Proceedings of APSIPA Annual Summit and Conference 2019 18-21 November 2019, Lanzhou, China

IoT-based Predictive Maintenance for Smart

Manufacturing Systems
Leong Chee Him∗ , Yu Yong Poh† and Lee Wah Pheng‡
∗ Tunku Abdul Rahman University College, Kuala Lumpur, Malaysia
E-mail: [email protected]
† Tunku Abdul Rahman University College, Kuala Lumpur, Malaysia

E-mail: [email protected]
‡ Tunku Abdul Rahman University College, Kuala Lumpur, Malaysia

E-mail: [email protected]

Abstract—Manufacturers have been practicing traditional pre- explained in Section III, the results obtained from the method
ventive maintenance for many years. However, it is not cost- are shown and discussed in Section IV. Section V concludes
effective. To avoid ineffective maintenance routine and costs, the paper.
manufacturers can leverage Industrial IoT and data science. This
paper presents a method to optimize the manufacturing processes II. P RELIMINARY S TUDIES
by using IoT-based predictive maintenance. It illustrates how an
IIoT solution can be used to predict a manufacturing defect. In this section, preliminary studies are presented. Brief
The data is collected from multiple smart sensors stored on explanations of Industry 4.0, preventive maintenance and
this welding machine. It is monitored using Statistical process predictive maintenance are included in this section.
control methods. Machine learning algorithms are applied to
reveal hidden correlations in the data sets and detect abnormal A. Industry 4.0
data patterns. The recognized data patterns are then reflected in
predictive models, classification approaches are used to identify As a part of a project carried out by the high tech strategy
the type of manufacturing processes, namely normal and welding of the German government, Industry 4.0 is the term that they
problem. The variables that contribute the most to the failure use to promote the computerization in Industry. Industry 4.0
are identified.
is treated as the next phase of digitization of Industry. There
I. I NTRODUCTION are four pillars in Industry 4.0, they are the advancing in
Manufacturers have been practicing traditional preventive data accessing, computing power, and connectivity between
maintenance for many years. Traditional preventive main- devices; the increasing capabilities of analytic and business-
tenance has two methods: condition-based and age-based. intelligence; Advanced human-machine interaction with Touch
Maintenance tasks or replacement will occur if the measure User Interface (TUI), Virtual Reality (VR) and Augmented
of degradation of the machine monitored is found exceeding Reality (AR) systems; and the improvement in establishing
preset threshold [1] or the operation cycle or age of the connection between digital world physical world, such as
machine is longer than preset period [2]. Both methods are advanced robotics and 3-D printing [4], [5].Industry 4.0 is
not cost-effective; Spare parts need to be always available and not only include industry but an overall transformation where
breakdown may occur before preset maintenance period. As machines will be redefined in the way of they communicate
Industry 4.0 was introduced by German government in 2011, with each other and carry out individual process by using
manufacturers have moved fast to improve their process to intelligent engineering and digitization [6].
meet up with the market expectation [3]. One of the pillars
B. Preventive Maintenance
of Industry 4.0 is extension of uses of Internet to connect
industrial assets, such as machines, devices, sensors, and Preventative maintenance is one of the most adopted main-
people, or known as Internet of Things (IoT). By leveraging tenance strategy nowadays. Condition-based and age-based
Industrial IoT and data science, predictive maintenance is are the two major methods in preventive maintenance. For
enabled to manufacturers, which can be practiced to extend the condition-based case, a machine will be monitored on its
operation cycle of machines till the end of its useful lifespan. degradation measure in a constant time interval. The machine
In this paper, we purpose a method to optimize the man- will be replaced if the measure is found exceeding preset
ufacturing process with IoT-based predictive maintenance. It threshold [1]. For age-based case, a machine will be replaced if
illustrates how an IIoT solution can be used to predict a the operation cycle or age of the machine is longer than preset
manufacturing defect . period [2]. Preventive maintenance has significant downsides,
This paper is organized as follows: Relevant preliminary one of them is the maintenance or replace of the machine
studies are presented in Section II. While proposed method is will make the machine unavailable, which will affect the
production schedule. Another downside is the spare parts need
This work is supported in part by Kian Joo Can Factory Berhad. to be always available for maintenance, which increases the

978-1-7281-3248-8©2019 APSIPA 1942 APSIPA ASC 2019

Proceedings of APSIPA Annual Summit and Conference 2019 18-21 November 2019, Lanzhou, China

inventory costs [7], [8]. Besides, maintenance tasks are carried • Flow rate of cooling water in section B241 of the machine
out based on the time of operation of the machine, breakdown Every 500 milliseconds, a programmable logic controller
may happen before the maintenance tasks are carried out (PLC) will send the sensor data to a supervisory control
or parts are replaced before end of their lifetime, which is and data acquisition (SCADA) system, SIMATIC WinCC, via
inefficient in term of costs [8]. Ethernet to monitor and record in a local database. The data
C. Predictive Maintenance sets used in this experiment are the sensor data collected from
April to May 2019. The failures of the welding machine during
Predictive maintenance is one of the maintenance methods, the period were recorded in manual (paper) format by the
which fully utilizes the parts to the limit of their service life by operators of the machine. There were 18 cases of failures,
data analysing and diagnosis on the parts and predicting their namely ’welding problem’, occurred during the period.
remaining service life [9]. Predictive maintenance is able to
ensure the maximum time interval between maintenance tasks B. Goals and Performance Metrics
and minimize the occurrence and cost of unscheduled down-
In this paper, we proposed to use one of the machine
time by machine-train failures [8]. In predictive maintenance,
learning algorithms in Scikit-learn library, decision tree, to
by measuring and analyzing deterioration data from sensors
identify the hidden correlations in the data sets and detect
or surveillance system, trend value can be managed, decision
abnormal data patterns. Scikit-learn uses an optimised version
making can be done based on the trend value for performing
of the CART (Classification and Regression Trees) algorithm,
maintenance tasks, and condition can be monitored through
which is very similar to C4.5, but supports numerical target
local or online system in real time [9]. In industry 4.0, predic-
variables (regression) and does not compute rule sets [12].
tive maintenance plays an important role to ensure and increase
This experiment evaluate how an IIoT solution can be used
effectiveness of maintenance tasks [10]. While advanced data
to predict a manufacturing defect, by addressing the features
accessing, widen connectivity and advanced analytical tools
that contribute the most to the failure. Precision, Recall and
are promoted by Industry 4.0, predictive maintenance is en-
F-measure (f1-score) of failure class are used to define the
abled as large amount of data can be monitored, pushed into
prediction performance where we used to identify the features
analytical tools to predict machine failure, downtime, overload
that contributed the most to the failure. While precision is
or any other problems [11].
the ratio of true negative to all predicted negative, recall is the
III. E XPERIMENTAL S TUDY ratio of true predicted failures to all true failures in the test set
In this section, we explain how the data is collected and the and F-measure is the harmonic mean of precision and recall.
details of the proposed method. While different importance can be assumed by precision or
recall depending on context of application, F-measure allows
A. Data Collection to take both in consideration with a single metrics [13].
The machine used in this experiment is a can welding
C. Experimental Workflow
machine, sensors were installed in different section of the
machine. The variables are described below: All the features The sensor data of the welding machine operated for
in the data set as described below: hours without any failure occur was extracted and labeled
• Current passing the welder as ’normal’ data set. For each failure, the sensor data of 20
• Air pressure of air circulation in the machine motor minutes before the failure occur was extracted and labeled as
• Vibration of the machine motor ’welding problem’ data set. All the ’welding problem’ data
• Pressure of nitrogen tank sets for the preset time interval were combined as a single
• Pressure of incoming cooling water ’welding problem’ data set. Then the ’welding problem’ data
• Pressure of incoming chilling water set combined with the ’normal’ data set as a single data set
• Temperature of chilling water in section B80 of the to used in this experiment. The data set was split into train
machine and test set with presets of features that may contribute to
• Temperature of chilling water in section B81 of the the failure. Data set of features was normalized if the features
machine had different units and scales. Decision tree model was built,
• Temperature of chilling water in section B82 of the trained with train set and tested the prediction with test set.
machine The presets of features used in this experiment are described
• Temperature of cooling water in section B240 of the below:
machine • All features (AF)
• Temperature of cooling water in section B241 of the • Temperature of all sections in the machine (TA)
machine • Flow rate of all sections in the machine (FRA)
• Flow rate of chilling water in section B80 of the machine • Air pressure of air circulation and vibration of the ma-
• Flow rate of chilling water in section B81 of the machine chine motor (AP & Vib)
• Flow rate of chilling water in section B82 of the machine • Flow rate of different pair of sections in the machine
• Flow rate of cooling water in section B240 of the machine • Temperature of different pair of sections in the machine

1943
Proceedings of APSIPA Annual Summit and Conference 2019 18-21 November 2019, Lanzhou, China

Precision, recall and F-measure for failure class of each model V. C ONCLUSION
were recorded and explained in Section IV. In this paper, we proposed the method of using decision
IV. E XPERIMENTAL R ESULTS tree algorithm to identify the feature that contributes the most
to failure and predict the failure. The proposed method gave
In this section, the results obtained from the proposed
acceptable result of 0.86, 0.94 and 0.90 for precision, recall
method for each preset of features are shown.
and F-measure respectively, and temperature is identified as
The values of precision, recall and F-measure of failure
the variable that contributes the most to failure while section
class for all features, temperature features, flow rate features,
B80 and B81 are the sections that may be the cause to failure.
and air pressure and vibration features, are shown in Table I.
The results shows the decision tree is able to predict failure R EFERENCES
with high value of precision and recall. Even flow rate has [1] W. Wang, “A model to determine the optimal critical level and the mon-
the highest value of recall but the value of precision and F- itoring intervals in condition-based maintenance,” International Journal
measure are the lowest among the feature sets. From the result, of Production Research - INT J PROD RES, vol. 38, pp. 1425–1436, 04
2000.
temperature seems to be the feature that contributes the most to [2] G. J. Glasser, “The age replacement problem,” Technometrics, vol. 9,
failure with the highest value of precision and F-measure, and pp. 83–91, 02 1967.
second highest value of recall. The values of precision, recall [3] L. Bassi, “Industry 4.0: Hope, hype or revolution?” in 2017 IEEE 3rd
International Forum on Research and Technologies for Society and
and F-measure of failure class for flow rate and temperature Industry (RTSI), Sep. 2017, pp. 1–6.
of different pair of sections in machine are shown in Table [4] J. Lee, E. Lapira, B. Bagheri, and H. an Kao, “Recent advances and
II and Table III. The result shows the temperature of section trends in predictive manufacturing systems in big data environment,”
Manufacturing Letters, vol. 1, pp. 38–41, 2013.
B80 and B81 has the highest value of F-measure. [5] T. K. Sung, “Industry 4.0: A Korea perspective,” Technological Fore-
casting and Social Change, vol. 132, pp. 40–45, 2018.
TABLE I [6] A. A. Pranab K.Muhuria, Amit K.Shuklaa, “Industry 4.0: A bibliometric
R ESULT OF FAILURE CLASS PREDICTION FOR analysis and detailed overview,” Engineering Applications of Artificial
DIFFERENT PRESETS OF FEATURES Intelligence, vol. 78, pp. 218–235, 2019.
Features Precision Recall F-measure [7] W. B. Wim J.C.Verhagen Lennaert, “Predictive maintenance for aircraft
AF 0.90 0.95 0.93 components using proportional hazard models,” Journal of Industrial
TA 0.86 0.94 0.90 Information Integration, vol. 12, pp. 23–30, 12 2018.
FRA 0.54 0.98 0.69 [8] R. K. Mobley, An Introduction to Predictive Maintenance, 2nd ed.
AP & Vib 0.76 0.85 0.80 Butterworth-Heinemann, 2002.
[9] J. Y. M. Saeid Mokhatab, William A. Poe, Handbook of Natural Gas
Transmission and Processing: Principles and Practices, 3rd ed. Gulf
Professional Publishing, 2018.
TABLE II [10] M. Bengtssonab and G. Lundstrmb, “On the importance of combining
R ESULT OF FAILURE CLASS PREDICTION FOR FLOW
the new with the old One important prerequisite for maintenance in
RATE OF DIFFERENT PAIR OF SECTIONS IN MACHINE
Industry 4.0,” Procedia Manufacturing, vol. 25, pp. 118–125, 2018.
[11] N. F. A. Alejandro Germn Frank, Lucas Santos Dalenogare, “Industry
Features Precision Recall F-measure 4.0 technologies: Implementation patterns in manufacturing companies,”
B80 & B81 0.78 0.16 0.27 International Journal of Production Economics, vol. 210, pp. 15–26, 4
B80 & B82 0.54 0.97 0.69 2019.
B80 & B240 0.52 0.96 0.68 [12] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
B80 & B241 0.52 0.96 0.68 O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vander-
B81 & B82 0.54 0.98 0.69 plas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duch-
B81 & B240 0.78 0.16 0.26 esnay, “Scikit-learn: Machine learning in Python,” Journal of Machine
B81 & B241 0.53 0.95 0.68 Learning Research, vol. 12, pp. 2825–2830, 2011.
B82 & B240 0.54 0.97 0.69 [13] D. Apiletti, C. Barberis, T. Cerquitelli, A. Macii, E. Macii,
B82 & B241 0.54 0.98 0.69 M. Poncino, and F. Ventura, “istep, an integrated self-tuning en-
B240 & B241 0.68 0.19 0.29 gine for predictive maintenance in industry 4.0,” in 2018 IEEE Intl
Conf on Parallel Distributed Processing with Applications, Ubiqui-
tous Computing Communications, Big Data Cloud Computing, So-
cial Computing Networking, Sustainable Computing Communications
TABLE III (ISPA/IUCC/BDCloud/SocialCom/SustainCom), Dec 2018, pp. 924–931.
R ESULT OF FAILURE CLASS PREDICTION FOR
TEMPERATURE OF DIFFERENT PAIR OF SECTIONS IN
MACHINE
Features Precision Recall F-measure
B80 & B81 0.86 0.94 0.90
B80 & B82 0.91 0.79 0.85
B80 & B240 0.93 0.71 0.81
B80 & B241 0.78 0.92 0.84
B81 & B82 0.92 0.78 0.85
B81 & B240 0.81 0.94 0.87
B81 & B241 0.92 0.78 0.85
B82 & B240 0.94 0.73 0.82
B82 & B241 0.78 0.93 0.85
B240 & B241 0.92 0.66 0.77

1944