Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
ScienceDirect
Available ScienceDirect
online
Available
Procedia atonline
CIRPwww.sciencedirect.com
at www.sciencedirect.com
00 (2019) 000–000
www.elsevier.com/locate/procedia
ScienceDirect
Procedia CIRP 00 (2019) 000–000
ScienceDirect www.elsevier.com/locate/procedia

Procedia CIRP 00 (2017)

Procedia 000–000
CIRP 81 (2019) 369–374
www.elsevier.com/locate/procedia
52nd CIRP Conference on Manufacturing Systems
52nd CIRP Conference on Manufacturing Systems
Data Analytics for Manufacturing Systems – A Data-Driven Approach for
Data Analytics for 28thManufacturing
CIRP Design Process Systems
Optimization
Conference, May 2018,– A Nantes,
Data-DrivenFrance Approach for
Process Optimization
A new methodology to analyze
Florian Ungermann* a
, Andreas theKuhnle
functional
a
, Nicoleand Strickerphysical
a
, Gisela architecture
Lanzaa of
existing products for
Florian Ungermann*
wbk anofassembly
Institute
a
, Andreas
Production
a
oriented
Kuhnle
Science,
a
product
, Nicole
Kaiserstrasse 12, 76185 Karlsruhe, family
Stricker a
, Gisela identification
Germany Lanzaa
* Corresponding author. Tel.: +49 1523 wbk Institute
fax:of+49
Production
721 608 Science, Kaiserstrasse
address:12, 76185 Karlsruhe, Germany
a
9502624; 45005. E-mail [email protected]
Paul Stief *, Jean-Yves Dantan, Alain Etienne, Ali Siadat
* Corresponding author. Tel.: +49 1523 9502624; fax: +49 721 608 45005. E-mail address: [email protected]
École Nationale Supérieure d’Arts et Métiers, Arts et Métiers ParisTech, LCFC EA 4495, 4 Rue Augustin Fresnel, Metz 57078, France
Abstract
* Corresponding author. Tel.: +33 3 87 37 54 30; E-mail address: [email protected]
Abstract
In the course of digitalization many small and medium-sized companies face the challenge of using the existing database for process optimization
in manufacturing. Furthermore, the demand-oriented expansion of the database is a great challenge. A lack of competencies, limited financial
In the course of digitalization many small and medium-sized companies face the challenge of using the existing database for process optimization
resources and historically grown data structures, which show a strong heterogeneity and lack of transparency, are the central obstacles. A specific
in manufacturing. Furthermore, the demand-oriented expansion of the database is a great challenge. A lack of competencies, limited financial
Abstract
approach, how data analytics projects for process optimization should be carried out in manufacturing, is presented. In particular, the question
resources and historically grown data structures, which show a strong heterogeneity and lack of transparency, are the central obstacles. A specific
which sensors should be implemented to expand the database is answered. The approach is applied exemplarily for a manufacturing line.
Inapproach, how dataenvironment,
today’s business analytics projects for process
the trend towardsoptimization
more product should
varietybeand
carried out in manufacturing,
customization is unbroken.isDue
presented.
to this In particular, the
development, the question
need of
which
agile andsensors should be production
reconfigurable implemented to expand
systems the database
emerged to cope iswith
answered.
various The approach
products and isproduct
appliedfamilies.
exemplarily
To for a manufacturing
design and optimize line.
production
© 2019 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
systems
© 2019 as Thewell as to choose
Authors. Publishedtheby
optimal product
Elsevier
(http://creativecommons.org/licenses/by-nc-nd/3.0/) Ltd. matches, product analysis methods are needed. Indeed, most of the known methods aim to
© 2019
This
analyze The
is aan Authors.
open
product access Published
article
orresponsibility
one product by Elsevier
under the on
family CCthe Ltd. This islevel.
BY-NC-ND
physical an open
license accessproduct
article under the however,
CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
Different families, may differ largely in terms of the number and
Peer-review under of the scientific committee of the 52nd CIRP Conference on Manufacturing Systems.
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
nature of components. This fact impedes an efficient comparison and choice of appropriate product familySystems.
Peer-review under responsibility of the scientific committee of the 52nd CIRP Conference on Manufacturing combinations for the production
Peer-review
system. A new under responsibility
methodology of the scientific
is proposed to analyze committee
existing of the 52nd
products CIRPofConference
in view on Manufacturing Systems.
Keywords: data-analytics; process optimization; manufacturing systems; procedure modelstheir functional and physical architecture. The aim is to cluster
these products in new assembly oriented product families for the optimization of existing assembly lines and the creation of future reconfigurable
Keywords: data-analytics; process optimization; manufacturing systems; procedure models
assembly systems. Based on Datum Flow Chain, the physical structure of the products is analyzed. Functional subassemblies are identified, and
a functional analysis is performed. Moreover, a hybrid functional and physical architecture graph (HyFPAG) is the output which depicts the
1. Introduction
similarity between product families by providing design support to both,challenge. productionItsystem
is associated
plannerswith financialdesigners.
and product expenditures [7] (e.g.
An illustrative
example of a nail-clipper is used to explain the proposed methodology. Anfor industrial case studysensors
additional on two product
and families
data of steering columnsand
preprocessing) of
1. Introduction challenge. It is associated with financial expenditures [7] (e.g.
thyssenkrupp Presta France is then carried out
The ongoing digitalization enables new optimizationto give a first industrial evaluation of the
organizational proposed approach.
©potentials
2017 The Authors. Published by Elsevier B.V. With a growing
for additionalbarriers (e.g. and
sensors series data
releasepreprocessing)
of a manufacturing and
The for
ongoing manufacturing
digitalization companies.
enables new optimization line might
organizationalbe lost, when
barriers additional
(e.g. series sensors
release ofare
a implemented).
manufacturing
Peer-review
number ofunder responsibility
sensors implemented of theinscientific committeesystems
manufacturing of the 28th
a CIRP Design Conference
Therefore, production 2018.engineers have to estimate which
potentials for manufacturing companies. With a growing line might be lost, when additional sensors are implemented).
large amount of data is generated. This data is an systems
number of sensorsDesign
implemented in manufacturing importanta additionally gathered data
Therefore, production could potentially
engineers have to contribute
estimate whichto the
Keywords: Assembly; method; Family identification
resource to maintain competitiveness in volatile
large amount of data is generated. This data is an important global markets intended process optimization before changing
additionally gathered data could potentially contribute to the the production
[1]. However,
resource the existing
to maintain data is often
competitiveness insufficiently
in volatile global used for
markets equipment or resources.
intended process optimization before changing the production
information acquisition and process optimization
[1]. However, the existing data is often insufficiently used for in production There are established process models on how to approach
equipment or resources.
management
1.information
Introduction [2]. Amongand
acquisition manufacturing
process optimizationcompanies only 5.5%
in production data-analytics
of the product
There projects,
range andsuch
are established as the CRISP-DM
characteristics
process models on howmodel
manufactured [8, 9].
and/or
to approach
of the available database is used for
management [2]. Among manufacturing companies only 5.5% process optimization However,
data-analytics projects, such as the CRISP-DM model [8,and
assembled these
in this models
system. are
In neither
this backed
context, the by
main practical
challenge in
9].
applying
Due todata-analytics
the fast [1]. Central
development
of the available database is used for process optimization reasons
in the for
domainthis are
of production-related
modelling and methods
analysis is and
now tools,
not nor
only
However, these models are neither backed by practical and answer
to cope the
withquestion
single
historicallydata-analytics
communication
applying grown
and data structures,
an ongoing
[1]. trendwhich
Central reasons show
of digitization
for thisa large
and
are of the potential
products, a limited
production-related of additional
product range
methods data.tools,
and Furthermore,
or existing thethe
product
nor answer design of a
families,
question
heterogeneity
digitalization, and lack
manufacturingof transparency,
enterprises
historically grown data structures, which show a large limited
are data
facing access
importantand problem-specific
but also to be able adequate
to analyze database,
and to which
compare integrates
products
of the potential of additional data. Furthermore, the design of a not
to only
define
missing product
challenges
heterogeneity in and tracing
today’s information
lack ofmarket [3,limited
environments:
transparency, 4]. Furthermore,
a continuing
data access and the sensor
new data but
product
problem-specific also
families. the knowledge
It can
adequate be observed
database, of manufacturing
whichthatintegrates
classical notprocess
existing
only
data
tendencysets used
towards for data-analytics
reduction of are
product often not
development
missing product tracing information [3, 4]. Furthermore, the of sufficiently
times and experts,
product is not
families areexplicitly
regrouped formulated
in function [10].
of
sensor data but also the knowledge of manufacturing process The
clients or approach
features.
high sets
shortened
data quality
usedand
product itdata-analytics
is difficult
forlifecycles. forarecompanies
In addition, often notisto
there anassess
of their
increasing
sufficiently presented
However,
experts, is in this
not paper
assembly addresses
oriented
explicitly productthese
formulated deficits.
[10].areAThe
families methodology
hardly to find.
approach
problem-specific
demand of adequacy
customization, being[5, at 1].
the The
same
high quality and it is difficult for companies to assess their demand-oriented
time in a global is developed,
On the productwhich can
family be
level,applied as
products a guideline
differ
presented in this paper addresses these deficits. A methodology mainlyfor indata-
two
enhancement
competition of
with databases
competitors drivenall by
over
problem-specific adequacy [5, 1]. The demand-orienteda specific
the use
world. case
This in order
trend, analytics
main projects
characteristics: (i) within
the number process
of optimization
components
is developed, which can be applied as a guideline for data- and (ii) of
the
to
whichachieve
is a higher
inducing quality
the level
development and amount
from
enhancement of databases driven by a specific use case in order of data
macro is
to a great
micro manufacturing
analytics projects within process optimization ofa
type of components systems.
(e.g. Based
mechanical, on expert
electrical, knowledge,
electronical).
markets, results in diminished lot sizes
to achieve a higher quality level and amount of data is a greatdue to augmenting Classical methodologies
manufacturing systems. Based considering on mainly
expert single productsa
knowledge,
product varieties (high-volume to low-volume production) [1]. or solitary, already existing product families analyze the
2212-8271 © 2019 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
To cope with this augmenting variety as well as to be able to
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
product structure on a physical level (components level) which
identify
2212-8271 possible
© 2019 The optimization
Authors. Published potentials
by Elsevier in
Ltd. the
This is existing
an open
Peer-review under responsibility of the scientific committee of the 52nd CIRP Conferenceaccess causes on difficulties
article under the CC BY-NC-ND
Manufacturing regarding
Systems.license an efficient definition and
(http://creativecommons.org/licenses/by-nc-nd/3.0/)
production system, it is important to have a precise knowledge comparison of different product families. Addressing this
Peer-review under responsibility of the scientific committee of the 52nd CIRP Conference on Manufacturing Systems.
2212-8271 © 2019 The Authors. Published by Elsevier Ltd.
This is an©open
2212-8271 2017access article Published
The Authors. under theby CC BY-NC-ND
Elsevier B.V. license (http://creativecommons.org/licenses/by-nc-nd/3.0/)
Peer-review
Peer-review under
under responsibility
responsibility of scientific
of the the scientific committee
committee of the of theCIRP
28th 52ndDesign
CIRPConference
Conference2018.
on Manufacturing Systems.
10.1016/j.procir.2019.03.064
370 Florian Ungermann et al. / Procedia CIRP 81 (2019) 369–374
2 Florian Ungermann et al. / Procedia CIRP 00 (2019) 000–000

selection of additional sensors is part of the approach. could be implemented. Bahrepour [19] adds a step called “data
Furthermore, the evaluation and selection of process validation” between the phases of data preprocessing and
optimization measures are included. modeling of the classical CRISP-DM approach, to avoid the
The paper is structured as follows: the next section deals input of faulty data into the data mining model. Nevertheless,
with existing process models for implementing data-analytics the validation of the database is only based on an expert
projects, approaches for expert-based process modelling and discussion and there is no description of how and which
Key Performance Indicators (KPIs) for the evaluation of measures could be taken.
optimization measures. In Section 3 the developed approach is Even though the need for evaluation and adjustment of an
presented. An exemplary application using the example of a existing database is recognized, a design of a problem-specific
real manufacturing line is given in Section 4. The last section adequate database and also considering manufacturing expert
provides a summary. knowledge is not explicitly formulated in existing approaches.

2. State of the Art 2.2. Approaches for Expert-Based Process Modelling

An overview of existing process models for data-analytics Manufacturing processes are naturally characterized by
projects is presented subsequently. Furthermore, approaches complex cause-effect relationships and a not negligible
for expert-based process modeling are discussed. In Section 2.3 underlying degree of uncertainty. The control and monitoring
KPIs are considered as a basis for evaluating process of these processes are therefore often highly dependent on
improvement measures. extensive expert knowledge. Besides the process data, which is
gathered by sensors, the human experience is of equal
2.1. Implementing Data-Analytics Projects for Process importance in order to improve manufacturing processes [9].
Optimization Hence, it is essential to take this knowledge into account, when
applying data-analytics in manufacturing systems [18].
According to Russom [11] data-analytics is defined as Tools like Ishikawa analysis, FMEA or DoE are supportive
advanced analytic techniques that operate on a large database. and powerful techniques when expert knowledge and
Data-analytics is therefore generally the statistical and influences on the manufacturing process need to be recorded
mathematical analysis of data, also including data mining. Data systematically. In order to map processes and thus prepare and
Mining is used to extract knowledge out of data that is arrange existing process data, various classical approaches
generally applicable, not trivial, new, useful and exist. For example, Value-Stream Mapping (VSM) focuses on
understandable. For this knowledge retrieval process, machine the entire value chain from raw material to the customer and
learning algorithms are nowadays broadly and successfully captures the material and information flows [21].
used. [12, 13] For the application of data-analytics, different Enhancements of VSM include the following: mapping with
holistic end-to-end process models exist. The KDD-Model is IT-landscape [22], informational wastes regarding data
one process model for data mining [12]. It is an interactive and processing [23] and labeling of self-regulating processes [24].
iterative process following nine steps. In addition to the actual The details of the actual manufacturing processes, however,
data mining phase, the procedure also includes the steps of remain a black box. Marbán et al. [9] define a modelling
preparing the data and evaluating the results. The KDD model language to formalize manufacturing process chains as a pre-
forms the basis for quite a few more recent process models like phase of CRISP-DM, however existing control loops and the
for instance the KDD-Roadmap [14], which additionally information on sensors are not considered. Effect plans can be
determines resources required for the project or marketing used to represent control loops within and between the
specific developments [15, 16]. The well-established CRISP- individual machines and formalize knowledge regarding the
DM procedure model for data-analytics projects by Shearer [8] manufacturing processes [25].
was developed based on the KDD-Model [17]. It includes the All in all there is no holistic process representation of value
iterative steps business understanding, data understanding, data streams, control loops and sensors in manufacturing.
preparation, modeling, evaluation and deployment [8].
Nevertheless, neither KDD nor CRISP-DM are backed by well- 2.3. KPI Models for the Evaluation of Optimization Measures
defined methods and tools that facilitate a practical domain-
specific application in manufacturing. In particular, there is no KPIs are used to measure the performance of a
description of how the results of the data-analytics model can manufacturing system and to evaluate improvement measures
be used to derive and evaluate technical optimization measures. [26]. KPI systems were initially used just from a financial
The process model according to Marbán et al. [20] perspective (e.g. the well-known Du-Pont System). Later, KPIs
combines software engineering methods with CRISP-DM to were applied to include also quality aspects and other non-
extend it. In addition to the actual data-analytics process, financial systems [27]. Kang et al. [28] proposed a hierarchical
aspects of project management and organizational processes structure for KPIs and enables the identification of mutual
are taken into account (e.g. training of employees and relationships. For this purpose they separated basic and
maintenance of the data-analytics model). However, in many aggregated KPIs and analyze pairwise relationships among
cases in practice there is no perfect database available in the these KPIs. Today many other KPI systems can be found,
first place. So, Reinhart [18] motivates the necessity of an generally focusing on either a special area of manufacturing or
additional data acquisition step, but does not explain how it providing a general framework without domain-specific KPIs.
 Florian Ungermann et al. / Procedia CIRP 81 (2019) 369–374 371
Florian Ungermann et al. / Procedia CIRP 00 (2019) 000–000 3

In order to obtain a comprehensive understanding of a 3.1. Process Landscape and Target Values
discrete manufacturing system, a generic KPI model was
developed by Stricker et al. [29]. This KPI model consists of In the first phase Process Landscape and Target Values a
142 KPIs and is hierarchically structured. All KPIs are comprehensive current state analysis of the considered
interconnected by their mathematical interdependencies. Base manufacturing system is performed. The aim is to gain a
KPIs make up the bottom of the hierarchy (level 0). They are comprehensive understanding of the process. First, the system
fundamental KPIs that can directly be measured on the shop boundary needs to be defined. Based on an advanced VSM
floor (e.g. process times, number of defect parts). Aggregated technique, value stream, data landscape, process control loops
KPIs on higher levels are derived from these base KPIs [29]. and sensor technologies are mapped in an integrated process
Machine-specific process parameters (e.g. torque, speed) are map. This integrated process map consists of four linked layers,
not part of the network. which are arranged hierarchically.
An analytical extension of this KPI model by process The first and superordinate layer has the highest
parameters is complex and time-consuming, since all technical aggregation level. It is similar to the VSM and compromises
interdependencies must be explicitly, mathematically the manufacturing processes and the material flows. The single
formulated. Alternatively, machine learning methods can be process steps and machines are explicitly stated.
used to “learn” the connections between process parameters In the second layer the process steps and machines with
and KPIs based on historical data. Jinsong et al. [30] and Shin respect to the control loops are specified. All existing quality-
et al. [31] apply neural networks to predict the development of related control loops, including the associated machine
single KPIs based on process parameters. Ding et al. [32] use parameters, are mapped. The identification of the quality-
the partial least squares method in this context to identify critical control loops is carried out in a FMEA. The control
correlations between process parameters and key figures. loops can refer to just one single process step or across several
In summary, existing methods are often used to predict steps. Each control loop has defined incoming data streams,
individual KPIs, e.g. in quality control. However, they are not which originate from sensors.
used in a holistic network of KPIs, which helps to identify and The so-called landscape of sensors is described in the third
evaluate measures to improve processes. layer. This layer includes the location of sensors. For each
sensor characteristic attributes are recorded such as:
3. Methodology measurement value, frequency of measurement and output data
format. A sensor does not necessarily have to be a technical
By combining process parameters with a KPI model for the unit, but can also be represented by a person that manually
evaluation of optimization measures, a specific process model records and transmits measured values (e.g. control panel).
for data-analytics projects in manufacturing is introduced in The fourth and last layer compromises the data landscape.
this section. It addresses the shortcomings of existing All databases in which the sensor data is stored and processed
approaches outlined in Section 2. The method is intended for are listed in this layer. Furthermore, the data connections are
the application in existing manufacturing systems. It consists illustrated in order to be able to identify system breaks.
of five iterative steps, wherein the following key questions are Based on the integrated process map, critical process steps
addressed: Which overall KPIs, from the manufacturing system can easily be identified (e.g. high reject rates or long processing
point of view, are the target values of the data analytics project? times). These are in the focus of the subsequent optimization
How to evaluate and if needed extend an existing database? phase. In addition, the target KPI (e.g. OEE of the entire
What additional data requirements can be met and manufacturing system) is defined based on the KPI model (see
implemented when involving expert knowledge? Which Section 2.3).
process improvement measures can be derived and evaluated In this first phase of the approach it is crucial that all
participants, e.g. manufacturing process experts, production
planner and data-analytics experts, are involved. In this way a
comprehensive understanding of the considered manufacturing
process is likely to be achieved and the objectives are selected.

3.2. Data Requirements and Collection

The aim of the second phase Data Requirements and

Collection is the examination and selective extension of the
existing database. First, the Ishikawa method is used to
determine the most relevant influencing factors on the
manufacturing process. Based on the Ishikawa analysis, the
integrated process map is used to determine which influencing
factors are already captured by sensors and which are also
Fig. 1. Approach for data-based process optimization. included in existing control loops. For this purpose, an
evaluation of the proportion of recorded influencing factors for
based on the data analytics outcome? The approach is shown in the respective classes of the Ishikawa method is performed.
Fig. 1 and explained in the following subsections.
372 Florian Ungermann et al. / Procedia CIRP 81 (2019) 369–374
4 Florian Ungermann et al. / Procedia CIRP 00 (2019) 000–000

In a next step the quality and information content of the 3.4. Process-Specific Extension of the KPI Model
already collected data is examined. In addition to an explorative
analysis and visualization of the data, the variance in the data In the fourth phase of the approach, the KPI model (see
is investigated. If the variance is low with regard to the Section 2.3) is extended by the set of process parameters
optimization target defined in the first phase, which means the defined before. As shown in Fig. 2, the process parameters
parameter do not vary or there is no statistical relationship, an form an additional hierarchy level (level -1) below the base
increase of the variance can be achieved by a Design of KPIs (level 0). While the KPIs of the KPI model are all
Experiments (DoE). The result of a DoE is a well-defined analytically linked with their mathematical relations, the
dataset with high variance and eventually also additional connections between process parameters and the base KPIs are
process knowledge confirming the experiences of the experts. computed by a learning method (e.g. neural network). For this
If this is not possible, additional data can alternatively be purpose, the process parameters data is linked to the
generated from existing process knowledge or with the help of characteristics of the base KPIs via their time stamp over a
a simulation model. period of time. The connections are first learned based on a
The identified influencing factors, which are not recorded, training data set and then validated on a test data set. This
serve as a starting point for the integration of additional extended KPI model forms the basis for the selection of
sensors. The relevance of the influencing factors is evaluated measures in the next phase.
by a utility analysis and the feasibility of the recording is
evaluated based on a feasibility analysis. A distinction is made 3.5. Selection and Evaluation of Improvement Measures
between target and ideal data requirements. Target data
requirements are those that must be met. Ideal data For the selection of improvement measures, a distinction is
requirements represent an option for the future. made between process-specific measures and measures that are
In the last step, suitable sensor systems are selected and generally valid for interlinked manufacturing lines. The latter
implemented. The main questions in this phase are: Which are listed in a catalogue which is derived from general
measurands have to be recorded? Which installation space and knowledge on manufacturing performance improvement
interfaces are available? Which specifications should the measures such as integration of a buffer or employee training.
measurement signal with respect to the data interpretation? The The catalogue is categorized in four different categories:
sensor capability is investigated by a prototypical personal-, organizational-, material- and equipment-related
implementation. Sensor capability means that the sensor measures. Each measure is described and the prerequisites as
directly measures the addressed measurand, i.e. the influencing well as assumptions for implementation are recorded. While
factor, and that the measurement uncertainty is sufficient the generic catalogue can be used for all interlinked
according to the requirements. If so, the sensors are manufacturing systems, it is necessary to derive process-
implemented and the process map is adapted accordingly. specific measures with a direct effect on the identified process
parameters in expert workshops. Process-specific measures
3.3. Data Analysis are, for example, the design or adaptation of control loops or a
new set of fixed manufacturing parameters. The selection is
In the third phase Data Analysis the resulting database is based on the integrated process map (see Section 3.1). Weak
transformed, cleaned and a plausibility check of the measured points are systematically analyzed in the map, by a similar
values is carried out, e.g. physically impossible values or approach as the seven types of waste in lean management, and
incorrect data formats are identified. Identified outliers are suitable optimization measures are derived. Useful generic
replaced and missing data points are integrated via estimation measures are also preselected in this way. The resulting list of
procedures. The available series of process data is then linked measures is supplemented by a cost estimation and forms the
to the product-related quality data and the targeted KPIs. In basis for the following evaluation step.
doing so, a time series of process data values is assigned to each For each measure, the effect on the target KPIs is examined
workpiece. For further investigations, a suitable aggregation, using the expanded KPI model (see Fig. 2). This is done
e.g. standard deviation, median, mean etc., of these time series qualitatively for the generic measures. This means that a
is identified, which represents the influence on the quality data statement is made if the measure contributes to improving the
and targeted KPIs. A correlation matrix is used to identify those target value or not. For example, an additional training of
parameters that have the greatest influence on the KPIs. In this employees (M3) leads to an improvement of the OEE. The
way, the characteristic features of the measured process effect of a process-specific measure (M4) can be quantitatively
parameters are identified. The influencing factors determined described by changing the process parameters value. Using the
in the second phase can thus be confirmed in their importance learning method and the mathematical connections in the KPI
and new correlations can be discovered. It is also possible that model, the effect on the target KPIs can be calculated (see Fig.
no significant correlation can be found. Then it is necessary to 2). As a result, the measures are ranked according to their cost-
jump back to the phase Data Requirements and Collection and
adapt according to additional influencing factors of the
Ishikawa analysis.
As a result, a set of process parameters is identified that
provides a high potential for deriving improvement measures.
 Florian Ungermann et al. / Procedia CIRP 81 (2019) 369–374 373
Florian Ungermann et al. / Procedia CIRP 00 (2019) 000–000 5

benefit ratio. The promising measures are recommended for 4.3. Data Analysis
implementation. The new data points serve as further train data
for the learning method. The ongoing operation of the For analysis, an example data record is generated from the
manufacturing system continuously adjusts the KPI model. new and existing sensors. Since the new sensors are not yet
integrated into the manufacturing system, two data sets are
4. Application to Industrial Use Case created over a period of four weeks. The time stamp is used to
merge both data sets. To identify a suitable representation of
the time series, a correlation analysis is performed. The
correlation matrix for the liquid pressure is shown in Fig. 3. Q1
represents the product quality data. The mean value has the
highest correlation with the quality date Q1 (-0.35). The mean
value is therefore a suitable indicator to evaluate the level of
the pressure and thus forms the additional input variable for the
KPI model. The correlation analysis furthermore confirms the
relevance of all additional sensors, which were identified based
on expert knowledge.

4.4. Extension of the KPI Model and Derivation of

Improvement Measures
Fig. 2. Extended KPI model for selection of optimization measures.
The extension of the KPI model and derivation of
The introduced approach is applied exemplarily to a improvement measures results first in the extension of the KPI
manufacturing line. On this line, a thermal, ablative process is model. The additional process parameter liquid pressure is
applied to metallic parts. The line consists of five stations, is determined analytically as a learning method is not yet required
highly automated and rigidly interlinked. for just one single parameter.
The selection of optimization measures leads to the
4.1. Process Landscape and Target Values outcome that an additional control loop, which keeps the

The primary goal of the process improvement is the OEE

improvement by reducing the number of rejects. The integrated
process map includes all five stations. At the first station the
key product features are defined. At the second and third station
fine machining is performed and at the fourth station, multiple
product quality tests are carried out. At the fifth and last station,
a data matrix code is applied. A total of five quality-critical
control loops can be identified, including the sensors from the
second layer of the integrated process map. Most of the sensor
data is stored in the internal machine control system. Access is
possible via a process data acquisition system. Since the reject
and rework rate at the first station is the highest, the focus of
further analysis is on this process step. The high reject rate
implies a large potential for improvement by optimizing the
process parameters. Fig. 3. Correlation matrix of liquid pressure and quality data.

4.2. Data Requirements and Collection pressure of the liquid constant at a high level, should be
installed. In addition to the already integrated sensors, the
A total number of 42 influencing factors can be identified required control logic is built up with a separate
within the Ishikawa analysis on the first station based on expert microcontroller. Targeted employee training is another selected
knowledge. However, just 13 of these influencing factors are measure to improve the OEE, an influencing factor related to a
currently recorded and only one quality critical control loop human operation. A positive effect of these measures on the
making use of them is implemented. The utility and feasibility OEE can be shown with the help of the KPI model and has been
analysis result in a desirable enhancement of the database by confirmed in first prototypical implementations. In order to
eight parameters (+19%). The ideal database represents 19 derive further measures, a neural network will be set up in a
additional parameters. It is recommended to record these next step to identify complex and non-linear dependencies.
influencing factors by means of additional sensors. In the
following, the data requirement liquid pressure is focused as an 5. Summary
example. The aim is to achieve a constant pressure within the
process to ensure a higher robustness of the considered thermal, The objective of this paper is to present a methodology for
ablative manufacturing process. The liquid pressure can be data-driven process optimization in manufacturing. In particu-
recorded with a pressure sensor, which can easily be integrated
lar, the question how to use and extend an existing database.
in the manufacturing line.
374 Florian Ungermann et al. / Procedia CIRP 81 (2019) 369–374
6 Florian Ungermann et al. / Procedia CIRP 00 (2019) 000–000

The proposed methodology includes a five step procedure. First [13]Runkler, T. A., 2015. Data Mining. Modelle und Algorithmen intelligenter
a comprehensive current state analysis is performed. Based on Datenanalyse, Springer Vieweg, Wiesbaden.
[14] Debuse, J. C. W., La Iglesia, B. de, Howard, C. M., Rayward-Smith, V. J.
this analysis further data requirements are derived and suitable , 2001. Building the KDD Roadmap. A Methodology for Knowledge
sensors are implemented. These additional data are used to Discovery, in: Industrial Knowledge Management. A Micro-level
identify improvement measures, which are evaluated by a Approach, ed. R. Rajkumar, Springer-Verlag, London, p. 179–196.
process-specific KPI model. The approach is validated using [15] Pan, D., 2010. Data Mining Process Model for Marketing and CRM, in:
the example of a manufacturing line. Measures to improve International Conference on Machine Learning and Cybernetics (ICMLC
2010), p. 433–438.
OEE were derived for a thermal, ablative process step. [16] Büchner, A. G., Mulvenna, M. D., Anand, S. S., Hughes, J. G., 2000. An
Internet-Enabled Knowledge Discovery Process.
Acknowledgements [17] Mariscal, G., Marbán, Ó., Fernández, C., 2010. A survey of data mining
and knowledge discovery process models and methodologies, in: The
This research is funded by the German Federal Ministry of Knowledge Engineering Review 25(02), p. 137–166.
[18] Reinhart, F., 2016. Industrial Data Science. Data Science in der
Education and Research (BMBF): "Produktionsbezogene industriellen Anwendung, in: Industrie 4.0 Management 32, p. 27–30.
Dienstleistungssysteme auf Basis von Big-Data-Analysen [19] Bahrepour, M., 2018. The Forgotten Step in CRISP-DM and ASUM-DM
(ProData)" (funding code 02K16C08). Furthermore, the Methodologies, Luminis, Amsterdam.
authors are indebted to the Baden-Württemberg Stiftung for the [20] Marbán, Ó., Mariscal, G., Menasalvas, E., Segovia, J., 2007. An
financial support of this research project by the “Elitepro- Engineering Approach to Data Mining Projects, Intelligent Data
Engineering and Automated Learning, ed. H. Yin, P. Tino, E. Corchado,
gramme für Postdocs”. W. Byrne & X. Yao, Springer-Verlag, Berlin, p. 578–588.
[21] Rother, M., Shook, J., 1999. Learning To See - Value Stream Mapping to
References Create Value and Eliminate Muda, Brookline, Lean Enterprise Institute,
Massachusetts.
[1] Groggert, S., Wenking, M., Schmitt, R.H., Friedli, T, 2017. Status quo and [22] Roessler, M. P., Kleeberg, I., Kreder, M., Metternich, J., Schützer, K.,
Future Potential of Manufacturing Data Analytics – an Empirical Study, in: 2015. Enhanced Value Stream Mapping: Potentials and Feasibility of IT
IEEE International Conference on Industrial Engineering and Engineering Support through Manufacturing Execution Systems.
Management (IEEM), p. 779-783. [23] Meudt, T., Rößler, M. P., Böllhoff, J., Metternich, J., 2016.
[2] Wang, K., Tong, S., Eynard, B., Roucoules, L., Matta, N., 2007. Review on Wertstromanalyse 4.0, in: ZWF Zeitschrift für wirtschaftlichen
Application of Data Mining in Product Design and Manufacturing, in: Fabrikbetrieb 111(6), p. 319-323.
Fourth International Conference on Fuzzy Systems and Knowledge [24] Lass, S., Theuer, H., Hennig, G., Schumacher, J., 2013. Modellierung
Discovery (FSKD), p. 613-618. intelligenter Produktionssysteme, in: Productivity Management (1)2013.
[3] Uhlemann, T. H.-J., Lehmann, C., Steinhilper, R., 2017. The Digital Twin: [25] Böge, A., 2011, Handbuch Maschinenbau: Grundlagen und Anwendungen
Realizing the Cyber-Physical Production System for Industry 4.0, in: der Maschinenbau-Technik, Springer Vieweg, Wiesbaden.
Procedia CIRP 61, p. 335–340. [26] International Organization for Standardization, 2014. Automation systems
[4] Harvard Business Review, 2017. The Enterprise Lacks a Big Data Strategy and integration - KPIs for manufacturing operations management - Part 2:
for IoT Transformation, Harvard Business School Publishing. Definitions and descriptions, no. 22400-2.
[5] Joch, A., Moreno, K., 2017. Data & Advanced Analytics – High Stakes, [27] DGQ, 1990. Qualitätskennzahlen und Qualitätskennzahlen-Systeme, no.
high Rewards, Forbes Insights, Ernst&Young. 14-23, Beuth Verlag, Berlin.
[6] Groggert, S., Elser, H., Ngo, Q. H., Schmitt, R. H., 2018. Scenario-based [28] Kang, N., Zhao, C., Li, J., Horst, J. A., 2016. A Hierarchical structure of
Manufacturing Data Analytics with the Example of Order Tracing through key performance indicators for operation management and continuous
BLE-Beacons, in: Procedia Manufacturing 24, p. 243–249. improvement in production systems, in: International Journal of Production
[7] Kampker, A, Heimes, H., Bührer, U., Lienemann, C., Krotil, S., 2018. Research 54(21), p. 6333–6350.
Enabling Data Analytics in Large Scale Manufacturing, in: Procedia [29] Stricker, N., Echsler-Minguillon, F., Lanza, G., 2017. Selecting KPIs for
Manufacturing 24, p. 120–127. production with a linear programming approach, in: International Journal
[8] Shearer, C., 2000. The CRISP-DM Model: The New Blueprint for Data of Production Research 55(19), p. 5537-5549.
Mining, in: Journal of Data Warehousing 5(4), p.13-22. [30] Jinsong, B., Yuan, G., Xiaohu, Z., Jianguo, Z., Xia, J., 2017. A Data
[9] Marban, O., Mariscal, G., Segovia, J., 2009. A Data Mining & Knowledge Driven Model for Predicting Tool Health Condition in High Speed Milling
Discovery Process Model, in: Data Mining and Knowledge Discovery in of Titanium Plates Using Real-Time SCADA, in: Procedia CIRP 61, p.
Real Life Applications, p. 438 – 453. 317-322.
[10] Wieland, U., Fischer, M., 2013. Zur methodischen Vorbereitung von [31] Shin, S.-J., Woo, J., Rachuri, S., 2014. Predictive Analytics Model for
Data-Mining-Projekten unter Verwendung von CRISP-DM im Kontext Power Consumption in Manufacturing, in: Procedia CIRP 15, p. 153-158.
diskreter Produktionsprozesse, in: Tagungsband zum fünften Workshop [32] Ding, S. X., Yin, S., Peng, K., Hao, H., Shen, B., 2013. A Novel Scheme
der GI-Fachgruppe Business Intelligence, p. 1613–1673. for Key Performance Indicator Prediction and Diagnosis With Applicatio
[11] Russom, P., 2011. Big Data Analytics, in: TDWI Best Practices Report, to an Industrial Hot Strip Mill, in: IEEE Transactions on Industrial
TDWI Research. Informatics 9(4), p. 2239-2247.
[12] Fayyad, U., Piatetsky-Shapiro, G., Smyth, P., 1996. The KDD process for
extracting useful knowledge from volumes of data, in: Communications of
the ACM 39(11), p. 27–34.