A Hybrid Machine Learning Approach

for Predictive Maintenance in Smart Factories

of the Future

Sangje Cho1(&), Gökan May1, Ioannis Tourkogiorgis1,

Roberto Perez2, Oscar Lazaro3, Borja de la Maza4,
and Dimitris Kiritsis1
1
EPFL, ICT for Sustainable Manufacturing, EPFL SCI-STI-DK,
Station 9, 1015 Lausanne, Switzerland
[email protected]
2
GF Machining Solutions, Meyrin, Switzerland
3
Innovalia, Bilbao, Spain
4
TRIMEK, Zuia, Álava, Spain

Abstract. Advanced technologies based on Internet of Things (IOT) are

blazing a trail to effective and efficient management of an overall plant. In this
context, manufacturing companies require an innovative strategy to survive in a
competitive business environment, utilizing those technologies. Guided by these
requirements, the so-called predictive maintenance is of paramount importance
and offers a significant potential for innovation to overcome the limitations of
traditional maintenance policies. However, real shop-floors often have obstacles
in providing insights to facilitate the effective management of assets in smart
factories. Even if a significant amount of machine and process data is available,
one of the common problems of these data is the lack of annotations describing
the machine status or maintenance history. For this reason, companies have
limited options to analyse manufacturing data, despite the capability of
advanced machine learning techniques in supporting the identification of failure
symptoms in order to optimize scheduling of maintenance operations. More-
over, each machine generates highly heterogeneous data, making it difficult to
integrate all the information to provide data-driven decision support for pre-
dictive maintenance. Inspired by these challenges, this research provides a
hybrid machine learning approach combining unsupervised learning and semi-
supervised learning. The approach and result in this article are based on the
development and implementation in a large collaborative EU-funded H2020
research project entitled BOOST 4.0 i.e. Big Data Value Spaces for COmpeti-
tiveness of European COnnected Smart FacTories.

Keywords: Industry 4.0
Predictive maintenance
Machine learning

Big data
Asset management
Smart factories
Sustainable manufacturing

© IFIP International Federation for Information Processing 2018

Published by Springer Nature Switzerland AG 2018. All Rights Reserved
I. Moon et al. (Eds.): APMS 2018, IFIP AICT 536, pp. 311–317, 2018.
https://doi.org/10.1007/978-3-319-99707-0_39
312 S. Cho et al.

1 Introduction

As of today, modern industries require efficiency and convenience for management of

the entire Product Life Cycle(PLC), in order to overcome intensely competitive busi-
ness environment. Advances of IoT opens an efficient way to innovative predictive
maintenance strategies in smart manufacturing environments, and these advanced
technologies generate industrial big data. To exploit this big data, trend-oriented pre-
dictive maintenance tasks are carried out based on actual condition of machinery to
avoid occurrence of failures, since advanced enabling technologies and wireless
technologies open innovative capability to monitor very details of machines’ status and
behaviour. In that regard, even if predictive maintenance allows increasing business
values and smart machining services, big data from various technologies should be
effectively managed for interoperability via standards for merging and transformation
of data. According to Lee, a successful shift toward more intelligent machines can be
addressed considering five distinct issues as follows: Manager and Operator Interac-
tion, Machine Feet, Product and Process Quality, Big Data and Cloud, and Sensor and
Controller Network [1]. To address these issues, predictive maintenance pilots in the
EU-funded H2020 research project entitled BOOST 4.0 [2] provides product innova-
tion through a data-driven approach. These pilots are elaborated to integrate digital
platforms and industrial things so as to foster collaboration considering the features of
Industry 4.0 including (i) horizontal integration through value networks to facilitate
inter-corporation collaboration, (ii) vertical integration of hierarchical subsystems
inside a factory to create flexible and reconfigurable manufacturing system, and
(iii) end-to-end engineering integration across the entire value chain to support product
customization [3].
Meanwhile, data analytics through advanced machine learning techniques has been
improved with the development of strong hardware and useful algorithms, supporting
engineers to find trends and symptoms of failures in order to carry out maintenance
tasks optimally. However, types and formats of data vary significantly depending on
data sources. Most companies do not have the competence for management of such big
data, and often record data without tags describing machine status and/or maintenance
history. These constraints limit the application of machine learning algorithms and thus
supervised learning and semi-supervised learning cannot be performed for data ana-
lytics. For this reason, this research provides a hybrid machine learning approach
combining unsupervised learning and semi-supervised learning based on the devel-
opment and implementation in the BOOST 4.0 project.

2 A Main Architecture for the Predictive Maintenance Pilot

Thanks to recent scientific and technological developments, most industrial practices

try to employ a predictive maintenance policy instead of conventional Maintenance
(i.e. corrective and/or preventive maintenance). According to Sullivan et al., inde-
pendent surveys indicate that this predictive maintenance policy can lead to high return
on investment, reduction in maintenance cost, elimination of breakdown, reduction in
downtime and increase in production [4] since conventional maintenance policies incur
 A Hybrid Machine Learning Approach for Predictive Maintenance 313

low reliability of machines or needless maintenance tasks. However, the main con-
straints for application of predictive maintenance in BOOST 4.0 can be summarized as
no maintenance history and heterogeneous data. Inspired by these constraints, the
predictive maintenance pilot in BOOST 4.0 deals with a business case as follows:
• The target product is a milling machine in the shop floor
• Products produced by a milling machine can be measured by a Coordinate Mea-
suring Machine (CMM) in the same shop floor
• Providers of milling machines and CMM machines are different, and machine data
is collected by each machine provider
• The results of measurement are useful for the milling machine, but currently there is
no intersection between two kinds of data
• The Data formats of the milling machine and the CMM machine are heterogeneous
• Milling machine data does not have maintenance indicator.
This pilot study consists of the following steps: (i) the acquisition and storage of
data, (ii) data analytics on operating data and monitoring, (iii) continuous evaluation
and prediction of the health status of the equipment as cyber-physical system (i.e.
transform descriptors extracted previously in relevant behaviour models, allowing to
represent the ways of functioning of the machine and the evolution of the equipment
condition over time for detection and prognosis of failures), (iv) decision-making
support by considering the context of use of the equipment.

Fig. 1. An architecture for the predictive maintenance pilot

To address the concern of the acquisition and storage of data, application of the
open source platform for the smart digital future considers data management (See
Fig. 1). This open platform has the capacity of context management to merge
heterogeneous data from milling machines and CMM machines. A system adapter in
the open source platform will read all the data from milling and CMM machines.
314 S. Cho et al.

Tracking ID of a product produced by milling machine, it will recognize relevant

measurement data. Afterwards, it will update attribute values of defined entities which
have the context of the shop floor. Context data will be managed by Context Blocker
within IDS connectors. Context data from the shop floor will be accessible through IDS
connect which is a back-end processor communicating with the milling machine pre-
dictive maintenance system. Depending on data security and requirements, IDS con-
nector will send relevant data to the predictive maintenance system. The Predictive
maintenance system will exploit distributed data for maintenance planning &
scheduling. This system is in charge of data analytics on operating data and monitoring,
evaluation and prediction of the health status, and decision-making support. On the
other hand, the open source platform allows exploitation of a specific part of milling
machine data for the CMM precision management system which is a part of an
operation management pilot. In this context, this study focuses on the predictive
maintenance application.

3 A Hybrid Approach of the Predictive Maintenance Pilot

This chapter describes the details of the predictive maintenance system for the pilot. As
mentioned above, milling machine data have no maintenance indicator of any events.
Therefore, available data limits application of machine learning algorithms and thus
supervised learning and semi-supervised learning are not available for data analytics.
For this reason, this research provides a hybrid machine learning approach combining
unsupervised learning and semi-supervised learning (Fig. 2).

Fig. 2. A hybrid approach for the predictive maintenance pilot

The hybrid approach is comprised of two parts; a training part and a running
part. The main role of the training part is to provide a probabilistic model to identify the
 A Hybrid Machine Learning Approach for Predictive Maintenance 315

status of machines. As a hybrid approach, this part has unsupervised learning module
linked to soft-class assignment and semi-supervised learning module linked to hard-
class assignment. Soft-class assignment initiates all the classes as inputs of an unsu-
pervised learning module to overcome no maintenance indicator whereas hard-class
assignment provides consistency of classes.
Based on the assignment modules, new data from milling machines could be used
to identify if the machine status is normal or not. To provide a detailed explanation of
procedures of data analytics, roles of each module are described below.
Receiving unlabelled training data sets (D1u) which means they have no mainte-
nance indicators from the open source platform, the training part will classify all the
data set through an unsupervised learning module. Where the class indexes are k and
pk ¼ NP k =N (Nk: a number Dataset P of class k, N is a total number of datasets),
pk ; lk ; k for each class k and lD1u ; D1u are estimated. And then, where xn 2 D1U ,
P
each of pk ; lk ; k will be updated through the Expectation-Maximization (EM) [5]
algorithm as follows:
P
pk Nðxn jlk ; k Þ
E step : cnk ¼ K ð1Þ
P P
pj Nðxn jlj ; j Þ
j¼1

1 X N
M step : lnew ¼ c xn ð2Þ
k
Nk n¼1 nk

Xnew 1 X N
new T
¼ c ðxn  lnew
k Þðxn  lk Þ ð3Þ
k Nk n¼1 nk

Nk
pk ¼ ð4Þ
N

The log likelihood can be estimated as follows:

( )
X X
N X
K X
ln pðXjl; ; pÞ ¼ ln pk Nðxn jlk ; Þ ð5Þ
n¼1 k¼1 k

In the case of P½lk 2 D1u   w (where, w is a predefined acceptance parameter), the

k class will be soft-assigned to a normal class because normal data set is overwhelming
in all the data sets. Otherwise, it will be soft-assigned to an abnormal class. When new
data sets (D2) from active milling machines are delivered to the running part, the
running part will estimate arg max P½xt 2 Ck  ðxt 2 D2 Þ. If max P½xt 2 Ck   w, d2 will
k P
be assigned to Class k, and pk ; lk ; k will be updated. Otherwise, d2 will be reported
as unidentified data set. The visualization module will display all the graphs with
significant values to deliver results of analytics (See Fig. 3). Maintenance engineers
will validate these results through dashboard and will give maintenance annotations
316 S. Cho et al.

(D1L) to abnormal/unidentified data sets. These maintenance annotations will go to the

semi-supervised Plearning module this module will initiate classes based on D1L. To
update pk ; lk ; k of each class k, where xn 2 D1U ðD1 ¼ D1U [ D1L Þ, E step of EM
algorithm is as follows:
8 P
> pk Nðxn jlk ; Þ
<PK
k
P if dn 2 D2U
E step : cnk ¼ pj Nðxn jlj ; Þ ð6Þ
>
: j¼1
j

1 ðk ¼ yðnÞÞ if dn 2 D2L

M step and estimation of the log likelihood can be estimated following Eqs. (2),
(3), (4) and (5).

Fig. 3. A dashboard of predictive maintenance system

Return values from the hybrid approach are presented in a dashboard (See Fig. 3).
Considering how to show results of the proposed approach, the dashboard is comprised
of three kinds of aspects, i.e. a summary of the machines, machines` KPI, and machine
details. A summary of the machines shows a graph representing a rate of available
machines, and abnormal machines with probability of failures. Moreover, machines`
KPI indicates meaningful information such as failure probability, failure events, aging,
 A Hybrid Machine Learning Approach for Predictive Maintenance 317

warning events, and so on. Machine details display status of each machine. The hybrid
approach allows extension of visualisation depending on validation process.
The resulting predictive maintenance system will be demonstrated in the milling
machine scenario, and then, its scope will be extended with illustration of all the
predictive maintenance pilots in BOOST 4.0 for a wide range of application.

4 Discussion and Concluding Remarks

The main purpose of this research was to provide a novel predictive maintenance
approach for the predictive maintenance pilot of BOOST 4.0. This study addressed the
problems caused by no maintenance annotations and heterogeneous data sources. To
resolve the issue in an efficient way, this paper included an architecture exploiting the
open source platform for the smart digital future for merging heterogeneous data from
milling machines and CMM machines through context management. In addition, the
hybrid predictive maintenance approach was proposed to overcome constraints of no
maintenance annotations.
Accordingly, the implications for knowledge and practice could be summarized as
follows: (i) the shop floor enables a high-value service for users of the equipment by
avoiding downtimes through predictive knowledge. This service brings results of
minimizing the total cost and offers optimization of material usage, and (ii) through
increased equipment availability, as well as the manufacturing technology as a whole,
will help customers to gain a competitive advantage where much unforeseen downtime
reduces the profitable production time.
As for future work, the proposed approach will be implemented and validated on
not only the milling machine case but also other predictive maintenance pilots within
tasks of BOOST 4.0, demonstrating its capacity and potential to support maintenance
engineers and machine operators.

Acknowledgements. This work has been carried out in the framework of BOOST 4.0 Project,
which has received funding from the European Union’s Horizon 2020 research and innovation
programme under grant agreement Nº 780732.

References
1. Lee, J., Kao, H.A., Yang, S.: Service innovation and smart analytics for industry 4.0 and big
data environment. Procedia Cirp 16, 3–8 (2014)
2. BOOST4.0 Project. http://boost40.eu/. Accessed 06 Apr 2018
3. Wang, S., Wan, J., Li, D., Zhang, C.: Implementing smart factory of industrie 4.0: an outlook.
Int. J. Distrib. Sens. Netw. 12(1), 3159805 (2016)
4. Sullivan, G., Pugh, R., Melendez, A.P., Hunt, W.D.: Operations & Maintenance Best
Practices-A Guide to Achieving Operational Efficiency (Release 3) (No. PNNL-19634).
Pacific Northwest National Laboratory (PNNL), Richland (2010)
5. Bishop, C.M.: Pattern Recognition and Machine Learning. Springer, New York (2006)