement
improvement in
in rotogravure
rotogravure manufacturing
manufacturing
Deep
Deep learning
learning enabling
enabling quality
quality improvement
improvement in
in rotogravure
rotogravure manufacturing
manufacturing
Daniel Schmidt , Roman Gevers , Jörg Schwiep , Joaquín Ordieres-Meré , Javier Villalba-Diezb,c,d
Daniel Schmidt
a,b
a,b
a
, Roman Gevers aa
a
a
b
b
, Javier Villalba-Diezb,c,d
a,b
a,b aa , Jörg Schwiepaaa , Joaquín Ordieres-Meréb b b,c,d
b , Javier Villalba-Diezb,c,d
Daniel Schmidt a,b , Romana Matthews
Gevers
Daniel Schmidt , Romanaa Matthews
Gevers ,
, Jörg
Jörg
International Schwiep
Schwiep
GmbH, ,
, Joaquín
Joaquín
Gutenbergstraße 1-3, Ordieres-Meré
Ordieres-Meré
48691 Vreden, Germany. , Javier Villalba-Diezb,c,d
b Escuela International GmbH, Gutenbergstraße 1-3, 48691 Vreden, Germany.
Matthews International GmbH, Gutenbergstraße 1-3, 48691 Vreden, Germany.
Técnica Superior deaaIngenieros
Matthews Industriales, Universidad Politécnica de Madrid. José Gutiérrez Abascal 2, 28006 Madrid. Spain.
bb Escuela
Técnica
b Escuela
Técnica Superior
Superior
c Fakultät de
de Matthews International
Ingenieros
Ingenieros International
Industriales,
Industriales,
GmbH,
GmbH, Gutenbergstraße
Gutenbergstraße
Universidad
Universidad Politécnica
Politécnica
1-3,
1-3,
de
de
48691
Madrid.
Madrid.
Vreden,
48691 José
Vreden,
José
Germany.
Germany.
Gutiérrez
Gutiérrez Abascal
Abascal 2,
2, 28006
28006 Madrid.
Madrid. Spain.
Spain.
b Escuela
Técnica Superiorfür
deManagement
Ingenieros und Vertrieb,
Industriales, Campus Schwäbisch-Hall,
Universidad Politécnica de Hochschule
Madrid. JoséHeilbronn,
Gutiérrez 74523, Germany.
d DepartmentTécnica
Escuela c
c Superior
Fakultät
Fakultät deManagement
für
für Ingenieros Industriales,
Management und
und Vertrieb,
Vertrieb,Universidad
Campus
Campus Politécnica de Madrid.
Schwäbisch-Hall,
Schwäbisch-Hall, Hochschule
Hochschule Gutiérrez Abascal
JoséHeilbronn,
Heilbronn, Abascal
74523,
74523,
2,
2, 28006
28006 Madrid.
Germany.
Germany. Madrid. Spain.
Spain.
of Artificial
c Intelligence,
c Fakultät
dd Department of Artificial für Universidad
Management und Politécnica
Vertrieb, de Madrid,
Campus Campus deHochschule
Schwäbisch-Hall, Montegancedo, 28660 74523,
Heilbronn, BoadillaGermany.
del Monte, Madrid, Spain.
d Department of Fakultät
Artificial für Management
Intelligence,
Intelligence, und Vertrieb,
Universidad
Universidad Campus
Politécnica
Politécnica de Schwäbisch-Hall,
de Madrid,
Madrid, Campus deHochschule
Campus de Heilbronn,
Montegancedo,
Montegancedo, 28660 74523,
28660 BoadillaGermany.
Boadilla del Monte,
del Monte, Madrid,
Madrid, Spain.
Spain.
d Department of Artificial Intelligence, Universidad Politécnica de Madrid, Campus de Montegancedo, 28660 Boadilla del Monte, Madrid, Spain.
Department of Artificial Intelligence, Universidad Politécnica de Madrid, Campus de Montegancedo, 28660 Boadilla del Monte, Madrid, Spain.

Abstract
Abstract
Abstract
Abstract
Abstract
Advances in Computer Vision have helped the manufacturing industry achieve superior quality norms with a minimal inspection time due to
Advances
Advances in
in Computer
Computer Vision
Vision have
have helped
helped the
the manufacturing
manufacturing industry
industry achieve
achieve superior
end ofquality
superior quality norms
norms with
with aa minimal inspection time due to
optical quality
Advances
Advances
optical in
quality
surveillance
in Computer
Computer Vision
Vision
surveillance
systems.
have
systems.
These inspections
have helped
helped
These the most often
the manufacturing
manufacturing
inspections most
take place
industry
industry
often take
at the
achieve
achieve
place at superior
superior
the end of
the value
quality
quality
the
chain,
norms
norms
value with
with
chain, aa minimal
insuring
insuringminimal
minimal
the
inspection
the quality
inspection
inspection
quality
time
standards
time
standards
due
time due
due
of
to
of the
to
to
the
optical
optical quality
manufactured
quality surveillance
pieces. The
surveillance systems.
downside
systems. These
to this
These inspections
approach
inspections is most
that
most often
defective
often take
take place
parts can
place at
at the
still
the end of
continue
end of the
the value
through
value chain,
the
chain, insuring
value chain.
insuring the quality
Wasting
the quality a standards
lot of of the
resources
standards of the
optical qualitypieces.
manufactured
manufactured surveillance
pieces. The
The systems. to
downside
downside These
to this
this inspections
approach
approach is
is most
that
that often take
defective
defective place
parts
parts can
canatstill
the continue
still end of thethrough
continue value chain,
through the
the insuring
value
value chain.
chain.the qualityaa standards
Wasting
Wasting lot
lot of
of of the
resources
resources
and increasingpieces.
manufactured the leadThetime. To avoidtothis
downside thisdrain, the machines
approach is in the value stream should only through
produce the
error-free parts or at leastadetects them. An
manufactured
and increasing pieces.
leadThe downside tothis
thisdrain,
approach is that
that defective
defective parts can
can still
partsstream still continue
continue through the value
value chain.
chain. Wasting
Wasting lot
lot of
leastadetectsof resources
resources
and
optical
and quality the
increasing
increasing the lead
inspection
the lead
time.
time. To
To avoid
system
time. To avoid this
at every
avoid this drain, the
the machines
production
drain, the machines
step would
machines
in
in the
in addvalue
the
the value
a high
value priceshould
stream
stream should only
onlythis
cost. For
should only
produce
produce
reason,
produce
error-free
existing parts
error-free
error-free parts or
or
sensors
parts or
at
at least
should
at least detects them.
them. An
detect unwanted
detects them. An
and increasing
optical
optical quality
quality the lead
inspection
inspectiontime. To
system
system avoid
at
at this
every
every drain, the
production
production machines
step
step would
wouldin the
add
add value
a
a stream
high
high price
price should
cost.
cost. only
For
For produce
this
this reason,
reason,error-free
existing
existing parts or
sensors
sensors at least
should
should detects
detect
detect them. An
An
unwanted
unwanted
states.
optical With structured
quality data,
inspection a
systemperson
at with
every specific
productiondomain
stepknowledge
would add could
a rate
high this.
price This
cost. is
For a tough
this task,
reason, as a lot
existing of unknown
sensors factors
should can
detect influence
unwanted
optical
states.
states. quality
With
With inspection
structured
structured system
data,
data, a
a at every
person
person with
with production
specific
specific stepknowledge
domain
domain would addcould
knowledge a highrate
could price
rate cost.
this.
this. This
ThisFor
is
is a
athis reason,
tough
tough task,
task, existing
as
as a
a lot
lot sensors
of
of unknown
unknown should detect
factors
factors can
can unwanted
influence
influence
each step.
states. WithTherefore,
structured this
data,paper proposes
aa person with steps todomain
improve quality incould
rotogravure manufacturing using deep learning. Further research will be
states.
each WithTherefore,
step. structured this
data,paperperson with specific
proposes specific domain knowledge
knowledge could rate
rate this. This
This is
is aa tough
this.manufacturing task,
task, as
toughusing as aa lot of
of unknown
lotlearning.
unknown factors
factors can
can influence
influence
each step.
conducted
each step. Therefore,
in the coming
Therefore, this
this paper
months
paper to expandsteps
proposes
proposes these to
steps
steps to improve
results.
to Thequality
improve
improve quality
proposed
quality
in
in
in
rotogravure
rotogravure
procedures will
rotogravure be applied tousing
manufacturing
manufacturing
deep
deep
live data
using deep oflearning. Further
Further
a rotogravure
learning. Further
research
research
manufacturing
research
will be
willsite
be
each step. in
conducted
conducted Therefore,
in the
the coming
comingthismonths
paper proposes
months to
to expand
expand steps
these
these to improve
results.
results. The
The quality
proposed
proposed in rotogravure
procedures
procedures manufacturing
will
will be
be applied
applied using
to
to live
live deepof
data
data learning.
of a
a Further
rotogravure
rotogravure research will
manufacturing
manufacturing
be
willsite
be
site
and the effectiveness
conducted in the of this
coming approach
months to will be
expand analysed.
these results. The proposed procedures will be applied to live data of a rotogravure manufacturing site
conducted
and
and the
the in the coming
effectiveness
effectiveness of
of months
this
this to
approach
approach expand
will
will these
be
be results.
analysed.
analysed. The proposed procedures will be applied to live data of a rotogravure manufacturing site
and the
and the effectiveness
effectiveness ofof this
this approach
approach will
will be
be analysed.
analysed.

©c 2020
2020TheThe Authors.
TheAuthors.
Authors. Published
Published by Elsevier
by ElsevierLtd. Ltd.

c 2020
cc 2020 The Authors. Published
Published by
by Elsevier
Elsevier Ltd.
Ltd.
This isisan open access article under the CC BY-NC-ND licenselicense
(http://creativecommons.org/licenses/by-nc-nd/4.0/)

 c 2020
This
This
This2020
is
is
The
The
an
an
Authors.
anopen
open
Authors.
open access
access
Published
access article
Published
article
article
by
under
under
Elsevier
under
by Elsevier
the
the CC
CC
Ltd.
theBY-NC-ND
CC
Ltd. BY-NC-ND
BY-NC-ND license
license
(https://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review
This is an
Peer-review under
openunder responsibility
access article of
under
responsibility the
the
of scientific
CC
the BY-NC-NDcommittee
scientific of
license
committee the FAIM 2020.
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
of the FAIM
This is an
Peer-review
Peer-review open
underaccess article
responsibility under
of the CC BY-NC-ND
scientific license
committee of 2020.2021.
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
the FAIM
Peer-review under
Peer-review under responsibility
responsibility of of the
the scientific
scientific committee
committee of of the
the FAIM
FAIM 2020.
2020.
Keywords: Quality improvement; Quality control and inspection; Jidoka; Industry2020.
under responsibility of the scientific committee of the FAIM 4.0; Machine learning
Keywords: Quality
Keywords: Quality improvement;
improvement; Quality
Quality control
control and
and inspection;
inspection; Jidoka;
Jidoka; Industry
Industry 4.0; Machine learning
Keywords: Quality
Keywords: Quality improvement;
improvement; Quality
Quality control
control and
and inspection;
inspection; Jidoka; Industry 4.0;
Jidoka; Industry 4.0; Machine
4.0; Machine learning
Machine learning
learning

1. Introduction formed at the end of production. Based on the generally high

1.
1. Introduction
Introduction formed at
formed the end of production. Based on the generally high
1.
1. Introduction
Introduction costs of at
formed
formed anthe
at
atan the
the
end
optical
end
end of
of
of
production.
quality controlBased
production.
production. system.
Based
Based
on
on the
the generally
Therefore,
the generally
onTherefore,generally
high
a defec-
high
high
costs
costs of
of an optical
optical quality
quality control
control system.
system. Therefore, aa defec-
defec-
Today’s manufacturing industry is increasingly subject to an tive
costs
costs part
of
of still
an
an travels
optical
optical along
quality
quality the value
control
control stream,
system.
system. and
Therefore,
Therefore, is remade
a
a or
defec-
defec-
Today’s manufacturing industry is increasingly subject totive
an
tive part
part still travels along the value stream, and is remade or
Today’s
international
Today’s manufacturing
competition, industry
Today’s manufacturing
manufacturing because is
industry
industry isof
is increasingly
falling transit
increasingly
increasingly
subject
subject
subject
to
and com-
to an
to an
tive
an part still
if possible
tive part still
still
travels
fixed.
travels
along
along the
Depending
travels along onvalue
the
the the defect
value
value
stream,
stream,
stream,andand the is
and
and is
remade
ispoint during
remade
remade
or
or
or
international competition, because of falling transit and if
com-
if possible fixed. Depending on the defect and the point during
international
munication
international
international
competition,
costs, and
competition,
competition,a fasterbecause
because
because
of
transportation
of falling
falling
of fallingof of transit
goods
transit
transit
and
[1].
and
and thepossible
com-
Prod-
if
com-
if
com- production,
possible
possible
fixed.
fixed.
fixed.this
Depending
this can have on
Depending
Depending on
on
the
significant
the defect
the defect
defect
and
effects
and
and the
the
the
point
on the lead
point
point
during
time
during
during
munication
munication costs,
costs, and
and aa faster
faster transportation
transportation of goods
goods [1].
[1]. the
Prod-
the
Prod- production,
production, this can
can have
have significant
significant effects
effects on
on the
the lead
lead time
time
ucts have become
munication
munication costs, more
costs, and
and standardized,
aa faster
faster transportation
transportation as hasofofthe technology
goods
goods [1]. and
[1]. Prod- to
the
the
Prod- the waste.
production,
production, this can have significant effects on
this can have significant effects on the lead time the lead time
ucts
ucts have
have become
become more
more standardized,
standardized, as
as has
has the
the technology
technology and
to
and
to the
the waste.
waste.
produce
ucts
ucts have
have these
become
become[2]. Adding
more
more to this,
standardized,
standardized, it is becoming
as
as has
has the
themore and
technology
technology more
and
toWhile
the
andWhile
to the general belief is common, that higher automa-
the waste.
waste.
produce
produce these
these [2]. Adding
[2]. Adding
Addingtoto to this,
tocut it
this,the is
it is
is becoming
becoming more more
more and more
and more
more While the
the general
general belief
belief is
is common,
common, that
that higher
higher automa-
automa-
important
produce
produce for
these
these companies
[2].
[2]. Adding to this,
this, it
it environmental
is becoming
becoming more footprint
and
and tionWhile
by
more naturally
While the improves
the general
general the quality
belief
belief is
is common,
common, in manufacturing,
that
that higher the ef-
higher automa-
automa-
important
important for
for companies
companies to
to cut
cut the
the environmental
environmental footprint
footprint tion
by
tion
by naturally
naturally improves
improves the
the quality
quality in
in manufacturing,
manufacturing, the
the ef-
ef-
reducing
important
important energy
for and material
for companies
companies to
to cut usage.
cut the Only through this
the environmental
environmental footprint
footprint fects
is a sus-
tion
by
tion
by aren’t
naturally
naturally always
improves
improves positive
the and
quality
the and
qualityat least
in on its
manufacturing,
in manufacturing, own shouldn’t
the
the ef-
ef-
reducing
reducing energy
energy and
and material
material usage.
usage. Only
Only through
through this
this is
is fects
aa sus-
fects
sus- aren’t
aren’t always
always positive
positive and at
at least
least on
on its
its own
own shouldn’t
shouldn’t
tainable
reducing
reducing use
energy
energyof resources
and
and material
material achievable
usage.
usage. [3].
Only
Only These
through
through conditions
this
this is
is a
a be seen
in-
fects
sus-
fects
sus- as a sufficient
aren’t
aren’t always
always step towards
positive
positive and
and at
at a least
higher
least on
on quality
its
its own
own in manufac-
shouldn’t
shouldn’t
tainable use of resources achievable [3]. These conditions be
in-
be seen as
seen[8]. a sufficient
as Sometimes step towards a higher quality in manufac-
tainable
tensify
tainable
tainable
usepressure
the
use of resources
of resources
usepressure to reduce
of resources
achievable
the costs,
achievable
achievable
[3].butThese
[3]. These
[3].but also to
These
conditions
increase the
conditions
conditions
in-
turing
be
in-
be
in-seen as
seen[8]. aaa sufficient
sufficient step
as Sometimes
sufficient step
step towards
whentowards
towards aaa higher
the complex higher
higher
qualitybetween
interplay
quality in manufac-
in
qualitybetween
manufac-
in manufac-ma-
tensify
tensify the
the pressure to
to reduce
reduce the
the costs,
costs, but also
also to
to increase
increase turing
the
turing
the [8]. Sometimes when
when the
the complex
complex interplay
interplay between ma-
ma-
speed
tensify
tensifyof the
the
the delivered
pressure
pressure to
to goods.
reduce
reduce An
the
the enormous
costs,
costs, but
but factor
also
also to
to
to fulfill
increase
increase chine
these
turing
the
turing
the and
[8]. human
Sometimes
[8]. Sometimes operator
when isn’t
the
whenisn’t fully
complex
the complex taken into
interplay
interplay consideration,
between
between ma- ma-
speed of the delivered goods. An enormous factor to fulfill chine
these
chine and
and human
human operator
operator isn’t fully
fully taken
taken into consideration,
speed
demands
speed
speed
of
of the
of thewilldelivered
the delivered
delivered
goods.
be that only
goods.
goods.
An
An enormous
high-quality
An enormous
enormous items factor
factor
factor
to
to fulfill these
are produced.
to fulfill this can
chine
fulfill these
chine
these even
and
and human
humanexpand the problems
operator
operator isn’t
isn’t because
fully
fully taken
taken ofinto
the consideration,
into
into complexity of
consideration,
consideration,
demands will be that only high-quality items are produced. this
this can
can even expand the problems because of the complexity of
demands
Optical
demands will
demands
will be
quality
will be that
be that only
control
that only high-quality
systems
only high-quality have
high-quality items items
helped
items areare produced.
improve
are produced.
produced. the de-
this
this can even
the task
can [9]. This
even
even
expand
expand
expand
the
the problems
allows
the two general
problems
problems
because
because
because
of
of the
strategies
of theoncomplexity
the how to han-
complexity
complexity
of
of
of
Optical
Optical quality
quality control
control systems
systems have
have helped
helped improve
improve the
the the
de-
the
de- task
task [9].
[9]. This
This allows
allows two
two general
general strategies
strategies on
on how
how to
to han-
han-
livered
Optical
Opticalproducts
qualitycondition
quality control
control systems in manyhave
systems havefields of manufacturing
helped
helped improve
improve the dle
the de-in
the
the
de- the
task
task complexity
[9].
[9]. This
This to improve
allows
allows two
two the
general
general quality that
strategies
strategies don’t
on
on demand
how
how to
to to
han-
han-
livered
livered products
products condition
condition in
in many
many fields
fields of
of manufacturing
manufacturing dle
in
dle
in the
the complexity
complexity to
to improve
improve the
the quality
quality that
that don’t
don’t demand
demand to
to
the last
livered
livered years.
products
products Through
condition
condition the inpossibilities
many
inpossibilitiesfields
many fieldsof ofof digitalization,
manufacturing
of digitalization,
manufacturingad- and
ad-
dle
in
dle
in shouldn’t
the
the complexity
complexity be applied
to on
improve
to improve their
the own.
quality
theown. that don’t
quality that don’t demand to demand to
the last years. Through the and
and shouldn’t
shouldn’t be applied on their
the
the
the
last
vances
last
lastin
years.
in imaging
years.
years.
Through
hardware,
Through
Through
the
the possibilities
and Computer
possibilities
the possibilities
of
of digitalization,
Vision (CV).
digitalization,
of digitalization,
ad-
andThe
The
ad-
and
ad- first isbe
shouldn’t
shouldn’t be
betoapplied
reduce on
applied
applied thetheir
on
on their
their
own.
complexity
own. of a process step by
own.
vances imaging hardware, and Computer Vision (CV). The The
The first
first is
is to
to reduce
reduce the
the complexity
complexity of aa process step by
vances
sectors
vances
vances
in
in imaging
concerned
in imaging
hardware,
imaginginclude
hardware,
hardware, a wide and
and Computer
andrange
Computer
Computer from foodVision
Vision
Vision
(CV).
[4,(CV). The
analyzing
5] to solar
The The
The first
firstitit is
and
isand to using
reduce
to using tools
the such as leanof
the complexity
reducetools complexity of aa process
management
process
process step
ofmanagement
step
step
by
(LM)
by
by
sectors concerned
sectors concerned include
include aaa wide
wide range
range from food
from food [4,(CV).
[4, 5] to
5] The
analyzing
to solar
solar
analyzing it and using tools such
such as
as lean
lean management (LM)
(LM)
cell [6]
sectors
sectors and fabric
concerned
concerned [7] inspection.
include
include a wide
wide The
range
range check
from
from is usually
food
food [4,
[4, only
5]
5] to
to [10,
per-
solar 11,
analyzing
analyzing
solar 12]. it
it The
and
and aim
using
using of LM
tools
tools is
such
such to find
as
as less
lean
lean complex
management
management ways(LM)
(LM)of
cell
cell [6] and
[6] and fabric [7]
fabric [7] inspection.
inspection. The The check
check is is usually only
usually only [10,
per-
[10,
per- 11,
11, 12].
12]. The
The aim
aim of
of LM
LM is
is to
to find
find less
less complex
complex ways
ways of
of
cell
cell [6]
[6] and
and fabric
fabric [7]
[7] inspection.
inspection. The The checkcheck is is usually
usually only achieving
[10,
only per-
[10, 11, the
12].
11, 12]. planned
The aim results
of LM [13,
is to14,find15, 16,
less 17,
complex18, 19].
ways The
of
per-
achieving
achieving the The aim results
planned of LM [13, is to14, find15,less
16, complex
17,
17, 18, waysThe
19]. of
non-value the
achieving
achieving the
theadding planned
stepsresults
planned
planned results [13,
are removed
[13, 14,
[13, 14,
15,
and
15, 16,
continuous
16,
16, 17,
18, improve-
17, 18,
19].
19]. The
19]. The
non-value
non-value adding
adding stepsresults
steps are removed
are removed 14,and 15, continuous
and continuous 18, improve-
improve-The
2351-9789  c 2020 The Authors. Published by Elsevier Ltd. ment becomes
non-value
non-value adding
adding a focus.
steps
steps are
are removed
removed and and continuous
continuous improve-
improve-
2351-9789
2351-9789
2351-9789 cc ©2020

This is an open
2020
2020 The
The
access
The Authors.
Authors.
Authors.
article
Published
Published
Published
under by
by
byLtd.
Elsevier
Elsevier
the CC BY-NC-ND
Elsevier Ltd.
Ltd. ment
ment becomes
becomes aa focus.
focus.
license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
2351-9789
This
This isis
an
2351-9789 anc
open
c 2020
open The
access
2020 The Authors.
access
article Published
article
under
Authors. under
the CC
Published by
the Elsevier
CC
BY-NC-ND
by Elsevier Ltd.
BY-NC-ND
license
Ltd. license ment
ment becomes
becomes a
a focus.
(https://creativecommons.org/licenses/by-nc-nd/4.0/)
focus.
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open access article under
Peer-review under responsibility of
This is the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
scientific committee of the FAIM 2020.
This is an
an open
Peer-review
Peer-review
Peer-reviewopen access
under
under
access
under
article
article under
responsibility
responsibility of
under
responsibility
the CC
of the
BY-NC-ND
of the
the scientific
CC license
scientific
committee
BY-NC-ND
scientific of (http://creativecommons.org/licenses/by-nc-nd/4.0/)
committee
of the
license
committee FAIM of the FAIM 2021.
FAIM 2020.
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
the 2020.
Peer-review under
under responsibility
responsibility of
10.1016/j.promfg.2020.10.047
Peer-review of the
the scientific
scientific committee
committee of
of the
the FAIM
FAIM 2020.
2020.
 Daniel Schmidt et al. / Procedia Manufacturing 51 (2020) 330–336 331
Daniel Schmidt et al. / Procedia Manufacturing 00 (2019) 000–000 2

Another approach is to use tools such as artificial intelli-

gence (AI), which can handle a higher complexity as the prob-
lem it controls. AI aims to show intelligence by machines [20].
This includes a wide range of sub-fields with an even wider
range of goals. A big sub-field of AI is machine learning,
through which it is possible to learn correlations and patterns
from sample data. Within machine learning, deep learning has
seen an immense surge in the last years. Through it, it has been
possible to solve many complex challenges that assumed to be Fig. 1. Surface of an electromechanically engraved cylinder.
impossible to solve even with earlier AI tools [21].
The third option is to view LM and AI not merely as side-
by-side approaches, but as complementary and integrated parts
of each other. In this hybrid, both contribute their strengths. AI
can improve the decision-making process by offering new in-
sights. The LM framework is used to incorporate these results.
Through this, the human still stays in the focus and prevents the
system from becoming a black box.
As AI has become an ubiquitous tool in production, it is be-
coming an essential building block for many current improve-
ments. One of these developments can be found under the term
smart manufacturing [22]. In simple terms, it can be summa-
rized as methods that use resulting data from manufacturing to
improve the performance. This also comprises several subareas Fig. 2. Rotogravure printing.
such as predictive maintenance, quality control [23] and quality
improvement, which will be the key focus in this paper.
This paper aims to further examine the possibilities how
Deep Learning (DL) can be employed for quality improvement because of the many influencing parameters and the required
by proposing steps in rotogravure manufacturing using DL. The high accuracy. Therefore, there is enough room for Machine
results are produced on the example of the rotogravure man- Learning (ML) techniques, and DL in particular to be applied
ufacturing industry but can also be adapted to other areas of and to help improve the processes. As the first step to the anal-
manufacturing. ysis, the production steps are described.

2.2. Production process

2. Rotogravure manufacturing
2.1. Overview
Rotogravure cylinders are one of the most important printing
methods in the packaging industry and were developed in the
early 1890s [24]. Belonging to the family of intaglio printing,
which has an even longer history dating back to the fifteenth
century, making them one of the oldest printing technologies
[25]. The cylinder has cells in the cylinder (Figure 1), that fill
with ink. This ink gets released during printing, as seen in Fig-
ure 2, by being pressed against an impression cylinder. In its
current form, rotogravure cylinders still work through the same
principle as they did in the beginnings. Yet, manufacturing ro-
togravure cylinders has experienced many advances through the
help of digitalization.
This has allowed the rotogravure cylinder to keep its place as
the most important printing method when high print quality and
long print runs, starting at a run of at least 100.000 units [26],
are needed. These cases have been decreasing. More printing
designs are only used for shorter times. This in turn increases
the need to also become relevant for smaller printing runs.
Although manufacturing rotogravure cylinders has this ex-
tensive history, the production processes are still prone to errors
2
The general production process is highly standardized. The
key steps can be seen in Figure 3. Although the core of the
printing cylinder is made of a steel plate that is enclosed by a
copper layer, the standard process usually starts with used print-
ing cylinders. These get dechromed and decoppered (3-1). After
this, the new production can start. Before a fresh copper layer
can be added, the cylinder is degreased (3-2) to remove any
kind of contaminant that could have negative effects on the fur-
ther production steps. If this doesn’t happen correctly or if any
dirt remains on the cylinder surface, this can lead to defects.
In the next step, a layer of copper is added in a galvanic pro-
cess (3-3). Because of the earlier production step, and through a
multitude of influencing parameters of the coopering machine,
holes in the copper surface can emerge. These are already vis-
ible in this step. Still many more difficulties can arise in this
process step, that aren’t visible by the eye. The major points are
the copper hardness or a contamination of the copper that could
lead to further problems in the next production step.
The following step adds the desired illustration to the print-
ing cylinder (3-4). Depending on the requirements, a multitude
of methods are available to achieve this goal. The most used
method is the electromechanical engraving, that uses a diamond
stylus to remove small cells of varying sizes of copper from the
332 Daniel Schmidt et al. / Procedia Manufacturing 51 (2020) 330–336
Daniel Schmidt et al. / Procedia Manufacturing 00 (2019) 000–000
Fig. 3. The main steps of rotogravure cylinder manufacturing. On the left the
on-top view of the cylinder. In the middle a cross section of the cylinder and on
the right the description.
surface, that fill with ink during printing and release it on the
printing substrate with the help of a doctor blade and a rotary
press. Also, other methods are available, from which many use a
laser directly or indirectly to form cells on the cylinder surface.
In scarce cases, engraving can even be done by hand. During
the process, many kinds of defects can emerge. For the elec-
tromechanical engraving this could, for example, be a broken
or fractured stylus or defects of the cylinder surface.
In the last step, the imaged rotogravure cylinder receives a
chrome plating (3-5). The reason being that a copper surface
wouldn’t be hard enough for the pressure needed during the
printing process. In this last step, it is also possible that defects
such as holes on the cylinder surface can occur. A contamina-
tion of the cylinder surface or chrome could cause such flaws.
2.3. Evaluating potential savings
The potential savings of energy, resources, and time depend
on multiple factors. In rare cases it is possible that defects of
a production step are found before the last quality check. For
an easier comparison and as the percentage is currently small,
in the following it is considered that the defects only get found
after chroming in the last control step.
The two major factors for the environmental consequences
are the impacts of a defect and how often it occurs. The cost of
the defect can be measured by the impact on the various forms
of environmental waste, but also on further kinds of drain used
from a management point of view. To a specific degree, most in-
clude environmental waste. Perhaps every kind of waste found
in LM to a certain extent impact the environmental waste, as
it has been shown, that prior lean experience can be an impor-
tant predecessor for environmental management practices [27].
Still, both environmental waste and lean waste can sometimes
3
3
even stand in conflict [28]. Therefore, in the following, the view
on the waste is through the lens of environmental waste.
The environmental waste of a defect needs to be measured
by examining the impact it has on the production compared to
a defect free item. Only viewing the environmental waste, it is
composed of all the material and energy that is needed to fix the
defect. For some types of defects, it’s possible to fix a mistake
by correcting the current production step. For other defects it
can be necessary, that also earlier steps have to be repeated as
the defect can not be fixed.
To get a complete picture, it can be an interesting challenge
to estimate a realistic potential of how much a production step
could be improved. As this can usually just be an educated
guess and could lead to falsely dismissing production steps that
have a high potential, the basis of deciding which production
step has the highest potential to be improved, should only be
based on the resulting waste of a production step.
2.4. Potential savings of the production steps
In the following, the results using the discussed methods for
determining the potential savings of every manufacturing step
are examined. Defects are usually found at the end. Therefore,
the root cause for the defect can’t always be determined. This
raises the difficulty of finding the source of a defect, and in
which production step it occurs. Also, it increases the difficulty
of determining the exact potential waste savings for every pro-
duction step. Because of this and to show generic data, in the
following relative figures in the form of a proportion of the total
waste of each production step between 0-1 are given.
During dechroming and decoppering, the risk for resulting
defects are very low. Although imperfections in the resulting
copper layer can exist, these neither significantly interfere with
the next production steps nor directly lead to defects on the
cylinder surface, as long as these aren’t extreme. A rare de-
fect could be that the decoppering hasn’t been deep enough,
and cells of the earlier engraving are still visible. Although a
rare case, this would cause immense waste, as all production
steps would have to be repeated to produce a defect free ob-
ject. Therefore, the proportion of defective items through this
production step are rated with 0.1 in the following evaluation.
The degreasing step is also a low risk production step. Nev-
ertheless, big contaminations could lead to dramatic conse-
quences, as the copper wouldn’t be able to form a solid layer.
The degreasing step still only contributes to 0.1 of the added
waste.
The coopering phase has the highest associated risk for re-
sulting defects. During this production step, the major risks are
holes in the cylinder surface. This is also confirmed in [29] as
the highest waste producing category for rotogravure manufac-
turing. Sizeable holes in the surface could print. Even if these
are smaller or the copper hardness isn’t consistent, this could
lead to defects in the next production phase. Hence this produc-
tion step adds a proportion of 0.45 to the total.
For the engraving phase, the complexity of defects is the
highest, as the cells need to be placed with a very high accu-
racy and also with a consistent depth across the complete cylin-
 Daniel Schmidt et al. / Procedia Manufacturing 51 (2020) 330–336
Daniel Schmidt et al. / Procedia Manufacturing 00 (2019) 000–000
der surface. As a result, the range of unique types of defects is
very high whereas the most common is the fracture of the stylus
[30]. But also other defects such as distortions of the engraved
image in any direction or pin holes are possible. As a result, this
production step comprises 0.25 to the total waste.
In the last step, the chroming can also add defects to the item.
These could be small holes or an incomplete chrome plating.
Due to it being the last step and as most of the defects result-
ing from this production step can be fixed be dechroming and
rechroming the cylinder, the total waste is manageable. This
step is responsible for 0.1 of the total waste.
Because of the analysis, the highest waste saving potential
can be seen in the copper plating phase. This is the produc-
tion step responsible for the highest amount of waste. Also, if
defects are found at this stage, repairing the item has a much
lower cost than in later stages. As an added benefit, it can help
in analyzing the origin of a defect.
2.5. Methods for improving quality control
The previous analysis shows the best next target point for the
future work. The implementation allows for various strategies.
In this section three alternatives are described and rated based
on the costs and saving potentials.
Fig. 4. Three different possible quality checks for rotogravure manufacturing.
The currently deployed quality check on the left. An alternative with more vi-
sual inspections in the middle and a data based method on the right.
The different alternatives are visualized in Figure 4. On the
left (I) the quality check deployed at the investigated manufac-
turing plant can be seen, where only one visual inspection at the
end is used, that uses algorithms from computer vision and AI
to detect defects [23]. This kind of quality check has its equiv-
alents in other manufacturing areas, such as for laser welding
[27] or metal additive manufacturing [31]. Though visual in-
spection systems show outstanding results for the final quality
4
333
4
check, as discussed, this still doesn’t prevent internal waste to
accumulate, as defects are usually only found at the end.
The first alternative to the current workflow (Figure 4-II)
uses multiple instances of the visual inspection system. Al-
though it is possible to use a visual inspection system after ev-
ery production step, only the three steps with high waste were
chosen for this alternative. Nevertheless, a visual inspection
system could be used for each step. The visual inspection sys-
tem could ensure that only items with no defects are passed on
to the next production step. This would make it possible to fix
or discard an item without producing waste by adding the next
production steps to the already broken item. A downside to this
approach is the strong cost point for the visual inspection sys-
tems and that only the symptoms in the form of defective items
are considered and not the root cause in form of a not optimized
production step that only produces defect free items.
The second alternative (Figure 4-III) doesn’t use any further
visual inspection systems. Quality checks are data based. De-
tails on feasible ways how this can be achieved are discussed
in Section 3. Depending on the achievements of the data based
checks, it would also be possible to replace the final visual in-
spection system. This alternative would reduce the costs for
needed hardware and could help analyze the core problems in
each process step.
In the long term, the correlation between the defects found
by the inspection systems and the operational parameters will
make it possible to address 4-III, following the LM culture.
3. Deep Learning for rotogravure manufacturing
3.1. Data based alternatives
DL appears to be a promising tool to improve the manufac-
turing quality. Still, other data based strategies should be taken
into consideration. Through the rise of data generation and col-
lection, many types have been developed over the years. All
coming with advantages and disadvantages. Though in general
terms this includes LM focused systems such as Six Sigma or
Kaizen [32], the further focus lies on automated techniques that
process data and give an output which is used to improve the
quality.
A solution to a problem should never be more complex than
it needs to be. Therefore, a first step should almost always
be the visualization of the data and if possible simple models
could be developed and the correlations of the factors checked.
Through this, big influencing factors can already be determined.
Domain-knowledge can be a huge plus in this stage. Unfortu-
nately, this approach is not sufficient for most cases as a man-
ufacturing system is usually dynamic, uncertain, and complex
[33]. With the help of machine learning, it is possible to model
more complex systems. This gives an enormous advantage as
most processing steps have many influencing parameters that
result in a high dimensional relation.
Most ML algorithms can be grouped into two major cat-
egories. Supervised and unsupervised learning. Unsupervised
learning is used to detect patterns in unlabeled data sets and
334 Daniel Schmidt et al. / Procedia Manufacturing 51 (2020) 330–336
Daniel Schmidt et al. / Procedia Manufacturing 00 (2019) 000–000
generally needs little supervision by a human. Supervised learn-
ing however depends on a labeled data-set and learns to find the
connecting patterns between input and label. It learns from the
examples provided. If the data and the use-case make it possi-
ble, supervised learning usually achieves better results.
Through the data won from the vision based systems it is
possible to label the data with the information of existing de-
fects. This makes it possible to use supervised learning if the
use-case allows it.
The more concrete groups that are promising for the planned
results are classification and regression. Though the exact use-
case depends on the planned achievements, which will be fur-
ther discussed in Section 3.2. Classification algorithms could
be used to detect defective states. This would be the best choice
if the factors resulting in a defect are only temporary. Through
regression, prediction and forecasting could be implemented.
This would be more useful if the influencing parameters are
more continuous and need to be kept in check.
Since the 2000s DL, which belongs to ML, has seen a huge
surge in popularity. Though more traditional ML approaches
such as Support Vector Machines (SVM) have also shown suc-
cesses in manufacturing [34] [35], DL are most often superior.
The more traditional approaches require a manual feature ex-
traction, while DL networks are able to learn more complex
features in each layer which reduces the difficulty significantly
[22]. The advantages of deep neural networks have been proven
mathematically [36]. This is one of the reasons why it has seen
many successful implementations in manufacturing for the last
years [37] [38] and why it is chosen as the prime candidate for
the future work.
3.2. Roadmap to improve rotogravure manufacturing using DL
To test which algorithm is most suitable for a problem, the
required results and the data that can be used to train the neural
network need to be examined.
The eventual goal is to find ways that reduce the waste in
the production, which allows for distinct strategies. These can
range from detecting, with a high probability, that a defective
item was produced to further inspect the item for faults, to im-
proving the production step by adjusting controllable variables
that have an influence during the production through an au-
tomatic system. While the latter would be the ideal end state
where every production step would adjust itself to only produce
defect free items.
According to the Industry 4.0 maturity index from Schuh et
al. [39], the steps visibility, transparency, prognosis and auton-
omy can be described as a path towards Industry 4.0 or in more
general terms as an improvement to the manufacturing. This
can also be adapted to this case of quality improvement with
the help of DL as seen in Figure 5.
The first step increases the visibility of a system. Here
this could, for example, be done by visualizing the interact-
ing parameters through feature reduction with the help of auto-
encoders [40].
5
5
Fig. 5. Potential steps in order to improve waste saving through quality im-
provement aided by the possibilities of DL.
Creating transparency can achieve further potential. For ex-
ample, showing which parameters have a negative influence on
quality or how these parameters interact.
Making a prognosis can increase the potential. For exam-
ple, the staff could see the influence of the current state on the
system. With this knowledge, it’s possible to know which pa-
rameters have the biggest influence on the quality and should
be adjusted.
As the last step, a fully autonomous system can be imag-
ined, that adjusts its own parameters to only produce defect free
items.
Because the goal is to detect defects, it is necessary to know
when a defect has occurred. After the copper plating, the cylin-
der is checked by scanning the cylinder surface with the help
of a line-scan camera and an LED-light. Any holes, or other
anomalies can be detected through a deviation to the cylinder
surface. The check after the image transfer and the chrome
plating is more difficult but also starts by scanning the cylin-
der surface. Now it needs to be compared with the engraving
file through transformations and brightness adaptations as de-
scribed in [23]. These results are the target for all supervised
ML/DL algorithms in the training phase.
For the input data, it is necessary that it in more or less hid-
den form contains the information that a defect exists. It is desir-
able that no further delays or costs are added. These properties
to the current knowledge should be able to be achieved by the
data already produced by the machines in every production step.
These are usually a mix of different outputs from sensors and
parameters from the machine. In the case of the copper (Fig-
ure 4.3) and chrome plating (4.5) these would apply to the gal-
vanization process such as the temperature, the amount of ad-
ditives or the current strength. For the image transfer step (4.4)
these would be related to the speed of the machine, the pressure
to the stylus and any fluctuations in the electricity used. If it is
not possible to achieve the wanted results, added sensors could
be installed that contribute further information.
 Daniel Schmidt et al. / Procedia Manufacturing 51 (2020) 330–336 335
Daniel Schmidt et al. / Procedia Manufacturing 00 (2019) 000–000 6

4. Discussion and Conclusions

This paper proposes steps that should make it possible to

improve the quality in manufacturing and especially in ro-
togravure manufacturing for this case systematically. To ver-
ify the results, the authors aim to extend this work in the next
months by applying these concepts to real-life data of a ro-
togravure cylinder manufacturer. This should help to show that
a general systematic quality improvement and waste reduction
strategy is possible by using data and DL algorithms.
A roadmap of the planned next steps is shown in Figure 6 to
conclude this paper. It is not a simple transformation from 4-I to
4-III, but consists of multiple steps that build upon each other.
The first step will be to have a visual inspection system after the
copper plating. This gives the information if a cylinder has de-
fects and will be the target data for the DL system in this part.
The second step will be to record the data from the used ma-
chines for the copper plating. This will be the input for the DL
system, from which it will learn which parameters influence the
emergence of defects. In the next step, the DL system will add
visibility to the copper plating by showing which parameters
have the biggest influence on the emergence of defects. Adding
to this, the next step will allow transparency by identifying why
defects have occurred. Through integrating the prognosis in the
next step, the ability will be added to determine the future out-
come and gives the ability to take preventive countermeasures.
These steps will be repeated for all the production steps until
in the last step full automation for the quality control can be
achieved.

Acknowledgements

Joaquín Ordieres-Meré wants to acknowledge the Spanish

Agencia Estatal de Investigación, through the research project
Fig. 6. The planned next steps for the DL based quality improvement in ro-
with code RTI2018-094614-B-I00 into the “Programa Estatal
togravure manufacturing.
de I+D+i Orientada a los Retos de la Sociedad”.

References
[1] D Hummel. Transportation costs and international trade over time. Journal
of Economic Perspectives, forthcoming, 2008.
[2] John Williamson. Globalization: The concept, causes and consequences.
In Institute for International Economics, Keynote address at the Congress
of the Sri Lankan Association for the Advancement of Science, Colombo,
15th December, 1998.
[3] Arjen Y Hoekstra and Thomas O Wiedmann. Humanity’s unsustainable
environmental footprint. Science, 344(6188):1114–1117, 2014.
[4] Tadhg Brosnan and Da-Wen Sun. Improving quality inspection of food
products by computer vision—-a review. Journal of food engineering,
61(1):3–16, 2004.
[5] J_ Blasco, S Cubero, J Gómez-Sanchís, P Mira, and E Moltó. Development
of a machine for the automatic sorting of pomegranate (Punica granatum)
arils based on computer vision. Journal of food engineering, 90(1):27–34,
2009.
[6] Fu Zhuang, Zhao Yanzheng, Liu Yang, Cao Qixin, Chen Mingbo, Zhang
Jun, and Jay Lee. Solar cell crack inspection by image processing. In Pro-
ceedings of 2004 International Conference on the Business of Electronic
Product Reliability and Liability (IEEE Cat. No. 04EX809), pages 77–80.
IEEE, 2004.
6
[7] Hamed Sari-Sarraf and James S Goddard. Vision system for on-loom fabric
inspection. In 1998 IEEE Annual Textile, Fiber and Film Industry Technical
Conference (Cat. No. 98CH36246), pages 8–1. IEEE, 1998.
[8] Clifford Martin Hinckley. Combining mistake-proofing and Jidoka to
achieve world class quality in clinical chemistry. Accreditation and quality
assurance, 12(5):223–230, 2007.
[9] Sharon Parker and Gudela Grote. Automation, algorithms, and beyond:
Why work design matters more than ever in a digital world. Applied Psy-
chology, 2020.
[10] J. Villalba-Diez and J Ordieres-Mere. Improving manufacturing opera-
tional performance by standardizing process management. Transactions
on Engineering Management, 62(3):351–360, May 2015.
[11] Villalba-Diez, J., J. Ordieres-Mere, H. Chudzick, and P Lopez-Rojo. NE-
MAWASHI: Attaining Value Stream alignment within Complex Organiza-
tional Networks. In Procedia CIRP, volume 37, pages 134–139, Cranfield,
UK, 2015. Elsevier.
[12] Villalba-Diez, J., Ordieres-Meré, J., and Nuber, G. The HOSHIN KANRI
TREE. Cross-Plant Lean Shopfloor Management. In Procedia CIRP, vol-
ume 32, pages 150–155, Bochum, Germany, June 2015. Elsevier.
[13] J. Villalba-Diez, J. Ordieres-Mere, and S. Rubio-Valdehita. Lean Learning
Patterns. (CPD)nA vs. KATA. Procedia CIRP, 54:147–151, October 2016.
336 Daniel Schmidt et al. / Procedia Manufacturing 51 (2020) 330–336
Daniel Schmidt et al. / Procedia Manufacturing 00 (2019) 000–000 7

[14] Villalba-Diez, J. and Ordieres-Meré, J. Strategic Lean Organizational De- [34] Bernardete Ribeiro. Support vector machines for quality monitoring in a
sign: Towards Lean World-Small World Configurations through Discrete plastic injection molding process. IEEE Transactions on Systems, Man,
Dynamic Organizational Motifs. Mathematical Problems in Engineering, and Cybernetics, Part C (Applications and Reviews), 35, 2005.
2016:1–10, 2016. [35] Ali Azadeh, Morteza Saberi, Ahmad Kazem, Vahid Ebrahimipour, A Nour-
[15] J. Villalba-Diez. The Lean Brain Theory. Complex Networked Lean Strate- mohammadzadeh, and Zahra Saberi. A flexible algorithm for fault diagno-
gic Organizational Design. CRC Press. Taylor and Francis Group LLC, sis in a centrifugal pump with corrupted data and noise based on ann and
Boca Raton, FL, USA, 2017. support vector machine with hyper-parameters optimization. Applied Soft
[16] J. Villalba-Diez, I. DeSanctis, J. Ordieres-Meré, and F.E. Ciarapica. Lean Computing, 13, 2013.
Structural Network Resilience. In C. Cherifi, H. Cherifi, M. Karsai, and [36] Tomaso Poggio and Steve Smale. The mathematics of learning: Dealing
M. Musolesi, editors, Complex Networks & Its Applications VI: Proceed- with data. Notices of the AMS, 50, 2003.
ings of Complex Networks 2017 (The Sixth International Conference on [37] Te Han, Chao Liu, Wenguang Yang, and Dongxiang Jiang. A novel ad-
Complex Networks and Their Applications), number 689 in Studies in versarial learning framework in deep convolutional neural network for in-
Computational Intelligence, pages 609–619. Springer International Pub- telligent diagnosis of mechanical faults. Knowledge-Based Systems, 165,
lishing, Lyon, France, November 2017. 2019.
[17] J. Villalba-Diez. The HOSHIN KANRI FOREST. Lean Strategic Organi- [38] Jack Francis and Linkan Bian. Deep learning for distortion prediction in
zational Design. CRC Press. Taylor and Francis Group LLC, Boca Raton, laser-based additive manufacturing using big data. Manufacturing Letters,
FL, USA, 1st edition, 2017. 20, 2019.
[18] J. Villalba-Diez, J. Ordieres-Meré, M. Molina, M. Rossner, and M. Lay. [39] Günther Schuh, Reiner Anderl, Jürgen Gausemeier, Michael Ten Hompel,
LEAN DENDROCHRONOLOGY: COMPLEXITY REDUCTION BY and Wolfgang Wahlster. Industrie 4.0 Maturity Index: Die digitale Trans-
REPRESENTATION OF KPI DYNAMICS LOOKING AT STRATEGIC formation von Unternehmen gestalten. Herbert Utz Verlag, 2017.
ORGANIZATIONAL DESIGN. Management and Production Engineer- [40] Mayu Sakurada and Takehisa Yairi. Anomaly detection using autoen-
ing Review, 9(4):3–9, 2018. coders with nonlinear dimensionality reduction. In Proceedings of the
[19] Javier Villalba-Díez, Martin Molina, Joaquín Ordieres-Meré, Shengjing MLSDA 2014 2nd Workshop on Machine Learning for Sensory Data Anal-
Sun, Daniel Schmidt, and Wanja Wellbrock. Geometric Deep Lean Learn- ysis, pages 4–11, 2014.
ing: Deep Learning in Industry 4.0 Cyber–Physical Complex Networks.
Sensors, 20(3), 2020.
[20] Stuart J Russell and Peter Norvig. Artificial intelligence: a modern ap-
proach. Malaysia; Pearson Education Limited„ 2016.
[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classi-
fication with deep convolutional neural networks. In Advances in neural
information processing systems, pages 1097–1105, 2012.
[22] Jinjiang Wang, Yulin Ma, Laibin Zhang, Robert X Gao, and Dazhong Wu.
Deep learning for smart manufacturing: Methods and applications. Journal
of Manufacturing Systems, 48:144–156, 2018.
[23] Javier Villalba-Diez, Daniel Schmidt, Roman Gevers, Joaquín Ordieres-
Meré, Martin Buchwitz, and Wanja Wellbrock. Deep Learning for Indus-
trial Computer Vision Quality Control in the Printing Industry 4.0. Sensors,
19(18), 2019.
[24] Dusan Stulik and Art Kaplan. The Atlas of Analytical Signatures of Photo-
graphic Processes: Photogravure. Getty Conservation Institute, 2013.
[25] Charles Harrison. The printed picture in Renaissance Europe. Yale Uni-
versity Press in association with the Open University, 2007.
[26] Guido Hennig, Karl Heinz Selbmann, Stefan Mattheus, Ralph Kecke, and
Stephan Brüning. Laser Precision Micro Fabrication in the Printing Indus-
try. JLMN, 1:89–98, 2006.
[27] Ma Ga Mark Yang, Paul Hong, and Sachin B Modi. Impact of lean man-
ufacturing and environmental management on business performance: An
empirical study of manufacturing firms. International Journal of Produc-
tion Economics, 129(2):251–261, 2011.
[28] Sandra Rothenberg, Frits K Pil, and James Maxwell. Lean, green, and the
quest for superior environmental performance. Production and operations
management, 10(3):228–243, 2001.
[29] Andira Andira, Nurhayati Surbakti, and others. Menurunkan Problem Re-
ject Pinhole di Proses dengan Konsep DMAIC di PT. XYZ, Purwakarta.
Journal of Industrial Engineering, 3(2):119–128, 2019.
[30] Dhirender, Vikas Jangra, and Sandeep Kumar Garg. Analysis of various
defects during electromechanical engraving of gravure cylinder. Interna-
tional Journal of Engineering Sciences & Research Technology, 5(4):236–
239, April 2016.
[31] Wenyuan Cui, Yunlu Zhang, Xinchang Zhang, Lan Li, and Frank Liou.
Metal Additive Manufacturing Parts Inspection Using Convolutional Neu-
ral Network. Applied Sciences, 10(2):545, 2020.
[32] Gülser Köksal, İnci Batmaz, and Murat Caner Testik. A review of data
mining applications for quality improvement in manufacturing industry.
Expert systems with Applications, 2011.
[33] Thorsten Wuest, Daniel Weimer, Christopher Irgens, and Klaus-Dieter
Thoben. Machine learning in manufacturing: advantages, challenges, and
applications. Production & Manufacturing Research, 2016.
7