Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 55-10 (2022) 13–18
Knowledge
Knowledge Graphs
Graphs in
in Digital
Digital Twins
Twins for
for
Knowledge
Manufacturing
Knowledge Graphs
-
Graphs in
Lessons
in Digital
Learned
Digital Twins
from
Twins for
an
for
Manufacturing
Manufacturing -
- Lessons
Lessons Learned
Learned from
from an
an
Manufacturing
Industrial
Industrial Case
Case - Lessons
at
at Atlas
Atlas Learned
Copco
Copco from
Airpower
Airpoweran
Industrial Case
Industrial Case at Atlas Copco
at Atlas Copco Airpower
Airpower
∗ ∗ ∗
Bart
Bart MeyersMeyers ∗∗ Johan Johan Van
Van Noten
Noten ∗∗ Pieter
Pieter Lietaert
Lietaert∗ ∗∗
∗∗ ∗∗
Bavo
Bart
BavoMeyers Tielemans
Meyers
Tielemans ∗ Johan∗∗ Hristo
Van Hristov
Noten ∗ ∗∗ Davy
Pieter Maes
Lietaert
Maes ∗∗ ∗∗
∗∗ Hristo Hristov ∗∗ Davy
∗ ∗
Bart
Bavo Tielemans Johan
Klaas Van
Hristo Noten
Gadeyne
Hristov ∗ Pieter
∗ Davy Lietaert
Maes ∗
∗∗
Klaas
∗∗ Gadeyne ∗∗
∗∗
Bavo Tielemans Klaas Hristo Hristov
Gadeyne ∗ Davy Maes ∗
∗
∗ Klaas Gadeyne ∗
∗ CodesignS @ Flanders Make vzw, Gaston Geenslaan 8, 3001
∗ CodesignS @ Flanders Make vzw, Gaston Geenslaan 8, 3001
Heverlee,
∗ CodesignS
Heverlee,
∗
Belgium
Belgium (e-mail:
(e-mail: [email protected]).
@ Flanders Make vzw, Gaston Geenslaan 8, 3001
[email protected]).
∗∗ CodesignS @ Flanders Make vzw, Gaston Geenslaan 8, 3001
∗∗ AtlasBelgiumCopco Airpower
(e-mail: nv, Boomsesteenweg
Heverlee,
Atlas
Heverlee,
∗∗ Belgium Copco
Belgium Airpower
(e-mail: nv, Boomsesteenweg 957, 957, 2610
[email protected]).
[email protected]). 2610 Wilrijk,
Wilrijk,
∗∗ Atlas
∗∗ Belgium Copco (e-mail:
Airpower
(e-mail: [email protected]).
nv, Boomsesteenweg 957, 2610 Wilrijk,
[email protected]).
Atlas
Belgium Copco Airpower nv, Boomsesteenweg 957, 2610 Wilrijk,
Belgium (e-mail: (e-mail: [email protected]).
[email protected]).
Abstract:
Abstract: In In this
this paper
paper we we introduce
introduce an an architecture
architecture for for aa cognitive
cognitive digital
digital twin
twin thatthat uses
uses
an
an ontology-based
Abstract:
ontology-basedIn this knowledge
paper
knowledge we graph
introduce
graph to anto improve
an the
architecture
improve analysis
for
the analysis a of manufacturing
cognitive digital
of manufacturing systems.
twin that
systems. The
uses
The
Abstract:
architecture In
is this paper
evaluated inwe a introduce
use case at architecture
Atlas Copco for
Airpower,a cognitive
revolving digital
around twin an that uses
adaptive
an ontology-based
architecture is evaluated knowledge in a use graph to improve
casetoatimprove
Atlas Copcothe analysis
Airpower, of manufacturing
revolving systems.
aroundsystems.
an adaptive The
an ontology-based
measurement
architecture isstrategy knowledge
evaluated for inquality graph
a use case control. We
at Atlas the
report
Copco on analysis
the
Airpower, of manufacturing
requirements,
revolving aroundoutcomes, and The
an adaptive the
measurement
architecture strategy
isstrategy
evaluated for quality
inquality
aa use control.
case at We
Atlas report
Copco on the
Airpower, requirements,
revolving outcomes,
around and
an research.
adaptive the
resulting
measurement
resulting challenges
challenges of using
for
of using cognitive
a cognitive control.digital
We
digital twin,
report andon provide
the a roadmap
requirements, for future
outcomes, and the
measurement strategy for quality control. We twin,
reportand onprovide a roadmap for
the requirements, future research.
outcomes, and the
resulting
Copyrightchallenges
© 2022 Theof using aThis cognitive
is an opendigital twin,
access andunder
article provide
the CC a roadmap
BY-NC-ND forlicense
future research.
resulting challenges ofAuthors.
using a cognitive digital twin, and provide a roadmap for future research.
(https://creativecommons.org
Keywords:
Keywords: Intelligent
Intelligent manufacturing /licenses/by-nc-nd/4.0/)
manufacturing systems,
systems, Modeling
Modeling of of manufacturing
manufacturing operations,operations, Quality
Quality
assurance
Keywords:
assurance and maintenance
Intelligent manufacturing systems, Modeling of manufacturing operations, Quality
Keywords:and maintenance
Intelligent manufacturing systems, Modeling of manufacturing operations, Quality
assurance and
assurance and maintenance
maintenance
1.
1. INTRODUCTION plications,
INTRODUCTION plications, existing
existing production
production knowledge
knowledge remains remains largelylargely
1. INTRODUCTION untapped
plications,
untapped as as collecting
existing
collecting data
production is a cumbersome
knowledge
data is knowledge
a cumbersome remains process.
largely
process. A
A
1. INTRODUCTION plications, existing production remains largely
With the the advent
advent of of Industry
Industry 4.0 4.0 and and the the Internet
Internet of gap remains
untapped as between
collecting traditional
data is aknowledge-based
of gap remains between traditional knowledge-based produc- cumbersome produc-
process. A
With untapped
tion analysis as collecting
on the one data
hand is a
and cumbersome
data-based process.
production A
Things,
With the the potential
thetheadvent
advent of
of Industry collected
Industry data
4.0 dataand istheis
the deemed
Internet huge,of tiongap
gap remains between
analysisbetween
remains on the one traditional
hand and
traditional knowledge-based
data-based production
knowledge-based produc-
produc-
Things,
With potential of of collected 4.0 and deemed
1 Internet huge,of analysis driven
if used in
Things, intheartificial
potential intelligence
of collected
collected (AI)datasystems huge, tion
1 . In current
is deemed
deemed analysis
analysis driven onby theIndustry
one hand 4.0 on
on the
4.0and other
other hand.
data-based production
if used
Things, artificial
the potential intelligence
of (AI) systems
data is 1 .neural
In current
huge, tion analysis
analysis driven onby
bytheIndustry
one hand
Industry 4.0and
on
the
data-based
the other
hand.
production
hand.
state-of-the-practice,
if used in artificial AI methods
intelligence (AI) using
systems e.g., . In net-
current This paper
paper aims to bridge this gap by providing a
aims to bridge this gap by providing a digital
state-of-the-practice, AI methods using 1
e.g.,1 .neural net- analysis driven by Industry 4.0 on the other hand.
if used in artificial intelligence (AI) systems In current
works,
works, classifiers,
classifiers, etc.
state-of-the-practice, are
etc. AI used
used based
are methods based on
on the
using thee.g.,available
neural data
available net- This
data twin
This of
paperthe production
aims to bridgesystems.
this gapSuchby a digital
providing twin
digital
a digitalen-
state-of-the-practice,
to support
works, or
classifiers, automate
etc. AI
are methods
decisions
used based using
like
on e.g.,
predictive
the neural
available net-
main-
data twin paper
This of the aimsproduction
to bridgesystems.
this gapSuchby a digital twin
providing a en-
digital
to support
works, or automate
classifiers, etc. are decisions
used based like
on predictive
the available main-
data compasses
twin of
compasses the all data
production
allproduction and knowledge
systems.
data and systems.
knowledgeSuchSuchabout
about a the
digital production
twin
the production en-
tenance,
to support quality
or control,
automate machine
decisions setting
like optimization,
predictive main- twin
system, of the
and a digital twin en-
tenance,
to support quality
or control, decisions
automate machine like setting optimization,
predictive main- compasses
system, and allpresents
data and
presents it
it in
in the
the form
knowledge
form of aa knowledge
about
of the production
knowledge graph.
graph.
scheduling,
tenance, etc.
quality The success
control, of
machinethese are
setting limited to specific
optimization, compasses
The main all
goal data
of and knowledge about the production
scheduling,
tenance,
applications,
etc. The
quality
like
successmachine
control,
the use
of these setting
are limited to specific system,
optimization, The
system,main and
and goal of this
presents
presents thisit digital
in
in the
it digital
the
twin
form
twin of
form
in
in aathe
of the context
context of
knowledge
knowledge of this
graph.
this
graph.
scheduling,
applications,etc.
scheduling, likeThe
etc. The use of
thesuccess
success
image
of of
of
theserecognition
image
these
are limitedfor
recognition
are limited for
to
quality
to specific
quality The
specific paper
papermainis
main to support
goal
is to goal
support of the
this AI cycle:
digital
the digital
AI cycle: twinfind inand
findinand the analyze
context
analyze insights
of this
insights
control.
applications,
control. A largely
like
A largely unsolved
the use
unsolved of challenge
image
challenge is
recognitionusing
is using AI AI
for to gain
quality
to gain like The correlations of this
within the twin
production the context
system, and of this
learn-
applications, like the use of image recognition for quality paper is to
like correlationssupport the
within AI cycle:
thecycle: find
production and analyze
system, insights
andinsights
learn-
broader
control. insight
A largelyin the full
unsolved production
challenge domain,
is using like
AI predict-
to gain paper
ing, is to support
deployment, the
execution AI find and analyze
broader
control. insight
A largely in the full production
unsolved challenge domain,
is using like
AI predict-
to gain like ing, correlations
deployment,within execution the and validation
production
and validation of
of an
system, an andAI
AI model.
learn-
model.
ing the
the influence
broader insight in of
the production
full production conditions
domain, on the
on like end-of-
end-of- like
thepredict- The correlations
obtained within
knowledge the production
throughout the system,
execution and learn-
of
ing
broader
line
ing
influence
insight
performance.
the influence
in of
the
The
of
production
full production
reason
production
conditions
for this domain,
conditions is mainly
on
like
the
predict-
twofold.
end-of-
ing,
The
ing, deployment,
obtained knowledge
deployment, execution
execution and
throughoutvalidation
andcycles. of an
the execution
validation AI
of an AI model. of this
model.
this
line the
ing performance.
influence ofThe reason for
production this is mainly
conditions on thetwofold.
end-of- workflow
The obtained
workflow may initiate
may knowledge
initiate new new AI
throughout
AI cycles. the execution of this
First,
line
First, a purely
performance. data-driven
a purely data-drivenThe reason black-box
for
black-box this application
is mainly
application of AI
twofold. is The obtained knowledge
of AI is workflow may initiate new AI cycles. throughout the execution of this
line performance. The reason for thisof isinfluential
mainly twofold. The paper
not feasible,
First,
not a purely
feasible,
First, a purely
because the
data-driven
because large
the large
data-driven
amount
black-box
amount of
black-box application offactors
influentialoffactors
application AI is
AI is The
workflowpaper is organized as follows. Section 2 briefly dis-
mayis organized
initiate newas follows.
AI cycles. Section 2 briefly dis-
requires
not feasible,training
because withthe a large
large data
amount set
of of homogeneous
influential factors cusses
The the
the state-of-the-art
cussespaper is organized as in
state-of-the-art knowledge
follows.
in knowledge Sectiongraphs
graphs for
for man-
2 briefly dis-
man-
requires
not feasible,training
because withthe alarge
largeamountdata set of of homogeneous
influential factors The paper is organized as follows. Section 2 briefly dis-
data.
requires Large data
training sets
with aare not
large available
data set of in a high
homogeneous mix ufacturing.
cusses the
ufacturing. Section
Section 33 presents
state-of-the-art in
presents the industrial
knowledge
the industrialgraphs use
use case
for
case and
man-
and
data. Large
requires data with
training sets aarelarge not dataavailable
set of inhomogeneous
a high mix cusses the state-of-the-art in knowledge graphs for man-
low volume
Large (HMLV)
data setssetting. Experimental
available incampaigns to its context
context at Atlas
Atlas 33Copco Airpower. Section 44 introduces
data.
low
data. volume
Large (HMLV)
data are
are not
setssetting.
are not Experimental
available in
a high mix
campaigns
a expensive.
high mix to ufacturing.
its
ufacturing. Section
at
Section presents
Copco
presents the
the industrial
Airpower. Section use
industrial use case
case and
introduces and
obtain
low such
volume data
(HMLV) often
setting. infeasible
Experimental and too
campaigns to the
its
the knowledge
context
knowledge at graph
Atlas
graph architecture,
Copco Airpower.
architecture, a concrete
Section
a Section 4 realization
introduces
concrete4 realization
obtain
low such (HMLV)
volume data are setting.
often infeasible
Experimental and too expensive.
campaigns to its context at Atlas Copco Airpower. introduces
Second, AI systems are need
often to be using an
an ontology-based tool
tool stack,
stack,aaand how
how it was
was used
obtain
Second,such
obtain suchAI data
systems
data are need
often to be more
infeasible
infeasible more andtransparent
and
too expensive.
transparent
too expensive.
and
and the usingknowledge
the knowledge
graph architecture,
ontology-based
graph architecture,
concrete
and
concrete itrealization
used
realization
explainable,
Second, in
AI systems order to be trustworthy for production- in the
using use
an case. Section
ontology-based 5 discusses
tool stack, the
and added
how it value
was used and
explainable,
Second, AI in order need
systems to beto
need to
be more transparent
trustworthy
be more for production-
transparent
and in the use case. Section 5 discusses the added value
and using an ontology-based tool stack, and how it was and
used
wide decision
explainable, insupport.
order toThere
be is a need
trustworthy for
for white-box
production- AI the
in
the challenges
the use
challenges case. of
of this
Section
this tool
5
tool chain,
discusses
chain, and
and the Section
added
Section 6
6 presents
value
presentsand
wide decision
explainable, support.
insupport.
order There
toThere is
be trustworthya need for
for white-box
production- AI in the use case. Section 5 discusses the added value and
techniques,
wide decision that are understandable. is a In
need thefor current
white-box AI ap-
AI the final conclusion.
challenges of this tool chain, and Section 6 presents
techniques,
wide decision that are understandable.
support. There is a need In the for current
white-box AI the
AI ap- the final conclusion.
challenges of this tool chain, and Section 6 presents
techniques, that are understandable. In the current AI ap- the final conclusion.
techniques,
⋆ This research that are understandable. In the current AI ap- the final conclusion. 2.
was supported by Flanders
⋆ This research was supported by Flanders Make, the strategic
⋆
Make, the strategic
2. RELATED
RELATED WORK WORK
research
This center
research for
was the manufacturing
supported by industry.
Flanders Make, This thepaper was
strategic 2. RELATED WORK
research
⋆ This center for
research was the manufacturing
supported by industry.
Flanders Make, Thisthepaper was
strategic 2. RELATED WORK
partially funded by the
the DTDesign ICONindustry.
(FlandersThis Innovation
paper and Ontology
research
partially
research
center for
funded
center
Entrepreneurship
by
for the
the
FM/ICON
manufacturing
DTDesign
manufacturing
::
ICON
HBC.2019.0079)
(Flanders
industry. Innovation
This
project. paper
This
was
and
was
paper Ontology based knowledge graphs are
based knowledge graphs are notnot used
used a a lot
lot inin
partially funded
Entrepreneurship
partially funded
by the
FM/ICON
by
DTDesign
thethe
DTDesign
ICON
:: HBC.2019.0079)
ICON
(Flanders
(Flanders
Innovation
project. and
This paper
Innovation and
manufacturing
Ontology
manufacturing based yet.
yet. Although,
knowledge
Although, there
graphs
there are has
hasnot been
used
been some
a
some lot re-
in
re-
is partially
is partially
funded
Entrepreneurship
funded
by
FM/ICON
by the
European
::
European
Commission
HBC.2019.0079)
Commission
under
project.
under
the frame-
This
the paper
frame-
Ontology
search based
formalizing
manufacturing knowledge
yet. manufacturing
Although, graphs are
knowledge
there hasnot used
in
been a lot
ontologies,
some in
re-
Entrepreneurship
work
is program
partially FM/ICON
Horizon
funded by 2020,
the ::grant
HBC.2019.0079)
European agreement
Commission project.
number
under This paper
101000165,
the frame- search formalizing
manufacturing yet. manufacturing
Although, knowledge
there has in ontologies,
been some re-
for example
search in
in the
formalizing context
context of
themanufacturing of design Jinzhi
Jinzhi et
knowledge in al.
al. (2020),
in ontologies,
work
is program
partially Horizon
funded by the2020, grant agreement
European Commission number
under 101000165,
the frame-
project
work name
program ASSISTANT
Horizon 2020,(Beldiceanu et al. (2021)).
grant agreement number 101000165, for example
search formalizing manufacturing design
knowledge et (2020),
ontologies,
project
1
work name ASSISTANT
program
European Horizon
Commission 2020,(Beldiceanu
grant
(2018).
et al. (2021)).
agreement
A definitionnumber of 101000165,
Artificial quality
for example
quality control
controlin Leitão
the
Leitão et
context al.
et al.of (2012),
design
(2012), Jinzhianomaly
Jinzhi
anomaly et detection
al. (2020),
detection
project name ASSISTANT
1 European Commission (Beldiceanu
(2018). A et definition
al. (2021)). of Artificial for example in the context of design et al. (2020),
project namemain
Intelligence:
1 European ASSISTANT
capabilities
Commission (Beldiceanu
and scientific
(2018). A et definition
al. (2021)). of
disciplines. https://ec.
Artificial
Steenwinckel
quality
Steenwinckelcontrolet et al.
Leitão
al. (2018),
et al. and
(2018), (2012),
and especially
anomaly
especially also
also in
in the
detection the
Intelligence:
1
1 European main capabilities(2018).
Commission and scientific
A
europa.eu/futurium/en/system/files/ged/ai_hleg_definition_
disciplines. of
definition https://ec.
Artificial quality
context
Steenwinckelcontrol
of Industry
et Leitão
al. et
4.0, al.
see
(2018), (2012),
e.g.
and Sampath anomaly
especially Kumar
also detection
inet al.
the
Intelligence: main capabilities and scientific disciplines. https://ec.
europa.eu/futurium/en/system/files/ged/ai_hleg_definition_ context
Steenwinckel of Industry
et al. 4.0, see
(2018), e.g.
and Sampath
especially Kumar
also inet al.
the
Intelligence: main capabilities
of_ai_18_december_1.pdf. Lastandaccessed:
scientific2022-02-09
europa.eu/futurium/en/system/files/ged/ai_hleg_definition_
of_ai_18_december_1.pdf. Last accessed: 2022-02-09
disciplines. https://ec. (2019);
context Kourtis
of
(2019); Kourtis Industryet al. (2019);
4.0,
et al.4.0, see
(2019); Giustozzi
e.g. Sampath
Giustozzi et al. (2018);
Kumar
et al.Kumar
(2018); et Saha
et
Sahaal.
europa.eu/futurium/en/system/files/ged/ai_hleg_definition_
of_ai_18_december_1.pdf. Last accessed: 2022-02-09
context of Industry see e.g. Sampath
(2019); Kourtis et al. (2019); Giustozzi et al. (2018); Saha al.
2405-8963 Copyright © 2022 The Authors. This is an open access article under the CC BY-NC-ND license.
of_ai_18_december_1.pdf. Last accessed: 2022-02-09 (2019); Kourtis et al. (2019); Giustozzi et al. (2018); Saha
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2022.09.361
14 Bart Meyers et al. / IFAC PapersOnLine 55-10 (2022) 13–18

et al. (2019). Furthermore, there are ongoing efforts for starting up the CNC machine or after a changeover
creating formal ontologies for industry, such as IOF, see but may be stable after several parts have been pro-
Kulvatunyou et al. (2018), or the Ontocommons project 2 . duced, or when a certain pace can be maintained.
Other research shows how ontologies can be used in knowl- Machine bed temperature, coolant temperature, or
edge graphs Grangel-González (2019); Yahya et al. (2021). ambient temperature may be an influencing factor.
Some practical applications of ontologies in knowledge Gaining such insights is a collaborative effort: they
graphs for manufacturing do also exist, see Kalayci et al. are inspired by production engineers’ experience, and
(2020); Kharlamov et al. (2015); Liebig et al. (2019); need to be analyzed by data scientists, relying on
Williams et al. (2015), however, there does not seem to be data managed by data engineers. In this use case,
widespread use yet. In Hildebrandt et al. (2020) ontologies a descriptive analysis is conducted to determine the
are used for interoperability between systems in manu- correlation between temperatures and quality read-
facturing, based on a number of standards. In this work ings by generating several plots.
however, we focus supporting analysis of historic data. The (2) Reduce costs: use an adaptive measurement strategy
architecture used in this paper, is partially based on the that decreases the number of test measurements and
work in Lietaert et al. (2021) and Oakes et al. (2021), therefore reduces the cost while maintaining quality.
and specifically adapted for the Atlas Copco Airpower use The adaptive measurement strategy makes the num-
case. It allows to use an ontology based knowledge graph ber of CNC, CAQ and CMM measurements dynamic
approach for improving the analysis on this manufacturing when geometric variance is expected to be low, de-
example. The knowledge graph is used to build a digital pending on the influential factors. After determining
twin in the context of data analytics. Digital twins and these influential factors, a model is created to predict
big data analytics have a huge potential, also, as can be the quality (pass or fail the tolerance checks) of a part
found in My (2021), in the context of CNC machining. The based on temperature data. For this purpose, we use
study shows that AI and digital twins have huge potential, both point readings (momentary values) and addi-
shown by some one-off success stories, but that the self- tionally the first order derivatives of the temperature
optimization of process parameters at a cognitive level is readings. This derivative represents the momentary
not yet realized. The design and creation of digital twins in change of a temperature reading between two data
industry, based on semantic information models, has also points measured in absolute units (°C). This part
been documented, e.g., Hänel et al. (2020); Caesar et al. includes designing, training and implementing an AI-
(2020); Botkina et al. (2018); Rozanec et al. (2021); Jilg based predictive classification model. Next, AI models
and Stuckart (2021); Jiang et al. (2021). While we make will need to be developed that dynamically suggest
use of a similar type of information model in our work, the minimal measurement ratio to maintain quality
we specifically focus on a tool chain to make this data, standards. For example, more measurements may be
information and knowledge efficiently available to a data necessary after starting up the CNC machine or after
scientist. a change over, but measurements may be drastically
reduced once the machine is producing stable results.
3. ATLAS COPCO USE CASE The source data for this use case is stored in different
data sources. Built-in temperature sensors generate the
The Atlas Copco use case is based on the production of temperature readings in a tabular format with timestamp
a compressor air end, the core element of a compressor. and value. The quality measurements, including thresholds
This is a HMLV production, and many different variants are in a csv format. A part is within tolerance if all of
(up to 30), each consisting of about 20 components need its measurements are within their respective tolerances.
to be produced. The current use case focuses on a single Additionally, another tabular data set exists that counts
production step that involves a CNC (computer numerical all machined parts on the machine (the so-called cnc
control) machine to machine compressor air end hous- counter table). This table contains a part identifier, start
ings. CNC machines can precisely process parts based time, and stop time of the machine program. This table
on computerized instructions, which can be in the form serves as the link between the temperature readings and
of programming code or CAD models. Several different the quality measurement.
cutting tools are used to machine specific areas of the
raw material. These tools include mills, drills, taps, etc. The following data processing transformations are per-
Concrete goals of the use case: formed in order to obtain the consolidated data set as used
in the experiment:
(1) Gain insights: determine and explicitly model what
correlations exist with the geometric quality of a • data type parsing (e.g., text to datetime);
housing. Quality is measured for a fixed number of • data cleaning (known outlier values);
parts in one or more ways: (1) on the CNC machine it- • data filtering (e.g., a certain part type number);
self (basic checks), (2) by CAQ manual measurements Atlas Copco is transitioning in a smart and digital factory.
measuring geometric features (e.g., hole diameters), This results in more data being generated in all levels
and (3) on a CMM (coordinate measuring machine) of the organization, including the shop floor. In order to
where a multitude of features are measured (e.g., po- efficiently carry out the use case in such a setting, a number
sition rotor bore male outlet). If one of the measure- of challenges arise:
ments is outside of the tolerance interval, the housing
fails the quality test. Quality may be unstable after • Heterogeneous data. Data are stored in different
databases, in different formats, with different schema,
2 https://www.ontocommons.eu/, last accessed: 2022/02/09
 Bart Meyers et al. / IFAC PapersOnLine 55-10 (2022) 13–18 15

naming conventions, access credentials, etc. These and/or deploying insights as a streaming digital twin (e.g.,
data must be connected to perform complex analysis; as dashboard). Given the complexity of data, information
• Ad hoc analysis. Because the most significant in- and knowledge, there is a need for an “intelligent” digital
fluential factors are not known beforehand, varying shadow of the production system. For this reason, we
hypotheses must be investigated quickly. This means propose a design loop involving a cognitive digital twin
that different data must be efficiently retrieved. In (top of Figure 1). Whereas a conventional digital shadow
large organizations, retrieving a data set may take is a reflection of the physical asset by recording data and
weeks as it involves requests to data engineers, and maintaining models of its behaviour, this cognitive digital
long communication loops about what data is re- twin exceeds this notion by serving the following main
quired and how data should be interpreted; purposes for the data scientist workflow:
• (Tacit) knowledge. Data scientists must be able to tap
into production knowledge by production engineers, • providing an abstraction layer for data. Since data
and data knowledge by data engineers. A lack of are heterogeneous, up to 80% of the data scientist’s
common understanding of this often tacit knowledge time is spent on retrieving data, rather than on their
results in missed opportunities or wrong conclusions. core tasks. Our approach provides a single interface
to access all data in a uniform way;
4. KNOWLEDGE GRAPH-BASED ANALYSIS • formal modeling of data, information and
knowledge in a knowledge graph. Data, information
To improve leveraging existing data, information and and knowledge is modeled according to the problem
knowledge in Atlas Copco’s Intelligence Hub, we have domain, rather than the technical solution or rep-
evaluated the use of a cognitive digital twin centered resentation of data. Exploration and querying tools
around a knowledge graph. In this section, we introduce allow data scientists to acquire data, information and
the conceptual architecture, and a concrete realization and knowledge about domain concepts such as part, ma-
usage for the Atlas Copco use case. chine, operation, etc., and allow common understand-
ing of the domain between data engineers, production
4.1 Knowledge Graph Architecture engineers and date scientists;
• reasoning about knowledge. As knowledge is
Figure 1 shows the overview of the approach. The physical made explicit and is formalized in a knowledge graph,
system is shown on the left side. As we are interested this opens up possibilities of (automated) reasoning
in a holistic analysis, e.g., linking product quality to to generate new knowledge. While there are several
production circumstances, the physical twin comprises the possibilities (e.g., ontological reasoning, translation
production system and the products. A digital twin in to Bayesian networks, etc.) we only explored manual
the traditional sense Kritzinger et al. (2018) is shown in reasoning in this use case.
the center bottom, where the physical system is directly Therefore, our conceptual architecture revolves around a
observed, analysed, and influenced in an operational loop. knowledge graph as single access point for all data,
A major challenge in complex production environments is information and knowledge. The knowledge graph consists
how to design such a digital twin. How can insights on your of a domain model, that defines how all domain concepts
physical system be found and turned in a digital twin that are structured, a high level query API that allows
creates added value? As shown at the right side of Figure 1, the user (data scientist) to access data, information and
this requires efficient data science workflows for explor- knowledge in terms of the concepts defined in the domain
ing and understanding data, information and knowledge, model, and a mapping from this domain model to the
extracting this data, analyzing it (e.g., find correlations), different sources of information (databases, models, etc.).
creating insights (e.g., AI-based prediction), updating the
knowledge graph with obtained insights (e.g., correlation),

Physical twin TypeInfo Knowledge Data Scientist Workflow

MES e.a.
(production & product)

explore
query API

Consoli-
mapping

Raw Curated Explore

dated
information
query select ?serialNumber ?timestamp ?temp {
?housing a :CompressorHousing;

record knowledge
Extract
:idCompressorHousing ?serialNumber .
?cnc a :CNCHousingProcessStep;
:treatedCompressorHousing ?housing;
Gateway

Cognitive Digital Twin graph :machineBedTemperatureTS ?timepoint .

Simulations update information ?timepoint

dt:valueTimestamp ?timestamp;
dt:valueDouble ?temp .
}
order by ?serialNumber ?timestamp

Streaming Analyze
observe information
observe Analytics

influence Create
Dashboarding insight
/ Alerting deploy
Streaming Digital Twin

Fig. 1. Overview of the approach.

16 Bart Meyers et al. / IFAC PapersOnLine 55-10 (2022) 13–18

Fig. 3. Visualisation of a part of the knowledge graph for

the Atlas Copco use case.
Fig. 2. Realized architecture for the Atlas Copco case
• Link domain concepts to data. This can be done in
based on an ontology tool stack.
the data lake using curation techniques, and/or in an
OBDA mapping in Ontop as shown in Section 4.3.
In order to make our architecture usable in an industrial
context, we identified the following requirements: 4.2 Domain modeling
• virtual integration with existing data sources, rather
than hard curation to the knowledge graph structure; The domain model needs to define concepts and relations
• cloud-based deployment of the cognitive digital twin, in the problem domain. This is in contrast to the current
the knowledge graph and its explore and query access; state-of-the-practice, where the user interacts with the
• scalability in terms of data retrieval. There should technical solution domain where he needs to take into
not be a query execution penalty when introducing consideration data format, location, credentials, different
the approach; interfaces, implicit links between databases, etc.
• incremental adoption of the approach. It is not fea-
A part of the domain model, modeled as ontology, is visu-
sible to first create a complete cognitive digital twin
alised in Figure 3. While approaches exist to generate an
before being able to use it. Instead it should be possi-
ontology from a database schema, we argue that a domain
ble to gradually create a cognitive digital twin, while
model is best created by hand to reflect well the problem
reaping its benefits during this transition period.
domain, so that it is not biased by the technical data
The experimental setup and cognitive digital twin for this structure in the solution domain. The knowledge engineer
use case are shown in Figure 2. An experimentation sand- can start from a generic model for manufacturing (drawing
box is provided by Atlas Copco, which is accessible for re- from existing standards, and including generic domain
search. The setup is designed to be as similar as possible to concepts like Asset, ProcessStep and TimePoint), which
Atlas Copco’s production environment. A Databricks envi- we are still developing. In the domain model, ambient,
ronment is set up for data analysis. Alongside, a fragment coolant and machine bed temperature are time series data
of the production data is copied into the experimental envi- that originate from sensors that continuously record data.
ronment in a Databricks Filestore. Instead of accessing the In the solution domain, the temperature for a certain CNC
data directly using low-level SQL queries, the user (data machining activity can be retrieved by querying tempera-
scientist) is able to access data, information and knowledge tures between the start and finish timestamp of the ma-
from an ontology-based digital twin through the query API chining activity. However, in the problem domain, the cog-
(a SPARQL endpoint) in the Databricks environment. The nitive concept is simply temperature of a CNC machining
domain model is modeled as an ontology, and this digital activity. Therefore, the domain model in Figure 3 hides the
twin includes a tool stack including Protégé for creating technical dependency on timestamps, and explicitly links
and editing the knowledge graph, Ontop for mapping the CNCHousingProcessStep to MachineBedTemperature.
ontology to databases according to the OBDA principles,
and WebVOWL to visualise the ontology. We use several 4.3 Mapping to data
extensions of this tool stack to cater for specific function-
alities, such as GraphDB to include the ability to write In order to access existing data through the knowledge
to the knowledge graph, and RDF4J for federating the graph, the domain model must be mapped to the data.
writable part and the part of the knowledge graph that How to technically establish such a mapping is a highly
links to the data. complex problem given the heterogeneity of data in man-
ufacturing. In order to tackle this complexity, our architec-
Next to the data scientist workflow shown in Figure 1,
ture defines a framework that separates three orthogonal
the architecture requires a knowledge engineering workflow
aspects of the data: semantic data type, data persis-
to build the knowledge graph. The role of “knowledge
tence format, and data curation strategy.
engineer”, who is responsible for the workflow, needs to
be assigned. A knowledge engineer’s main tasks are: The semantic data type categorizes data in what their val-
ues semantically represent. We encountered the following
• Define and connect domain concepts in the ontology,
in the use case:
e.g., compressor housing, CAQ quality control step,
machine bed temperature, etc. This is detailed in • non-timed relational data, e.g., CAQ quality;
Section 4.2. • geometric data, e.g., CMM quality data;
 Bart Meyers et al. / IFAC PapersOnLine 55-10 (2022) 13–18 17

• event based data, e.g., CNC processing logs; especially important when going beyond the classification-
• time series data, e.g., temperature sensor values. based use case with binary result (within or outside of
tolerance) presented in this paper, to e.g., optimization of
Each of the above can be stored in different persistence
machine settings or prediction of performance values.
formats, such as relational databases, 3D CAD files, time
series databases, text files, etc. Some persistence mecha- Reasoning about knowledge. Although still a manual
nisms are specialized in specific kinds of semantic data: process, we were able to improve the reasoning process by
PostgreSQL is well suited for non-timed relational data, querying the knowledge graph and using the resulting data
while InfluxDB focuses on time series data. Often, an or- in data analysis tools like Keras.
ganization attempts to decrease the number of persistence
formats as a measure to deal with the heterogeneity. Existing data source. We were able to access all data
virtually in the use case, while still mapping the domain
Data curation strategy can be split up into categories: hard model correctly to the data. Therefore, no additional data
curation, virtual curation and referencing. Hard curation curation pipelines need to be set up. Instead, new data is
copies all data into the knowledge graph. This may be readily available.
preferred when query execution performance is essential.
Virtual curation keeps data in its original data persistence Cloud-based tooling. The cloud-based cognitive digital
format, and retrieves it when querying the knowledge twin is accessible through a query API and a visualization
graph. This is less intrusive and avoids unnecessary data interface with WebVOWL. The queries can be executed in
duplication and maintenance. The referencing curation cloud-based environments like Databricks.
strategy does not retrieve data when querying the knowl- Query execution. Queries in the use case were executed
edge graph, but returns a reference (query, location, code) within seconds, which is acceptable for ad hoc analysis.
that can be executed to obtain the actual data. This is a
suitable strategy when the less intrusive nature of virtual Incremental adoption. We were able to incrementally
curation is required, while performance is key: query and create the knowledge graph, and use low-level queries in
data retrieval are effectively decoupled. combination with SPARQL queries in Databricks during
development.
In our architecture, the combination of all three aspects
of the data dictates how the mapping between data and 5.2 Challenges
domain model can be achieved. A knowledge graph may
then comprise several different types of mappings together, Tool stack for the knowledge engineering workflow.
to fit well the heterogeneous data sources. Our ontology- The tool stack we use with Protégé, Ontop and VOWL
based stack allows linking to several other types of data, is not suitable out-of-the-box for industrial adoption as
images (e.g., by returning its reference), or simulation a knowledge graph for manufacturing. We identify the
models (e.g., through a user defined function calling a following improvements: (1) need for a (visual) exploration
REST-api). This is beyond the scope of the paper. tool specific to the domain model, (2) improved IDE for
creating mappings to data. Additionally, mechanisms for
5. LESSONS LEARNED integration with existing tool chains, production engineer-
ing workflows, security, etc. will need to be defined in order
In this section, we reflect on the lessons learned of using a to be usable in industry.
knowledge graph in an industrial use case. Query execution performance. While query execution
was reasonable in this use case, their low-level counterparts
5.1 Potential for Added Value took considerably less time to execute. It needs to be inves-
tigated under which conditions this may be problematic,
Abstraction layer for data. The knowledge graph and how to deal with this (e.g., by using a referencing data
dramatically improves ad hoc analysis, i.e., the iterative curation strategy).
search process for finding correlations and relevant features
Quality of data. Data quality remains an inherent con-
in data. By querying the knowledge graph using a language
cern in industry. Data may be lost during translation,
in the problem domain, the user of the knowledge graph is
may not exist, or may be hard to understand. While
shielded from the technical complexity and heterogeneity
the knowledge graph may help by explicitly modeling the
of the underlying data, leaving more time for analysis.
problem domain, it needs to be investigated how data is
Formal modeling of data, information and knowl- best prepared (cleaned, curated, as shown in Figure 1), so
edge. The knowledge graph models data, information and that it can be optimally mapped to the knowledge graph.
knowledge in the problem domain, including insights on
Guidelines on data mapping. While we present differ-
the data (see the running example on relating temperature
ent possibilities for data mapping depending on the nature
to a CNC machining activity) or data analysis results (fea-
of the data, it needs to be investigated what techniques are
tures on data, correlations, uncertainties, models, etc.). We
best used in which situations.
are establishing generic domain models and workflows to
automate updating the knowledge graph with acquired in- Systematic modeling of knowledge. We intend to
sights. Whereas before, data insights were spread through draw from existing standards for manufacturing (e.g.,
reports, notebooks or presentations, we are able to make SOSA, ISA-95) to improve the upper ontology for manu-
these insights queryable from a single source. This allows facturing, that is the starting point for creating an organi-
us to more easily and effectively evaluate discrepancies zation’s domain model. Furthermore, we aim to include an
between the physical twin and the digital twin. This will be explicit notion of experiments (physical experimentation,
18 Bart Meyers et al. / IFAC PapersOnLine 55-10 (2022) 13–18

data science experiment, production as experiment), as ence on Automation Science and Engineering (CASE), 758–763.
well as a notion of uncertainty that is inherent to the doi:10.1109/CASE49439.2021.9551554.
results of data experiments (e.g., probabilities, predictions, Jilg, D. and Stuckart, N. (2021). Conception and implementation
of a digital twin using an enterprise knowledge graph platform.
estimations). We also aim to explicitly model the evolution
doi:10.13140/RG.2.2.17553.92004.
of the knowledge graph, to be able to query its status at a Jinzhi, L., Junda, M., Zheng, X., Wang, G., and Kiritsis, D. (2020).
certain moment in time, and to support what-if analysis. Design ontology supporting model-based systems-engineering for-
Furthermore, we need to integrate our knowledge engineer- malisms.
ing workflow with production engineering workflows where Kalayci, E.G., González, I.G., Lösch, F., Xiao, G., ul Mehdi, A.,
knowledge is used and insights are acquired. Kharlamov, E., and Calvanese, D. (2020). Semantic integration
of bosch manufacturing data using virtual knowledge graphs. In
6. CONCLUSIONS AND FUTURE WORK Proc. of the 19th Int. Semantic Web Conf. (ISWC 2020), volume
12507 of Lecture Notes in Computer Science, 464–481. Springer.
Kharlamov, E., Hovland, D., Jiménez-Ruiz, E., Lanti, D., Lie, H.,
We presented a use case at Atlas Copco which we consider Pinkel, C., Rezk, M., Skjæveland, M., Thorstensen, E., Xiao, G.,
to represent well the generic transition to Industry 4.0, Zheleznyakov, D., and Horrocks, I. (2015). Ontology based access
both as a representative organisational context, as well as to exploration data at statoil. 93–112.
a representative use case itself. To tackle such problems, Kourtis, G., Kavakli, E., and Sakellariou, R. (2019). A rule-based
we introduce a conceptual architecture for cognitive digital approach founded on description logics for industry 4.0 smart
twins based on knowledge graphs. We evaluated the use factories. IEEE Trans. Ind. Informatics, 15(9), 4888–4899.
of an implementation based on ontologies in the use case. Kritzinger, W., Karner, M., Traar, G., Henjes, J., and Sihn, W.
The main conclusion that there is a potential for important (2018). Digital twin in manufacturing: A categorical literature
review and classification. IFAC-PapersOnLine, 51(11), 1016–
added value, but there are technical and tooling challenges
1022.
that need to be tackled to allow industrial adoption of Kulvatunyou, B., Wallace, E., Kiritsis, D., Smith, B., and Will,
knowledge graphs. C. (2018). The industrial ontologies foundry proof-of-concept
In future work, we will focus on modelling and reasoning project. APMS 2018 International Conference Advances in Pro-
duction Management Systems (APMS 2018).
on knowledge on the Atlas Copco case, to further improve
Leitão, P., Rodrigues, N., Turrin, C., Pagani, A., and Petrali, P.
the potential of using domain models. Furthermore, we (2012). Grace ontology integrating process and quality control. In
will focus on analyzing how to apply our approach (see IECON 2012 - 38th Annual Conference on IEEE Industrial Elec-
Figure 1 to industry-ready tools, and the trade-off between tronics Society, 4348–4353. doi:10.1109/IECON.2012.6389189.
the acceptable workflows and performance such tools may Liebig, T., Maisenbacher, A., Opitz, M., Seyler, J.R., Sudra, G., and
bring on the one hand, and the penalty with respect to the Wissmann, J. (2019). Building a knowledge graph for products
functionality of our approach on the other hand. and solutions in the automation industry. In KGB@ESWC.
Lietaert, P., Meyers, B., Van Noten, J., Sips, J., and Gadeyne, K.
REFERENCES (2021). Knowledge graphs in digital twins for ai in production. In
APMS 2021, 249–257. Springer International Publishing.
Beldiceanu, N., Dolgui, A., Gonnermann, C., Gonzalez-Castañé, G., My, C. (2021). The Role of Big Data Analytics and AI in Smart
Kousi, N., Meyers, B., Prud’homme, J., Thevenin, S., Vyhmeister, Manufacturing: An Overview, 911–921. doi:10.1007/978-981-15-
E., and Östberg, P.O. (2021). Assistant: Learning and robust 7527-3 87.
decision support system for agile manufacturing environments. Oakes, B.J., Meyers, B., Janssens, D., and Vangheluwe, H. (2021).
IFAC-PapersOnLine, 54, 641–646. Structuring and accessing knowledge for historical and streaming
Botkina, D., Hedlind, M., Olsson, B., Henser, J., and Lundholm, T. digital twins. In ODCM-DT 2021, volume 2941 of CEUR Work-
(2018). Digital twin of a cutting tool. Procedia CIRP, 72, 215– shop Proceedings.
218. doi:https://doi.org/10.1016/j.procir.2018.03.178. 51st CIRP Rozanec, J.M., Lu, J., Rupnik, J., Skrjanc, M., Mladenic, D.,
Conference on Manufacturing Systems. Fortuna, B., Zheng, X., and Kiritsis, D. (2021). Actionable
Caesar, B., Hänel, A., Wenkler, E., Corinth, C., Ihlenfeldt, S., and cognitive twins for decision making in manufacturing. CoRR,
Fay, A. (2020). Information model of a digital process twin for abs/2103.12854.
machining processes. In 2020 25th IEEE International Conference Saha, S., Usman, Z., Li, W., Jones, S., and Shah, N. (2019). Core
on Emerging Technologies and Factory Automation (ETFA), domain ontology for joining processes to consolidate welding
volume 1, 1765–1772. doi:10.1109/ETFA46521.2020.9212085. standards. Robotics and Computer-Integrated Manufacturing, 59,
Giustozzi, F., Saunier, J., and Zanni-Merk, C. (2018). Context 417–430. doi:10.1016/j.rcim.2019.05.010.
modeling for industry 4.0: an ontology-based proposal. Procedia Sampath Kumar, V.R., Khamis, A., Fiorini, S., Carbonera, J.L.,
Computer Science, 126, 675–684. Olivares Alarcos, A., Habib, M., Goncalves, P., Li, H., and Ol-
Grangel-González, I. (2019). A Knowledge Graph Based Integration szewska, J.I. (2019). Ontologies for industry 4.0. The Knowledge
Approach for Industry 4.0. Ph.D. thesis. Engineering Review, 34, e17.
Hildebrandt, C., Köcher, A., Küstner, C., López-Enrı́quez, C.M., Steenwinckel, B., Heyvaert, P., De Paepe, D., Janssens, O., Hautte,
Müller, A.W., Caesar, B., Gundlach, C.S., and Fay, A. (2020). S., Dimou, A., De Turck, F., Hoecke, S., and Ongenae, F. (2018).
Ontology building for cyber–physical systems: Application in Towards adaptive anomaly detection and root cause analysis by
the manufacturing domain. IEEE Transactions on Automation automated extraction of knowledge from risk analyses.
Science and Engineering, 17(3), 1266–1282. doi:10.1109/TASE. Williams, J.W., Cuddihy, P., McHugh, J., Aggour, K.S., Menon, A.,
2020.2991777. Gustafson, S.M., and Healy, T. (2015). Semantics for big data
Hänel, A., Schnellhardt, T., Wenkler, E., Nestler, A., Brosius, A., access amp; integration: Improving industrial equipment design
Corinth, C., Fay, A., and Ihlenfeldt, S. (2020). The development through increased data usability. In 2015 IEEE International
of a digital twin for machining processes for the application in Conference on Big Data (Big Data), 1103–1112.
aerospace industry. Procedia CIRP, 93, 1399–1404. doi:https:// Yahya, M., Breslin, J.G., and Ali, M.I. (2021). Semantic web and
doi.org/10.1016/j.procir.2020.04.017. 53rd CIRP Conference on knowledge graphs for industry 4.0. Applied Sciences, 11(11).
Manufacturing Systems 2020.
Jiang, Y., Chen, C., and Liu, X. (2021). Assembly process knowledge
graph for digital twin. In 2021 IEEE 17th International Confer-