Comput Syst Sci & Eng (2020) 2: 81–89 International Journal of

© 2020 CRL Publishing Ltd Computer Systems

Science & Engineering

Intelligent Choice of Machine

Learning Methods for Predictive
Maintenance of Intelligent
Machines
Marius Becherer∗, Michael Zipperle† and Prof. Dr. Achim Karduck‡
Furtwangen University

Machines are serviced too often or only when they fail. This can result in high costs for maintenance and machine failure. The trend of Industry 4.0 and
the networking of machines opens up new possibilities for maintenance. Intelligent machines provide data that can be used to predict the ideal time of
maintenance. There are different approaches to create a forecast. Depending on the method used, appropriate conditions must be created to improve the
forecast. In this paper, results are compiled to give a state of the art of predictive maintenance. First, the different types of maintenance and economic
relationships are explained. Then factors for the forecast are explained. Requirements for the data are collected and algorithms for machine learning are
presented. Based on the relationships found, a process model is presented that shows a fast implementation of the predictive maintenance for machines.

Keywords: Predictive Maintenance, Machine Learning, Artificial Intelligence, Smart Machines, Industry, Process Model

1. INTRODUCTION The predictive maintenance should determine the ideal time for
maintenance and thus increase the availability of the aircraft.
Maintenance is an integral part of the machine life cycle, Data is needed to create a prediction. Therefore, an
checking the machine status and repairing defects. In this infrastructure that collects and stores data of a machine by
way, service life can be extended, and maintenance repair sensors is needed. Specific data requirements arise during the
costs reduced. Especially with large machines, high costs for prediction generation. For example, these must correspond to
extensive maintenance work and machine failures can be the the context of the application case and, on the other hand, useless
result. For this reason, needs-based maintenance is essential or incorrect data should be excluded. The data then serves as
to minimize production downtime and avoid high maintenance the basis for a selected algorithm, which provides the forecast
costs due to machine component failures. The company for the maintenance time of a machine. There are different
Trenitalia invests approx. 1.3 billion euros for the maintenance algorithms for a forecast. For predictive maintenance, the main
of its trains [1]. At Airbus, there are different levels of algorithms used are those of machine learning. These have
maintenance for the aircraft, but the maintenance of an aircraft different characteristics and requirements. Which algorithm is
can take several weeks and thus cause high costs [2]. Predictive most suitable depends on the application.
maintenance is used to reduce costs and maintain efficiency. This article will present the possibilities of predictive main-
tenance and show how this process is structured and can be
∗ [email protected] implemented. For this purpose, section II first references related
† [email protected] works, which reflect the importance of predictive maintenance.
‡ [email protected] Then, in section III, the theoretical basics of maintenance are

vol 35 no 2 March 2020 81

 INTELLIGENT CHOICE OF MACHINE LEARNING METHODS FOR PREDICTIVE MAINTENANCE OF INTELLIGENT MACHINES

Figure 1 Maintenance times of the different stages.

discussed, introducing the different levels of maintenance and different maintenance levels [6]. Depending on the use case, the
their economic impact. Afterwards, section IV explains the arguments can be weighted more or less heavily.
process of predictive maintenance in more detail. The sub- Various factors are essential for the implementation of
processes of data collection, data pre-processing, algorithms predictive maintenance, such as IoT and significant data
and evaluation are examined in detail. Finally, in section V, the Architecture for predictive maintenance [7], the data and the
connection between data, algorithm and use case is explained. selected algorithm [8]. For the algorithms, supervised and
Based on these relationships, a procedural model is presented, unsupervised learning is used. These also have different
which shows the process for a quick implementation of predictive strengths and weaknesses [9]. Because of the different data and
maintenance. At the end of section VI, the gained knowledge is use cases, different algorithms from the field of machine learning
summarized, and an outlook on further development is given. can be used [10]. To implement predictive maintenance with less
data is a challenge. The research deals with the transfer of use
cases [11].
2. RELATED WORK Thereby algorithms of already existing use cases are used,
which are used in the same or similar contexts.
There is an excellent potential for optimization in the main-
tenance of machines. By using predictive maintenance, the
availability of a machine can be increased by up to 15% and 3. FUNDAMENTALS
costs can be reduced by up to 25% [3]. The report “Predictive
Maintenance - predict the unpredictable” shows that only a For a prediction, data are required. Intelligent machines use
small percentage of them predict the time of maintenance. sensors to record current characteristic values that describe the
The report also provides introductory information on predictive condition of a machine. Often the term digital twin is used, which
maintenance [4]. First of all, different maintenance levels and represents the machine in a digital image. As shown in Fig. 1,
areas of application are presented. This includes a survey result maintenance can be divided into different maintenance levels.
regarding predictive maintenance. Besides, two use cases in the The purest form is reactive maintenance. Here, no planning or
area of mobility and chemistry are described. further action takes place, but rather a reaction is made when
Further current figures on economic efficiency are contained in the machine fails. This leads to unplanned breakdowns and
a report by BHGE [5]. This shows for the oil recovery application can cause significant damage because smaller defects are not
that predictive maintenance reduces the failure rate and saves repaired early enough.
costs compared to other maintenance levels. Although predictive The next stage is preventive maintenance. Maintenance takes
maintenance has the lowest failure rate, it is not necessarily the place at fixed time intervals. Resource planning can be carried
best alternative. There are advantages and disadvantages to the out at an early stage, but maintenance costs increase because

82 computer systems science & engineering

M. BECHERER ET AL.

Figure 2 The optimum maintenance time.

maintenance can take place too often. However, maintenance are documented in the oil and gas industry. The unplanned
can also take place too rarely, which increases repair costs. This outages were reduced by half with the predictive maintenance
means that preventive maintenance does not allow for needs- compared to reactive and preventive maintenance. Furthermore,
based maintenance. With proactive maintenance, the ma- chine the costs for maintenance were halved since the prediction could
is monitored in real-time. Various events can be triggered determine the ideal time for maintenance. This not only reduced
when specific values are exceeded. This principle allows small unplanned outages but also saved costs.
errors to be fixed, thus preventing critical errors. Here, too,
unplanned breakdowns can occur, since even small errors lead
to machine downtimes and these are not detected in advance. 4. PROCESS
To be able to carry out proactive maintenance, sensors must
be installed on the machine to be inspected so that the current In the following, the process chain for implementing the
status can be recorded. The highest level of maintenance is predicted maintenance of a machine is explained. Predictive
predictive maintenance. A forecast is created based on a data maintenance is a complex process that can be divided into the
basis. The forecast is intended to provide timely information following four sub-processes:
about necessary maintenance before the machine breaks down.
Furthermore, the forecast can be used to optimize resource 1) Data collection
planning. For example, replacement machines can be provided,
or spare parts can be ordered in advance. 2) Data preprocessing
Two types of costs can be distinguished for maintenance: 3) Algorithm
preventive and repair costs, which are shown in Fig. 2.
Preventive costs are high at the beginning, as the machine is 4) Evaluation
not operated during permanent maintenance. If there is no
maintenance, the preventive costs are reduced. On the other These sub-processes build on each other so that the result
hand, there are repair costs. These are low at the beginning of one process step influences the quality of the next process
because the material wear of machine parts takes place over step. If, for example, incorrect data is captured, this affects
time. The longer the machine is in operation, the greater the the result of the algorithm. Furthermore, the design of the
probability of major breakdowns and increased costs. Finally, individual sub-processes depends on the respective application.
it is important to find the minimum maintenance costs. By This determines which data can be collected and which results
adding the preventive and repair costs, the total cost curve can be the forecast should deliver. It is, therefore, essential that the
established. The ideal maintenance time cannot be determined maintenance requirements are defined first. These requirements
by time or events but must be forecast. can then be used to determine individual sub-processes. In the
Many companies maintain reactively and preventively, as following sections, individual sub-processes are described in
shown in Fig. 3. Proactive and predictive maintenance requires more detail.
an infrastructure that captures behaviour and current data from
machine sensors. Depending on the machine, reactive and
preventive maintenance is more economical than investing in 4.1 Data Collection
infrastructure. Ultimately, for an assessment, the additional
downtime costs must be compared to the investment costs. This sub-process forms the basis of the predictive maintenance
An example that predictive maintenance leads to a reduction because no meaningful forecast can be made without data from
in costs and failure rate is shown in Fig. 4. These developments a machine. Data can be collected from different sources,

vol 35 no 2 March 2020 83

 INTELLIGENT CHOICE OF MACHINE LEARNING METHODS FOR PREDICTIVE MAINTENANCE OF INTELLIGENT MACHINES

Figure 3 Used maintenance methods [4].

Figure 4 Maintenance methods used [5].

which may differ depending on the application. The sources various sensors that are used to monitor the function-
provide either temporary data, which serves to test and compare ing of the machine. A key assumption in predicted
the machine condition or static data, which form the basis maintenance is that the condition of a machine will
for comparison. The following sources can be used for data deteriorate over time in routine operation. The data
collection: contains time-dependent features that detect this aging
pattern and anomalies that lead to a deterioration in
• Temporary data sources: machine condition. The time aspect of the data is
◦ Operating status of the machine: The sensors installed necessary for the algorithm to predict errors [8].
in a machine provide metrics about the current status.
This metrics such as the temperature or speed of an ◦ Monitoring sensors: These types of sensors are
engine can be shown on display. Machines contain installed subsequently and are used for machine

84 computer systems science & engineering

M. BECHERER ET AL.

monitoring. For example, acoustic sensors can be ◦ Distillation error: In many machines, data from
used to analyze the noise development of a machine. sensors are processed and combined before being
If the noise develop- ment does not correspond to the stored in a database. This reduces the complexity
normal behaviour, a possible fault condition of the of the data and the noise of the sensor data. However,
machine may exist [4]. this process can also result in errors or the aggregation
of data that is particularly important for predicted
• Static data sources: maintenance [12].
◦ Historical maintenance logs: Maintenance logsexist ◦ Error in data integration: The collected data comes
for machines that were previously maintained man- from different sources, which are stored in a
ually by an inspector. These maintenance logs database at different times using different methods.
document the machine status at a specific point in time Furthermore, in practice, it may be necessary to
and contain details of replaced components and repair combine data from different sources. When merging,
activities per- formed. Maintenance logs provide inconsistencies have to be resolved, and errors may
much information, especially about the behaviour of occur. Individual data sets may differ in their units
a machine, and are therefore an important source of and measurement period [12].
data for predictive maintenance [4].
• Quantity: There is no general answer to the question of
◦ Fault logs: Machines provide fault logs in which
how much data is needed for the predicted maintenance.
occurred faults are documented. These can be used
The quantity required depends on the application case or
to determine hardware or software errors.
the context of the problem to be solved. For predictive
◦ Machine metadata: The metadata of a machine maintenance, it is important to have data sets where the
includes the manufacturer, model, date of manufac- machines are in unexpected operating states. If, for
ture, date of commissioning, location of the system example, only data records of a machine in ideal condition
and other technical specifications such as throughput (without errors) are available, predictive maintenance
and errors per day[8]. cannot be performed based on this data. Therefore, data
records where a machine goes into a fault condition are
Thus there are several possibilities to collect machine data. particularly important[8].
However, not all data is the same. The accuracy of a fore- cast
for maintenance depends on the relevance, quantity, and quality This section described the data sources, the requirements for
of the data. This results in the following additional requirements these and possible sources of error. The next section describes
for them: methods for preparing this data so that an algorithm can use it
for predictive maintenance.
• Relevance: The data must be relevant to the problem to
be solved. This means that a complex machine cannot be
considered as a whole system. It is, therefore, necessary 4.2 Data Preperation
to divide a machine into components and to examine these
individually. A domain expert can divide a system into The aim of this sub-process is the pre-processing of the data,
components. For example, in a motor vehicle, the engine whereby the data sets are optimized and prepared for the
temperature cannot be used for the predictive maintenance algorithm used. The input format of the data is determined by
of the transmission [8]. the algorithm used. Besides, the sources of error presented in
• Quality: There are numerous sources of error in data col- section IV-A must be analyzed and excluded.
lection, which can ultimately lead to insufficient quality
• Temporary data: Temporary data reflects the current
of the data for predictive maintenance. It is therefore
machine status. Each data set includes the time and a
important to avoid these sources of error; methods for doing
detailed description of the behaviour and the associated
so are explained in the next section. Possible sources of
sensor val- ues. This machine data is recorded periodically,
errors are listed below:
but may be recorded too often or rarely. If they are
◦ Error in data entry: In many cases, data is still recorded too often, two consecutive measured values may
entered by humans or transmitted by voice. This be identical. In this case, these measured values can be
can lead to typing errors or misunderstandings during removed to reduce data complexity. In the other case, a
communication [12]. mathematical procedure can be used to supplement missing
measured values. For example, this can be done by
◦ Measurement error: Measured values are manually averaging two consecutive measured values to obtain an
measured and read by humans. Errors may occur average value.
when applying the measurement method or reading
the measured value. However, these errors can be • Static data: Static data can contain data records that are not
avoided by using sensors and automated data storage. relevant to the problem to be solved. These data records
Nevertheless, measurement errors can also occur with must be filtered out and removed. Also, data sets can
sensors, for example, if several sensors interfere with be combined to form a new data set, and this is called
each other or are influenced by the environment [12]. feature development. Furthermore, the data sets can be

vol 35 no 2 March 2020 85

 INTELLIGENT CHOICE OF MACHINE LEARNING METHODS FOR PREDICTIVE MAINTENANCE OF INTELLIGENT MACHINES

Table 1 Labeled data sets for supervised learning.

Time stamp Feature 1 Feature 2 Lifetime Lifetime <1 day
01.01.2018 0,1432 8,532 4 false
02.01.2018 0,1858 7,942 3 false
03.01.2018 0,2145 6,606 2 false
04.01.2018 0,3243 4,348 1 true

categorized, for example, according to repair measures. of set X can be assigned by the classification to an element
Thus, repair actions can be assigned to records that trigger of set Y in the form of a type or category [14]. Table I
them. shows labeled records of a specific system component. For
each record, there is a timestamp and values for certain
characteristics. The last two columns represent values to
4.3 Algorithms be predicted by supervised learning. The value for the
”life days” describes how many days this component will
In the previous section, the sub-processes data collection and be without maintenance. Whereas the value for ”lifetime
data preprocessing, which form the basis for the algorithm, were <one day” describes whether the component will fail in the
examined in more detail. This section shows possible algorithms next 24 hours. Both values can be predicted by classifying
that can use the collected and preprocessed data to predict the the data sets. These two values are not only an example
ideal time for maintenance. Which type of algorithm is best of the classification but also represent the added value
suited for a particular application case is explained later in the in the area of predicted maintenance. By stating these
context of the procedure model in section V. values, a system component can be maintained at the right
Algorithms from the field of machine learning are used for time, and system failures can be avoided [15]. Decision
maintenance prediction. The algorithms learn from example trees are suitable for the predicted maintenance. Fig. 5
data and recognize patterns in them. These patterns are stored shows the decision tree, which can be obtained from the
as mathematical functions in a model. A model always refers records of table I. Furthermore, Long Short-Term Memory
to a specific system and can be used on any number of such (LSTM) algorithms, which belong to the neural networks,
systems. In the case of similar systems, an existing model can be used [10]. LSTM algorithms are often used to
can be adapted so that it can also be used there. This means predict rare events in time series data. Since machine
that no new pattern recognition is necessary, and time can errors are rare, LSTM algorithms are suitable for predicting
be saved since the previous process steps no longer need to maintenance [15].
be run through, and the learning process does not have to be
carried out. Furthermore, the algorithms of machine learning • Regression: In regression, however, the quantity Y is
can be divided into two categories, supervised and unsupervised continuous. This means that an element of the set X can
learning. Ultimately, the goal of the two categories is identical, be assigned a numerical value to an element of the set Y
namely to learn a mathematical function f :X → Y . However, through regression [14]. Example algorithms for this are
the sets X (input data) and Y (output data) differ in their linear regression and regression trees[10].
nature. In the following, supervised and unsupervised learning,
as well as sample algorithms for these, will be examined in more 2) Unsupervised Learning: While in supervised learning
detail [13]. the algorithm has sufficiently large amounts of marked data
1) Supervised Learning: In supervised learning, an algorithm sets available, in unsupervised learning only unmarked input
has a sufficiently large amount of input and output data available. data is available. Unsupervised learning is intended to find
The data records are already selected; that is, an input X is hidden structures in the input data in order to assign them to
assigned an output Y. The task of the algorithm is to learn a different groups. This process is also called clustering. For
mathematical function with which the output Y can be generated example, a group can represent the normal behaviour of a
based on the input X [14]. The marking of the data is done by a component of a machine. If anomalies occur, an error and
domain expert. For this purpose, the domain experts must have necessary maintenance can be inferred. Unsupervised learning is
sufficient knowledge about the system behaviour. A significant particularly suitable for predictive maintenance of machines that
advantage of supervised learning is that the data sets can be do not have historical maintenance and error logs. A significant
divided into training and test data sets. The training records are advantage of unsupervised learning is that no domain experts
then used for supervised learning. The resulting model can then are required for data tagging. Thus, the process of unsupervised
be evaluated with the test data records. With regard to predicted learning is more automated and requires less manual preparation
maintenance, data sources such as historical maintenance logs of data sets than with supervised learning. The significant
and error logs are particularly important for this. The result of differences between classification and clustering are, on the
the monitored learning is either classification or regression of one hand, the data and on the other hand, that in classification,
the set X. the target categories are already given by the data tagging. In
clustering, on the other hand, the target categories are created
• Classification: In classification, quantity Y is discrete. This by the algorithm. This has the disadvantage that the learned
means that the elements of this set are clearly isolated from model is difficult to evaluate. For unsupervised learning, LSTM
each other and can be clearly separated. Thus, an element algorithms can be used. These can be used for supervised and

86 computer systems science & engineering

M. BECHERER ET AL.

Figure 5 Classification with decision tree.

unsupervised learning, depending on the implementation of the general requirements of the predicted maintenance. The general
loss function [15]. Furthermore, clustering algorithms such as requirements include the type, the possible components and the
hierarchical clustering, K-Means, and C-Means can be used to environment of a machine. On the other hand, the data supplied
detect anomalies in the data sets. They enable early prediction by the machine and which maintenance tasks are to be forecast.
of errors and thus timely maintenance of the component [16]. The requirements for the data described earlier must be observed.
In general, supervised learning provides a more accurate Based on the general requirements, an algorithm can be selected
prediction because the model can be more easily evaluated by to obtain the desired result. The conclusion to be drawn from
the test data. If the data of a machine component meets the this is that the design of the process chain presented in section
requirements of supervised learning, an algorithm from this area IV depends on the respective application of a machine. For
is preferred. example, if the use case changes, other data must be collected,
and the selected algorithm must be adapted. In the following, a
procedure for the implementation of Predictive Maintenance is
4.4 Evaluation presented. This process model is shown in Fig. 6 and is intended
to show the process of how a forecast for a use case can be
The evaluation represents the last sub-process of the process implemented with as little effort as possible.
chain and deals with the reaction to the result of the underlying The procedure is cyclical since individual steps have to be
sub-processes. By continuously monitoring the components optimized depending on the desired result. In the beginning, the
of a machine, continuous condition monitoring can be carried use case must be defined with exact requirements and measurable
out. With the help of an application directly on the machine goals. Measurable objectives are essential for the evaluation so
or centrally on a server, current statuses can be displayed. The that a correct evaluation of the results is possible.
trend of Industry 4.0 and the related networking of machines In the first iteration, existing models of supervised learning
offer numerous possibilities for central monitoring and predicted are analyzed. A model might already exist, which is used in
maintenance. If an algorithm detects an anomaly, various other the same or similar context. If such models exist, they can be
processes can be started. On the one hand, automatic resource tested and compared with the requirements and objectives. If the
planning for spare parts can be carried out. If a component objectives are not met, or no models exist, the second iteration
threatens to fail, it must be checked whether this component is is performed.
still available as a spare part in the warehouse. If this is not the The second iteration considers the available data that are
case, it must be ordered from the vendor. On the other hand, necessary for model creation. Based on this data, an algorithm
automatic resource planning can be created for the inspector so from the area of supervised or unsupervised learning can be
that he can maintain the machine in time and replace the affected selected. If the objectives are not achieved with the selected
components. Furthermore, in the event of an impending machine algorithm, other models or algorithms can be tested. Mainly, the
failure that can no longer be prevented, the employees can be steps described in Section IV can be followed to create a model.
scheduled in good time for other work. If the goals are still not achieved because the data quality and
quantity are too low, the third iteration is performed.
5. PROCEDURE MODEL In the third iteration, the existing infrastructure is adapted to
collect the required data. The data can be collected using built- in
In order to create a concept for predictive maintenance, a hardware such as temperature sensors. However, sometimes this
concrete use case is important. A use case defines the data is not enough, and additional hardware has to be mounted

vol 35 no 2 March 2020 87

 INTELLIGENT CHOICE OF MACHINE LEARNING METHODS FOR PREDICTIVE MAINTENANCE OF INTELLIGENT MACHINES

Figure 6 Procedure model to set up predictive maintenance.

to the machine. To determine whether additional hardware is evaluation. There are temporary and statistical data sources
required, first, only the built-in hardware is used, and a machine for data collection. Temporary data sources provide real-
learning model is implemented. If the evaluation of the model time data. Static data sources provide data over some time,
shows that predictive maintenance is inaccurate, additional such as maintenance and error logs. Essential requirements
hardware is required. Additional hardware are mostly sensors, for the data are relevance, quality, and quantity. The data
which can sense how the machine behalf, typical examples are must be transformed into a format that the algorithm can use.
temperature or sound sensors. In order to use the gathered Furthermore, incorrect and redundant data must be filtered
data from the sensors efficiently, the data is stored in a central out. Algorithms from the area of supervised and unsupervised
database. If the machine is already connected to a database, it is learning can be used for prediction. Supervised learning
recommended to integrate the sensor data into its data structure. is suitable for predicting the maintenance of machines with
Otherwise, the sensors can be connected to a central IoT gateway sufficient static data. It provides high accuracy because the
to store and analyze the data and to add additional business logic model can be evaluated using the data markers. Unsupervised
there. Since in this iteration, no historical maintenance and error learning is suitable for machines with insufficient static data.
logs are available, a model for unsupervised learning is suitable. The algorithm learns from the temporary data and performs
The data to train the model can be accessed from the database. clustering. If anomalies occur, an error can be inferred. Based on
the forecast, further processes can be started, such as optimizing
resource planning for spare parts and personnel. The collected
6. CONCLUSION findings show that the individual sub-processes of the predictive
maintenance depend on the use case. The process model
The paper explores the many factors that influence the predictive presented is based on these assumptions and is intended to
maintenance process. First of all, it is important to know the simplify the forecasting process for machines. The approach
different maintenance levels. It turns out that higher maintenance here leads from an existing solution to in-house development.
levels can save costs and reduce machine downtime. However, Ideally, existing solutions can be used and optimized.
this requires data that can be collected with the appropriate As discussed in the paper, currently, hardly any forecasts are
infrastructure. For the adapted infrastructure, the investment used to determine the lifetime of maintenance. The biggest
costs increase due to the additional hardware. Depending on challenges for companies are data collection or lack of re-
the application, the possible effects of a machine failure must be sources to develop their systems for predictive maintenance.
compared to the investment costs. Predictive maintenance leads One reason for this is that older machines are often not yet
to a reduction in the failure rate and costs. networked and therefore no historical data is available. Cloud
Predictive maintenance can be divided into four process vendors such as Amazon, Microsoft and IBM offer solutions
steps: data collection, data pre-processing, algorithms, and to easily network machines and collect their data centrally.

88 computer systems science & engineering

M. BECHERER ET AL.

They also offer existing models, or new models for predictive 8. M. Azure, “Predictive maintenance”, 2018. [Online]. Available:
maintenance can be created. The presented process model can https://docs.microsoft.com/en-us/azure/machine- learning/team-
be implemented with these cloud providers and also enables data-science-process/cortana-analytics-playbook-predictive-
smaller companies to implement predictive maintenance for their maintenance (visited on 12/14/2018).
machines. 9. P. Peeling, “Big data and machine learning for predictive
maintenance”, 2017. [Online]. Available: https://www.
matlabexpo.com/content/dam/mathworks/mathworks-dot-
com/images/events/matlabexpo/uk/2017/big-data-machine-
REFERENCES learning-for-predictive-maintenance.pdf (visited on 12/08/2018).
10. Frauenhofer-Gesellschaft, “Maschinelles lernen. eine analyse
1. SAP, “Creating a dynamic maintenance management zu kompetenzen, forschung und anwendung”, 2018. [Online].
system powered”, 2015. [Online]. Available: Available: https://www.fraunhofer.de/de/presse/presseinforma
https://www.sap.com/assetdetail/2015/12/b6caea0d-507c-0010- tionen/2018/juni/maschinelles-lernen-kompetenzen-forschung-
82c7-eda71af511fa.html (visited on 11/10/2018). anwendung.html (visited on 11/18/2018).
2. Airliners, “Wartung und Uberholung von flugzeugen folgen strik- 11. Context Awareness in Predictive Maintenance. Berlin: Springer
ten regeln”, [Online]. Available: http://www.airliners.de/wartung- International Publishing, 2016, pp. 197–211.
ueberholung-flugzeugen-regeln/35275 (visited on 10/20/2017). 12. J. M. Hellerstein, “Quantitative data cleaning for large
3. McKinsey, “Smartening up with artificial intelligence (ai) databases”, 2018. [Online]. Available: http://db.cs.berkeley.edu/
- what’s in it for germany and its industrial sector”, 2017. jmh/papers/cleaning-unece.pdf (visited on 12/14/2018).
[Online]. Available: https://www.mckinsey.com/ /media/McK- 13. Wikipedia. Maschinelles lernen, [Online]. Available:
insey/Industries/Semiconductors/Our%20Insights/Smartening% https://de.wikipedia.org/wiki/Maschinelles_Lernen (visited on
20up%20with%20artificial%20intelligence/Smartening-up-with- 01/08/2019).
artificial-intelligence.ashx (visited on 11/10/2018). 14. Maschinelles Lernen: Grundlagen und Algorithmen in Python.
4. PWC, “Predictive maintenance 4.0 - predict the unpre- Berlin: Springer International Publishing, 2018, pp. 197–211,
dictable”, 2017. [Online]. Available: https://www.pwc.nl/nl/ ISBN: 9783446457058.
assets/documents/pwc-predictive-maintenance-4-0.pdf (visited on 15. I. Downard, “Conquering the challenges of data preparation
11/10/2018). for predictive maintenance”, 2019. [Online]. Available:
5. BHGE, “The impact of digital on unplanned downtime”, https://www.infoq.com/articles/challenges-of-data-preparation-
2016. [Online]. Available: https://www.bhge.com/sites/default/ for-predictive-maintenance (visited on 12/27/2018).
files/2017-12/impact-of-digital-on-unplanned-downtime- 16. N. Amruthnath and T. Gupta, “A research study on
study.pdf (visited on 11/10/2018). unsupervised machine learning algorithms for early fault
6. embitel, “Does the future of industrial asset management detection in predictive maintenance”, 2018. [Online]. Available:
belong to predictive maintenance?”, 2018. [Online]. Available: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=83871
https://www.embitel.com/blog/embedded-blog/does-the- 24&isnumber=8387054 (visited on 12/27/2018).
future-of-industrial-asset-management-belong-to-predictive-
maintenance (visited on 12/12/2018).
7. W. Yu, T. S. Dillon, F. Mostafa, W. Rahayu, and Y. Liu, “A global
manufacturing big data ecosystem for fault detection in predictive
maintenance”, IEEE Transactions on Industrial Informatics, 2019.

vol 35 no 2 March 2020 89