Content uploaded by Christian Krupitzer
Author content
All content in this area was uploaded by Christian Krupitzer on Mar 29, 2020
Content may be subject to copyright.
A Survey on Predictive Maintenance for Industry 4.0
Christian Krupitzera, Tim Wagenhalsb, Marwin Z¨
uflea, Veronika Lescha, Dominik Sch¨
aferc, Amin Mozaffarind,
Janick Edingerb, Christian Beckerb, Samuel Kouneva
aSoftware Engineering Group, University of W¨urzburg, W¨urzburg, Germany
bChair of Information Systems II, University of Mannheim, Mannheim, Germany
cSyntax Systems GmbH, Weinheim, Germany
dMOZYS Engineering GmbH, W¨urzburg, Germany
Abstract
Production issues at Volkswagen in 2016 lead to dramatical losses in sales of up to 400 million Euros per week.
This example shows the huge financial impact of a working production facility for companies. Especially in the
data-driven domains of Industry 4.0 and Industrial IoT with intelligent, connected machines, a conventional, static
maintenance schedule seems to be old-fashioned. In this paper, we present a survey on the current state of the art in
predictive maintenance for Industry 4.0. Based on a structured literate survey, we present a classification of predictive
maintenance in the context of Industry 4.0 and discuss recent developments in this area.
Keywords: Predictive Maintenence, Forecasting, Anomaly Detection, Industry 4.0, Industrial IoT
1. Introduction
Maintenance has always been a severe cost driver
in the production industry. Studies show that depend-
ing on the industry between 15 and 70 percent of total
production costs originate from maintenance activities
[1]. Nevertheless, the majority of the production indus-
try still relies on outdated maintenance policies and fo-
cuses on an inefficient run to failure approach or statisti-
cal trend driven maintenance intervals [2].Thus, leading
to reduced production time and product quality as a re-
sult of inappropriate maintenance policies. According
to [2] surveys on maintenance show that about 33 cents
of every dollar spent on maintenance in the US is wasted
because of unnecessary maintenance activities. On the
other hand, comprehensive research with regard to mod-
ern maintenance policies using modern technologies is
conducted in different fields of academics such as com-
puter science, production and artificial intelligence. The
usage of well-developed sensors and prognostic tech-
niques allows a relatively reliable prediction of the re-
maining useful life of plant equipment. This so-called
predictive maintenance policy is especially of interest in
the environment of Industry 4.0 and severely enhances
the efficiency of modern production facilities.
Email address: christian.krupitzer@uni-wuerzburg.de
(Christian Krupitzer)
Predictive maintenance is based on the idea that cer-
tain characteristics of machinery can be monitored and
the gathered date be used to derive an estimation about
the remaining useful life of the equipment. Hence, this
kind of maintenance policy implicates several important
improvements in the manufacturing and maintenance
process which can severely reduce production costs [3].
First, predictive maintenance can reduce the number of
unnecessary maintenance activities as it is not based on
periodic maintenance intervals bound to average life-
time. Thus, potentially reducing the overall number
of maintenance activities over a machine’s life. Sec-
ond, not only can too early maintenance activities be
avoided, but also too late activities as equipment might
fail before the next periodic maintenance interval since
the intervals rely on average lifetime which likely in-
cludes significant positive but also negative deviations
from the mean. For example, due to the specific struc-
ture in which a certain component is deployed in larger
machinery. Both, the reduction of unnecessary mainte-
nance as well as the reduction of fatal breakdowns result
in increased productivity and reduced production down
time. Therefore, depending on the accuracy of the prog-
nostic method applied, predictive maintenance can be
considered as an overall improvement of efficiency in
contrast to conventional maintenance policies [4],[5].
Preprint submitted to February 20, 2020
arXiv:2002.08224v1 [cs.LG] 5 Feb 2020
1.1. Objective and Approach
The existing research on predictive maintenance is
comprehensive and dates back decades with still hun-
dreds of new published papers annually [6]. Neverthe-
less, as the environment is changing and new technolo-
gies become available at a more affordable price, there
is still a wide range of potential for new research in
the field of predictive maintenance. Especially in the
context of the Internet of Things (IoT) and Industry 4.0
the possibilities to integrate predictive maintenance and
connect it to other systems of the production process
are increasing. In order to identify the potential starting
points for further research the objective of the present
survey is to structure the complex of themes regarding
predictive maintenance and to put the comprehensive
amount of research into a transparent and comprehen-
sible framework.
The methodical approach of the the present survey is
a Structured Literature Review (SLR) based on the key-
words Predictive Maintenance. Thereafter, all relevant
attributes of the selected papers are captured for con-
structing the framework.
1.2. Structure of the paper
The remainder of the survey is structured as follows:
Section 2 includes the theoretical principals and foun-
dations of predictive maintenance. Next, Section 3 de-
scribes the methodical procedure of the literature review
and the framework construction. Section 4 explains in
detail the categories and the most important attributes
of the predictive maintenance framework. Section 5 fol-
lows a discussion of the framework. Finally, Section 6
concludes the survey by summarizing the main findings
and giving recommendations for future research.
2. Foundations of Predictive Maintenance
Predictive maintenance is a maintenance policy lead-
ing to improved efficiency since it allows an estimation
of the remaining useful life of the machinery. This ap-
proach is based on condition monitoring ideally con-
ducted by sensors, which allows a continuous moni-
toring process of relevant machine parameters such as
vibration and temperature. However, condition moni-
toring in isolation cannot be considered predictive as it
only allows to identify the parameter changes that occur
before a failure, but it does not allow to predict a rela-
tively narrow future period of time in which the param-
eter changes happen and thus the failure might occur.
Thus, it does not enable those responsible to schedule
maintenance activities ahead in an efficient way, such as
finding the most cost efficient time frame for shutting
down the production process for maintenance activities.
Consequently, a reliable prognostic technique is neces-
sary to transform the acquired data into valuable infor-
mation for failure prediction. Hence, prognostic tech-
niques are an important part of predictive maintenance
and will be an important component of the predictive
maintenance framework in Section 4.
In order to clearly differentiate predictive mainte-
nance from traditional conventional maintenance poli-
cies Section 2.1 describes the most relevant categories
of maintenance policies and their unique characteris-
tics. Finally, Section 2.2 briefly explains the modern
approach of e-maintenance which integrates predictive
maintenance into a broader context of manufacturing
processes.
2.1. Conventional Maintenance Policies
Generally, maintenance policies can be divided in
2 categories: Corrective maintenance and preventive
maintenance. A categorization of conventional main-
tenance policies is shown in figure 1. Corrective main-
tenance always takes place after a failure occurred. Af-
terwards, the repairing of the machine can be done im-
mediately or at some later point. A production plant that
uses this approach is following a run-to-failure manage-
ment based on the philosophy: If it ain’t broke, don’t fix
it [2]. However, this approach can result in severe reduc-
tion of manufacturing time ad costly repairs. Corrective
maintenance is also referred to as reactive maintenance
[2]. On the other hand, preventive maintenance poli-
cies, also called proactive maintenance, are attempted to
be executed before a fatal failure occurs. Thereby, pre-
ventive maintenance policies are divided into two cate-
gories. First, preventive maintenance where the main-
tenance activities are conducted on pre-scheduled inter-
vals based on historic average equipment lifetime [7].
On the contrary, condition-based maintenance monitors
the current condition of a machine and schedules main-
tenance activities based on the observations made [7].
Here, three distinct condition monitoring methods are
feasible: Monitoring on request, scheduled monitoring
and continuous monitoring [3], [8]. The first two meth-
ods are mostly based on inspection, while continuous
monitoring is generally implemented by sensors. The
drawbacks of predetermined maintenance were already
discussed in Section ??. The issue with condition-based
maintenance is that even with continuous monitoring
the acquired data only represents a snapshot of a ma-
chines condition. The approach does not allow to ef-
ficiently schedule maintenance activities ahead due to
missing knowledge about a machine’s or component’s
2
presumable future state. However, as has already been
mentioned and is shown in figure 1 the approach of pre-
dictive maintenance is based on the idea of condition
monitoring. Hence, predictive maintenance represents
an enhancement of mere condition-based maintenance.
2.2. E-Maintenance
Mere predictive maintenance is already an improve-
ment in contrast to conventional maintenance policies.
Nevertheless, a predictive maintenance strategy which
is carried out in isolation of other relevant processes still
has potential for optimization. Hence, according to [9]
the efficiency can be severely improved by connecting
multiple machines across manufacturing locations of a
production facility and monitoring them remotely with
wireless sensors, as described in Section ??. A web ap-
plication and cloud-based monitoring infrastructure al-
lows to synchronize the process of maintenance with
the overall operation even across production facilities.
Furthermore, connected processes such as maintenance
resources and even automated spare-part ordering can
be integrated in this system [9]. Precondition for such
an e-maintenance system is to aquire and process rele-
vant data in real time. [10] argue that an e-maintenance
can bee seen as an entirely integrated system since it
handles the monitoring, diagnosis, prognosis and con-
trol process. Thus, e-maintenance goes beyond mere
predictive maintenance [9]. The integrated approach of
e-maintenance matches the concept and idea of Industry
4.0 [11].
3. Methodical Procedure
The following section describes the methodology
used to select the relevant papers and construct the
framework. In total 150 papers are selected to derive
the framework. All the attributes are clustered into 10
categories that build the main layer of the framework.
3.1. Structured Literature Review
The Structured Literature Review (SLR) is a system-
atic approach to find the relevant literature in order to
answer a set of research questions by searching for pa-
pers based on predefined key words [12]. By searching
with specific key words the method allows to find the
most relevant literature to address the research ques-
tions. However, there is no guarantee that a SLR is
able to find all relevant literature [12]. Nevertheless,
the SLR has the advantage to tackle a specified topic
from numerous directions, allowing the researcher to
properly cover the important sub themes [12]. Thereby,
the SLR allows to uncover gaps in the existing primary
research and helps to reveal areas where additional re-
search might be needed [12].
As the field of predictive maintenance is extremely
broad and has been addressed comprehensively by re-
searchers for decades, the existing pool of papers is
enormous. Therefore, the only key words used in the
present survey to detect the relevant literature with a
Google Scholar search are Predictive Maintenance. As
the objective of the survey is to construct a framework
addressing the distinctive facets of predictive mainte-
nance this procedure results in the most favorable set
of important literature since a combined search of pre-
dictive maintenance and Industry 4.0 mainly results in
papers which address Industry 4.0 but only mention pre-
dictive maintenance as part of it without explaining a
specific approach in detail. Thus, with these papers the
development of a detailed framework with all facets of
predictive maintenance would not be possible. As al-
ready mentioned, the tremendous amount of research
addressing the topic of predictive maintenance results
in thousands of potentially relevant papers which can-
not be evaluated in a single survey. Thus, the present
survey does not claim to be entirely complete and only
represents an overview. The search algorithm of Google
Scholar runs very successful as the first papers of the
search are the most suitable papers for the topic, while
the quality and usefulness of the later papers is steadily
decreasing. This allows to identify appropriate litera-
ture by checking the list of results from the top down.
The search for academic papers via Google Scholar was
conducted in October 2018. A paper for the framework
was selected by first scanning the abstract to identify
the detail in which the topic of predictive maintenance
is covered in the paper. If a paper is explaining the
applied predictive maintenance approach and its con-
text, e.g. system size and condition monitoring in de-
tail, it is utilized to build the framework. The selected
papers are published between 1993 and 2018 while al-
most two thirds (97/150) are published during the last
decade. Since the research concerning predictive main-
tenance is a very broad field it is addressed by various
types of journals and conferences.
Another literature review method that was consid-
ered for the present survey is the so-called berry pick-
ing method. In this method the review starts by identi-
fying a starting paper that matches the addressed topic
and objectives [13]. The following step in the process
is the footnote chasing where the list of references of
the starting paper is checked for more relevant litera-
ture [13]. Furthermore, the search for relevant litera-
ture can be extended by checking not only the refer-
3
Figure 1: Maintenance Policies [7]
ences of the starting paper, but also checking the pa-
pers which cite the starting paper [13]. The drawback
of this method is that the papers are most likely con-
nected based on a similar sub theme that is addressed
by these papers, thereby potentially missing important
other subsections of the main topic. Of course, this is-
sue can be approached by choosing multiple starting pa-
pers to cover the missing sub themes. Nevertheless, the
amount of literature addressing predictive maintenance
is so broad and contains numerous sub themes that even
with an increased number of starting papers the berry
picking method would still miss relevant fields. Hence,
the SLR is considered as better method for the present
survey as it allows to find more diversified relevant lit-
erature in order to cover a much broader scope of the
topic. Thus, resulting in a more favorable result for
building the framework and addressing the first research
question specified in section 1.1.
3.2. Construction of the Framework
The construction of the framework starts by examin-
ing the first paper that was selected from the top down of
the potentially relevant literature of the Google Scholar
search. Notice that it is not important with which paper
this process is started as the whole process has an iter-
ative characteristic. While a paper is examined a table
is filled with data about the relevant attributes covered
in the paper. Thus, whenever an important attribute or
characteristic concerning the topic of predictive main-
tenance is identified a new column is added to the ta-
ble and the attribute is ticked for this paper. Thereafter,
every paper that is categorized in the table gets a sepa-
rate line to mark the attributes which are covered in this
paper. Thereby, the following papers are checked iter-
atively as their research content is searched for the al-
ready existing attributes. Additionally, the following pa-
pers are checked for additional relevant attributes which
are added to the table. The data grid that is generated by
this process represents the foundation of the predictive
maintenance framework. It is important to mention that
an attribute is merely considered as covered by a paper
if it is a direct research subject of the paper. Hence, a
literature review within a primary study is not consid-
ered sufficient to fulfill certain attributes. However, the
list of categorized papers also contains 10 surveys ([6],
[9], [14], [15], [16], [17], [18], [19], [20], [10]) where
the approach is obviously different and all referenced
attributes are marked in the table.
In the next step the raw data grid is structured by
grouping attributes which are similar to each other or
belong to the same sub theme. Subsequently, each clus-
ter is named by a term which represents the content of
the cluster. These umbrella terms build the main cate-
gories of the framework. If appropriate, the clusters can
have further subdivisions to allow for a finer classifica-
tion within the categories. Thus, the entire framework
can be best represented in form of a tree. The main cate-
gories of the predictive maintenance framework, as well
as the tree structure for every category are presented in
Section 4.
4. Framework for Predictive Maintenance
In section 3.2 the methodical approach for structur-
ing the framework was explained. The analysis of the
papers resulted in a grid which consists of a total of 69
different attributes. The framework that is built from
this data grid is described in the following sections. The
entire framework for predictive maintenance consists of
10 categories. These categories are Goals, Condition
Monitoring, Maintenance Scope, Degradation Process,
Fault Detection, System Size, Scheduling, Prognostic
Techniques, Data Handling and Evaluation. Figure 2
shows these 10 categories which represent the highest
level of the framework. The following sections of de-
scribe the categories in detail and show all the attributes
which are summarized within a certain category.
4
Figure 2: Categories of the Framework for Predictive Maintenance
4.1. Goals
The category of goals is generally used by researches
to support their motivation for choosing the research
topic of predictive maintenance. Thus, 139 out of
150 papers (93%) mention one or several goals of the
framework. During the analysis of the paper 8 goals
have been identified: Spare Part Inventory Reduction,
Prolong Machine/Component Life, Cost Minimization,
Minimize Downtime, Availability, Productivity, Relia-
bility and Safety. Some of these goals were already in-
troduced in section ?? but not explained in detail. Fig-
ure 3 shows all attributes belonging to the category of
goals. The number in parentheses is the frequency of
occurrence of a certain attribute out of a total of 150
papers.
Figure 3: Overview of the Category Goals
First, predictive maintenance has the potential to re-
duce the number of spare parts in stock since a reli-
able predictive maintenance system would allow for the
possibility to merely store the spare parts needed soon
instead of storing every spare part which could poten-
tially be necessary [8]. Hence, the predictive mainte-
nance system would reduce the number of spare parts in
stock and the overall storage size while still maintaining
a proper maintenance process and avoiding production
downtime due to unavailable spare parts.
Second, some researchers argue that predictive main-
tenance can in fact prolong the overall life of machines
and components [21]. This is because the predictive
maintenance system monitors the health condition and
predicts the remaining useful life of a machine and thus
reduces the risk of a fatal breakdown which might re-
duce the lifespan of the machine even in case of a thor-
oughly corrective maintenance action [21]. Further-
more, as predictive maintenance systems also help to
avoid unnecessary maintenance actions this can addi-
tionally prolong a machine’s life because every main-
tenance action puts the machine to risk since a mainte-
nance action and replaced components represent an in-
tervention in the structure of a machine [22].
The by far most frequently mentioned goal is the ob-
jective to minimize costs through a predictive mainte-
nance approach [23]. More than two thirds of the an-
alyzed papers mention cost minimization as one of the
main goals of predictive maintenance and thus their re-
search. The goal of cost minimization is often related
to other goals of a predictive maintenance system such
as minimizing downtime by avoiding fatal breakdowns
or prolong machine life and reduce the need for spare
part inventory. As a reliable predictive maintenance sys-
tem leads to an overall more efficient way of the main-
tenance process this results in an overall reduction of
costs. Hence, while the implementation of a predictive
maintenance system is costly and complex it is argued
to be a positive business case in the long run [23].
Fourth, since predictive maintenance systems allow
fault detection in the future based on monitoring data
and prognostic techniques the number of fatal break-
downs can be reduced [24]. With a conventional sched-
uled maintenance policy based on historic data about
the life span of a machine or component, there is still
the possibility that a machine or component might fail
before the next maintenance interval since the interval
is merely an average that includes outliers. In contrast
to that, a predictive maintenance approach monitors the
machine and its components regularly or even continu-
ously and triggers maintenance actions regardless of the
time a machine or component is in use [2]. Additionally,
a predictive maintenance system can also minimize the
downtime of a manufacturing line as it allows to plan
5
maintenance actions ahead and group certain mainte-
nance actions to reduce the number of production stops
for single maintenance actions [25]. Minimizing down-
time has a severe effect on reducing costs and increasing
productivity. Thus, almost half (47%) of the papers in
this survey address the objective of minimizing down-
time.
The goals of availability, productivity and reliability
are all connected to the goals of cost minimization and
minimizing downtime [26], [27], [22]. By decreasing
downtime a predictive maintenance system automati-
cally increases the total amount of time a machine can
be in an online state. The increased overall manufac-
turing time also increases the productivity of a man-
ufacturing plant as more products can be produced in
the same time while the monitoring component of the
predictive maintenance system always makes sure that
the equipment is running in a healthy condition. Hence,
the predictive maintenance approach keeps machinery
in a state where it produces products in a good quality
[28]. Thus, this leads to an increase in productivity. The
goal to increase the reliability of machines is strongly
related to the goal availability as these two objectives
are always addressed together. Additionally, availabil-
ity, productivity and reliability increase the output and
quality in a certain time interval and therefore minimize
production costs.
Finally, safety is a goal that is not directly related
to business performance indicators such as productivity
and costs. Nevertheless, it is still an important objective
in the context of predictive maintenance research since
it is addressed by almost a third (30%) of all papers in
the present survey. At present, most production lines
still require the interaction of humans and are not fully
automated. Therefore, humans work nearby or directly
with heavy machinery. Fatal breakdowns and unreliable
running conditions of these machines cause a potential
risk to the employees. Hence, as predictive maintenance
avoids fatal breakdowns and monitors the health condi-
tion of machines, the safety of employees is improved
[15].
4.2. Condition Monitoring
A predictive maintenance approach is based on the
collection of data from a machine or component which
indicates its health status and allows the prediction of
the residual useful life based on this monitoring data
[29]. The analysis of the 150 papers has identified 4 dif-
ferent attributes for the category of condition monitor-
ing: Inspection-Based Monitoring, Sensor-Based Mon-
itoring, Continuous Monitoring and Online/Real Moni-
toring. All attributes and their corresponding frequency
of occurrence are shown in figure 4.
Figure 4: Overview of the Category Condition Monitoring
First, the least addressed approach to monitor a ma-
chine’s condition and collect valuable data is inspection-
based monitoring (16%). With inspection-based moni-
toring the data is merely collected in inspection inter-
vals. However, the intervals are not predefined as it is
the case in conventional maintenance policies. The in-
tervals are adapted with respect to the observed and col-
lected data about a machine’s or component’s current
and predicted conditional state [6].
Second, since advancements in sensor technology
make sensors for various types of parameters more af-
fordable, the majority of studies base their research on
sensor-based monitoring. With sensor-based monitor-
ing different types of sensors for instance to observe
vibration and temperature are used to collect the rele-
vant data [30]. Note that sensor-based monitoring and
inspection-based monitoring are not mutually exclusive
as it is often the case that sensor equipment is necessary
to perform the inspection [24]. Generally, the usage of
sensor-technology is more appropriate for an integrated
predictive maintenance system as it is crucial for effi-
cient continuous monitoring.
As the term already indicates, continuous monitor-
ing is the continual collection of relevant monitoring
data to estimate the remaining useful life of a ma-
chine or component [31]. In contrast to inspection-
based monitoring the amount of data collected is sig-
nificantly higher since inspection-based monitoring is
merely a periodic snapshot of a machine’s conditional
state. Hence, inspection-based monitoring and contin-
uous monitoring are the only two attributes in this cat-
egory which are mutually exclusive. All other attribute
combinations are feasible.
Finally, online/real monitoring is a condition moni-
toring technique which allows the collection of data in
the running state of a machine and is addressed by the
majority of the analyzed papers (61%) [32]. Further-
more, it is the prerequisite for continuous monitoring as
a continual collection of data is merely feasible in the
running state of a machine. Therefore, researchers al-
6
ways address online monitoring when implementing a
continuous monitoring approach in their studies. How-
ever, online monitoring is also possible for inspection-
based methods, but in this case it does not resemble a
requirement.
4.3. Maintenance Scope
The category of maintenance scope is a category
which is generally not addressed frequently. Obviously,
for every maintenance action there has to be an assump-
tion about the maintenance scope but merely a few stud-
ies (13%) mention this category directly. Therefore, the
framework solely represents the number of times the
topic of maintenance scope is addressed directly. The
present survey makes no assumptions about the main-
tenance scope in cases where no maintenance scope
is mentioned directly. The analysis of the papers re-
vealed 3 different attributes for the category mainte-
nance scope: Perfect Maintenance, Imperfect Mainte-
nance and Grouping Maintenance Actions. Figure 5
shows all attributes belonging to the category of main-
tenance scope.
Figure 5: Overview of the Category Maintenance Scope
First, the attribute of perfect maintenance is based
on the assumption that every maintenance action con-
ducted at a machine or component restores the function-
ality and durability to its original level. Hence, into the
condition of a new machine [33]. This assumption fol-
lows an as good as new approach for every maintenance
action [33]. In contrast to this approach, the assump-
tion of imperfect maintenance is based on the premise
that a maintenance action cannot restore the functional-
ity and durability of a certain machine into an as good
as new state, but only into an as good as old condi-
tion [34]. Thus, the machine or component is still as-
sumed to be in the state of used equipment even after
maintenance. More studies make the assumption of im-
perfect maintenance compared to perfect maintenance
since they argue that it is a more realistic conjecture [8].
Generally, the category of maintenance scope is not very
widespread in the predictive maintenance research, pre-
sumably because the usually continuous monitoring and
prediction updates make the assumption about the main-
tenance scope obsolete.
The third and final attribute of the category of main-
tenance scope seems to be far more important for pre-
dictive maintenance but is as well not severely repre-
sented in the academic literature. The possibility to
group maintenance actions in an efficient way leads to
an overall cost reduction for maintenance activities as
downtime can be reduced [35], [36]. Precondition for
grouping maintenance actions is a holistic predictive
maintenance approach which monitors the entire manu-
facturing equipment simultaneously in order to identify
certain maintenance actions which are best conducted at
the same time [36]. While the assumptions about per-
fect and imperfect maintenance are mutually exclusive,
the attribute of grouping maintenance actions can be ad-
dressed in combination with the other two attributes of
the category. The issue that merely a few of the ana-
lyzed papers address the important attribute of grouping
maintenance will be discussed in Section 5.
4.4. Degradation Process
The category degradation process is addressed by a
third (34%) of the analyzed papers and covers either
the direct modeling of the degradation process of a ma-
chine or a predefined assumption about its deterioration
course. The attributes identified for this category are:
Degradation Modeling, Exponential Degradation As-
sumption, Random Failure Assumption, Linear Degra-
dation Assumption and Weibull Distribution Assump-
tion. Figure 6 shows the category degradation process
and all its attributes.
Figure 6: Overview of the Category Degradation Process
First, the process of degradation modeling is the
derivation of the deterioration course of a machine or
component based on relevant machine health indication
data such as vibration or temperature [37]. The rea-
son for this procedure is to gather information about a
certain machines typical conditional state over its life-
time [37]. The modeling of the deterioration process
is beneficial in a sense that having knowledge about
a machine’s degradation pattern can support a predic-
tive maintenance system to predict future breakdowns
7
Figure 7: Vibration-based Degradation Modeling [37]
more accurately. Nevertheless, the predictive mainte-
nance approach is not meant to merely rely on the av-
erage degradation of a machine to decide on the main-
tenance intervals. The main indicator is still the pre-
dictive maintenance system and its prognostic approach
that could be supported by the information of the degra-
dation modeling. Figure 7 shows an example for an
vibration-based degradation course where phase I rep-
resents the non-defective state and phase II the condi-
tional state close to failure [37]. Thus, the objective is
to model these phases in order to support the machinery
prognostics of the predictive maintenance decision with
regard to the timing and need for maintenance activities.
A few papers make an assumption about the degrada-
tion process instead of following a modeling approach.
However, since the predictive maintenance approach is
based on monitoring the conditional state and prognos-
tics, the mere assumption about the deterioration course
of a machine is not really relevant for a predictive main-
tenance approach. Hence, only a fraction of the aca-
demic literature mentions such an assumption prior to
the implementation of an predictive maintenance ap-
proach. The assumptions about linear degradation, ex-
ponential degradation and random failure are straight
forward [38], [29], [39]. A predictive maintenance ap-
proach would be most efficient in the prsence of random
failure since conventional maintenance policies would
fail most of the time. Additionally, the academic litera-
ture mentions a distribution called Weibull Distribution
[29]. It is a continuous probability distribution based on
adjustable parameters which is used to model the lifes-
pan of machines or components [29].
4.5. Fault Detection
A pure predictive maintenance approach solely fo-
cuses on the prediction of the future conditional state of
machinery and components to schedule maintenance ac-
tivities in an appropriate way and scope. Nevertheless,
a little bit over a third (36%) of the examined academic
papers additionally address the topic of fault detection,
meaning that the predictive maintenance approach does
not only attempt to predict the remaining useful life of
the machine, but also tries to identify the root cause of
the failure based on the collected data [5]. The cate-
gory of fault detection includes the following attributes:
Root Cause Analysis, Machinery Diagnostics and No
Fault Detection. All attributes are shown in figure 8.
Figure 8: Overview of the Category Fault Detection
Fault detection covers the additional function of di-
agnostics. Thereby, root cause analysis and machin-
ery diagnostics address the same issue. While most re-
searchers call it machinery diagnostics, some also refer
to it as root cause analysis or mention both terms. The
general idea is the processing of acquired monitoring
data to uncover the reasons for future failure. Thus, us-
ing vibration or other machine monitoring data for di-
agnostic purposes [14]. The feasibility and accuracy of
a fault detection approach depends on the level of moni-
toring activity which means that the more machine parts
and components are monitored separately, the better can
be identified where the root cause for a future failure
may lie [14].
4.6. System Size
Another relevant parameter which was identified dur-
ing the literature review is the category of system size.
This category addresses the extent to which the predic-
tive maintenance approach is applied or assumed to be
applied when implemented in real life. The analysis re-
vealed two attributes: Single-Component Systems and
Multi-Component Systems. Furthermore, the attribute
of multi-component systems is further divided in the sub
category of component dependencies which are divided
into the following attributes: Structural Component
Dependence, Economic Component Dependence and
Stochastic Component Dependence. Figure 9 shows all
attributes of the category. A little bit over two thirds
8
(70%) of the papers analyzed in the present survey ad-
dress the category of system size. Hence, this category
can be seen as important in the context of predictive
maintenance.
Figure 9: Overview of the Category System Size
First, single-component systems are defined by the
present survey as either single components, e.g. ex-
perimental studies which conduct laboratory tests with
merely single bearings, or single machines considered
solely in an isolated context [40]. This means that the
machine itself might consist of multiple components but
the total machine as a whole is not considered as a true
multi-component system. The definition chosen for this
category results in less true multi-component systems
compared to single-component systems (61% vs. 39%).
As already described in the previous paragraph, the
true multi-component system consists of multiple ma-
chines or separate components which together form or
are part of larger system, e.g. an entire manufactur-
ing line [41]. The implementation of an successful
predictive maintenance system is much more compli-
cated for these multi-component systems since more
data needs to be processed and dependencies between
the system’s components become relevant [41]. How-
ever, while there is a significant number of papers ad-
dressing multi-component systems, merely 8 of the 41
papers (20%) additionally address the topic of depen-
dence.
Dependence in multi-component systems can be di-
vided in three different types. First, economic depen-
dencies are such dependencies that enable cost reduc-
tion when parts of the system are maintained simultane-
ously, e.g. because for the maintenance of one com-
ponent other components have to be offline as well,
thereby reducing downtime when maintenance actions
for these components are conducted jointly [4]. Sec-
ond, stochastic dependencies are dependencies in con-
sequence of stochastic relations between components of
their deterioration process. Hence, the degradation of
one component affects the state of one or multiple other
components of the system [4]. Finally, structural depen-
dencies result from components which form a unified
part in a sense that the maintenance of one component
directly implies the maintenance of all structural depen-
dent components [4]. The most important of these three
types of dependencies seems to be the economic depen-
dence as all papers which cover the topic of dependence
at least mention this type. Furthermore, the existence of
these dependencies is the reason why single-component
predictive maintenance approaches are not simply scal-
able to a multi-component level but must be adapted
with regard to the affects of these dependencies [4].
4.7. Scheduling
The category of scheduling is addressed by almost
two third (62%) of the papers and therefore has an im-
portant role in the field of predictive maintenance. This
does not seem to be unusual as the main motivation be-
hind a predictive maintenance approach is to identify
the need and timing for maintenance activities in ad-
vance, allowing for an efficient scheduling of these ac-
tivities. The attributes identified in this category are:
Dynamic Action Scheduling, (Autonomous) Dynamic
Spare Part Availability and No Scheduling Included.
Figure 10 shows all attributes of the category and their
corresponding frequency of occurrence.
Figure 10: Overview of the Category Scheduling
First, dynamic action scheduling describes the pos-
sibility to dynamically adapt the maintenance sched-
ule based on new and processed condition monitoring
data [25]. This dynamic scheduling is only possible in
a predictive maintenance environment due to the fore-
cast of a machine’s future conditional state. Optimiza-
9
tion algorithms can be applied to define the most cost-
effective maintenance schedule and continuously update
this schedule when new machinery prognostics infor-
mation becomes available [25]. Hence, the maintenance
schedule is not static as it would be for conventional
maintenance policies but rather dynamic.
Additionally, a few papers cover the attribute of (au-
tonomous) dynamic spare part availability, where not
only the maintenance activities itself but also the neces-
sary spare part ordering is linked to the predictive main-
tenance system [36]. This broader and more integrated
approach better fits a modern idea of Industry 4.0 and
smart factory compared to an isolated policy. However,
over a third (38%) of the analyzed academic literature
does not address the category of scheduling in conjunc-
tion with predictive maintenance directly.
4.8. Prognostic Techniques
The category of prognostic techniques is clearly one
of the most important ones for predictive maintenance.
While all the monitoring and data acquisition is indis-
pensable the prognostic technique is what transforms
the raw data into valuable information. Note that since
a prognostic technique attempts to predict a prospective
failure of a machine or component the generated infor-
mation are just probabilities. For the predictive mainte-
nance framework of the present survey 3 attributes were
identified for the category of prognostic techniques:
Data-Driven Approach, Model-Based Approach and No
Prognostic Approach specified. Figure 11 shows all 3
primary attributes as well as their secondary attributes
and their corresponding frequency of occurrence.
The analysis of the 150 papers showed that the num-
ber and diversity of different techniques is enormous.
The majority of techniques is often found only once
or twice within all the papers. Thus, the present sur-
vey will not present and discuss the different techniques
in detail. In addition, the attribute no prognostic ap-
proach specified is added to the predictive maintenance
framework since not every paper mentions the applied
approach. For the cases in which the papers mention
their approach there is no single definition of how to
classify these techniques into the attributes data-driven
and model-based. Furthermore, some papers just men-
tion that they apply some kind of model-based or data-
driven approach but do not specify this approach further,
which makes it difficult to stick to one definition for all
papers.
It is important to take into consideration that all these
prognostic techniques rely on monitoring data for their
predictions. Hence, in a broad sense all these methods
Figure 11: Overview of the Category Prognostic Techniques
are somehow data-driven. However, in the present sur-
vey data-driven approaches are such techniques which
rely heavily on previous data input in order to enable
reliable predictions in the first place [42], [43]. Good
examples are (multiple) linear regression, logistic re-
gression, regression trees, random forest and the kriging
statistical technique. All these techniques are based on
the concept of regression [43]. However, for a reliable
regression model there must be a huge amount of train-
ing data to define the model before it can actually be
applied to new test data. The same holds for bayesian-
based models. Prior to work with the bayesian principle
of believe updates the a priori probabilities must be de-
rived from training data [43].
Beside these classical statistic methods the attribute
data-driven approach also has a sub attribute category of
prognostic techniques related to artificial intelligence.
Almost 25% of the papers mention an artificial intel-
ligence approach. Most frequently the concept of ar-
tificial neural networks. Artificial intelligence concepts
also belong to the group of data-driven approaches since
they require an enormous amount of training data, too
[18], [21]. In the case of an artificial neural network
nodes which are also called neurons are structured in
multiple layers where every neuron passes on a value
to all nodes in the next layer [44], [45]. Every value
is weighted by some real number that represents the
weight of the connection between two neurons [44],
[45]. The idea of the network is that a specific input
10
results in a specific outcome with a certain probability
[44], [45]. The network mostly depends on the weights
of every connection. However, finding these weights is
no easy task and requires a huge amount of training data
in order to build a reliable artificial neural network [44],
[45]. In total 73% of the 124 papers which mention a
prognostic technique, mention a data-driven approach.
As already mentioned, the difference between data-
driven and model-based approaches is not straightfor-
ward but rather vague. Generally, a data-driven prog-
nostic technique is based on a large amount of available
historic monitoring data while a model-based technique
is applied for new or untested systems which lack this
comprehensive measurement data [42]. Furthermore,
about 8% of the papers mention both the data-driven
and the model-based approach. In the predictive mainte-
nance framework of the present survey model-based ap-
proaches are defined as techniques which process moni-
toring data but do not require a great amount of training
data before they can be applied. However, the allocation
of techniques into this attribute is vague and often based
merely on the fact that a method does not fit the defini-
tion of a data-driven approach properly. Furthermore,
the framework shows that about a third (31%) of the
papers apply a model-based approach but the frequen-
cies of the sub attributes are very low. This is the case
since a lot of the papers in the model-based attribute
merely mention the usage of some type of model for the
machinery prognostics part. Thus, even more making
the allocation within this framework category a vague
process. However, the main purpose of this category is
not to find a generally valid classification of every sin-
gle prognostic technique as this is not feasible, but to
demonstrate the variety and number of techniques iden-
tified after analyzing just 150 papers of the research area
of predictive maintenance.
4.9. Data Handling
The category data handling deals with the amount of
data acquired by condition monitoring. Especially in
cases of continuous monitoring with multiple sensors
the amount of data collected by these sensors results in
an enormous amount of data to be handled [20], [11].
The analysis of the 150 papers covered in the present
survey reveals the following relevant attributes: Data
Fusion, Data Filtering, Storage Location and Data Ac-
cess. Additionally, the attributes of storage location and
data access are divided into the sub attributes of Local
Date Storage and Remote/Cloud Data Storage as well
as Local Data Access and Remote Data Access. Figure
12 shows all attributes of the category and their corre-
sponding frequency of occurrence.
Figure 12: Overview of the Category Data Handling
First, data fusion and data filtering are two methods
which deal with the issue of big data generated by con-
tinuous monitoring. The approach of data fusion uti-
lizes the idea to integrate multiple data sources in order
to generate more reliable data compared to any individ-
ual data source. In case of sensor-generated data, data
fusion combines the data of multiple sensors and the re-
sulting more reliable and more accurate data is then pro-
cessed and analyzed [18], [9]. On the other hand, data
filtering also deals with the issue of large amounts of
data by filtering uninformative data [46]. For the most
part continuous monitoring will generate data which
just shows that a certain machine is in a normal state
[47]. Therefore, data filtering models identify the use-
less data and only analyze the informative parts of the
total amount of data, thus making the data processing
more accurate and efficient [48]. However, merely the
minority (14%) of the papers directly address either data
fusion, data filtering or both.
Second, in contrast to data fusion and filtering, data
storage and access are attributes addressed by all papers.
This is the case since every paper includes some type of
monitoring which inevitably generates data that needs
to be stored and made accessible for further process-
ing. The analysis shows that for both, storage location
and data access, the remote alternative is chosen by only
about 20% of the papers and mostly in combination with
IoT, Industry 4.0 and cloud computing [49], [11], [9].
This fact is of interest as the remote approach seems to
be more suitable for an efficient predictive maintenance
system in an Industry 4.0 environment, especially with
a manufacturing structure that consists of multiple pro-
duction sites. Hence, the method of remote storage and
access is not covered as much as expected in the recent
academic literature. This issue will be discussed in Sec-
tion 5 of the present survey.
11
4.10. Evaluation
Finally, the framework of the present survey for
predictive maintenance is concluded by the category
of evaluation. This category covers the application-
oriented part of the studies. Most of the studies use
some sort of data to test their predictive maintenance
approach and to proof its feasibility. The following
attributes were identified for this category: (Numeric)
Simulation, Evaluation based on Real Data, Experiment
and Comparison with Conventional Maintenance Poli-
cies. Figure 13 shows all attributes of the category and
their corresponding frequency of occurrence.
Figure 13: Overview of the Category Evaluation
Note that none of the attributes are mutually exclu-
sive. Nevertheless, only 6 papers out of 150 choose
a combination of numeric simulation and either real
data or experiment. First, the numeric simulation is
an evaluation based merely on simulated date, e.g. a
Monte Carlo Simulation which generates data about
hypothetical failures which must then be identified by
the researcher’s predictive maintenance approach [50],
[51]. More than a third (37%) of the papers apply a
simulation-based evaluation method to test their predic-
tive maintenance techniques. Second, 37% of the papers
choose an evaluation based on real data. For this eval-
uation method the researchers acquire actual data from
companies who monitor their machines [52], [38]. The
data includes the monitoring data as well as informa-
tion about the corresponding state of the machine, e.g.
whether a failure occurred. This data can then be used to
validate the accuracy and performance of the introduced
predictive maintenance approach. The advantage of real
data is that it allows to test the predictive maintenance
approach with long term data and with data gathered
from multiple large and complex machines.
In contrast to that, the minority of papers (13%) con-
ducts experiments to validate their approach. Mostly,
the experiments are kept very small in a sense that
merely one component, e.g. bearings or a detached
engine are used as experimental subjects [29], [53].
Thus, experiments are usually limited to small scaled
test setups which makes them unsuitable to validate ap-
proaches that include multi-component systems and de-
pendencies as well as connections with other related
processes. Finally, about 20% of the authors addi-
tionally compare their predictive maintenance approach
with conventional maintenance policies as a benchmark.
These comparisons are generally based on comparing
total costs for different maintenance policies including
predictive maintenance [41], [4]. Thus, attempting to
prove the superior efficiency and cost reduction oppor-
tunities of predictive maintenance.
5. Discussion
The number of papers analyzed in the present survey
is merely a fraction of the available academic literature
addressing predictive maintenance. However, the anal-
ysis provides insides about which aspects the literature
mostly focuses on. This discussion concentrates on four
aspects that emerged during the construction and anal-
ysis of the predictive maintenance framework. First, an
interesting aspect in this context is the topic of depen-
dencies. Although 41 papers state that they take multi-
component systems into account, merely 8 papers ad-
dress the topic of dependencies. This implies that al-
though predictive maintenance is more efficient when
applied to an entire production system instead of sin-
gle machines and when taking dependencies into con-
sideration, the existing literature frequently focuses on
a too narrow view on predictive maintenance. Many re-
searchers either apply their approach merely to a single
machine or they do not focus on dependencies of any
of the three types described in subsection 4.6. Nonethe-
less, dependencies are severely important when attempt-
ing to design an efficient and integrated predictive main-
tenance system which can be applied to an entire manu-
facturing process. Thus, in this area seems to be a gap in
the existing research and further research should focus
on a broader scope for predictive maintenance systems.
Second, to date a minority of papers address mod-
ern possibilities of remote data storage and remote data
access. However, with the available technology and
improving internet connections for a faster transfer of
data, future predictive maintenance approaches should
use these technologies more severely in order to cre-
ate a more centralized, networked and therefore more
efficient predictive maintenance. This is especially of
interest for companies with multiple manufacturing lo-
cations. When all the information can be accessed and
processed at a remote location for every manufactur-
ing site, resources can be distributed more efficiently.
Furthermore, production facilities only need to install
the sensor technology for monitoring and data collec-
tion purposes as the data processing is conducted re-
12
motely. Hence, connecting multiple locations to a cen-
tralized predictive maintenance system has the potential
to create economies of scale. Nevertheless, since pre-
dictive maintenance is not simply scalable for any sys-
tem size [4], the centralized system needs to take into
account structural, stochastic and especially economic
dependencies. Hence, with regard to an implementation
of predictive maintenance in the context of Industry 4.0
remote data storage and access should come to the fore
of further research.
The third aspect which might point out a gap in the
existing research based on the sample of papers ana-
lyzed in the present survey is related to the category of
data handling as well (see subsection 4.9). Hardly over
20 papers address the topic of data fusion and data fil-
tering. This might be the case since a lot of research
focuses on single-component systems. However, with
an increasing number of sensors in use the amount of
data increases and makes methods such as data fusion
and data filtering necessary in order to process it effi-
ciently. Especially in an Industry 4.0 environment with
multiple integrated systems, big date becomes an issue.
Predictive maintenance systems not only generate an in-
creasing amount of data because of multiple machines
or even multiple manufacturing locations but addition-
ally due to improvements in sensor technology which
will allow to monitor even more parameters then be-
fore. Wireless sensors already enable the collection of
monitoring data where it was not possible to collect data
before due to fragile wires. The improvements in sen-
sor technology will support predictive maintenance sys-
tems in becoming more accurate. Nevertheless, in order
to keep predictive maintenance efficient with regard to
processing power usage and big data, further research is
needed in the area of feasible methods of data filtering
and data fusion.
Finally, no paper in the present survey covered the
topic of possibilities for small and medium sized com-
panies. The research shows opportunities of predictive
maintenance such as cost minimization, increased pro-
ductivity and prolonged machine life. However, none
of the papers in the present survey address the issue of
how firms can implement predictive maintenance in a
cost minimizing way. Thus, while the latest sensor and
data processing technology might be affordable for large
companies due to economies of scale, since predictive
maintenance can process data of multiple sensors, small
and medium sized companies are unlikely to implement
predictive maintenance due to high upfront costs and
complexity [28]. Nonetheless, the majority of firms are
small and medium sized companies. The researchers
show what is possible and do not focus on what is prac-
tical and can be implemented on a broad scale in order
to boost an economies productivity and use of resources.
Therefore, further research should focus on developing
predictive maintenance systems that are simple and af-
fordable to implement even for smaller companies.
To sum it up, the present survey reveals research gaps
in the context of Industry 4.0. The research often fo-
cuses on too narrow aspects instead of more broader
concepts as it would be necessary in an Industry 4.0 en-
vironment. The connection of predictive maintenance
to other processes such as spare part logistics should
be considered in further research. Hence, predictive
maintenance should be addressed as part of a larger in-
tegrated system. Furthermore, possibilities for a large
scale implementation are an issue as the literature shows
what is possible, but only few companies have yet suc-
cessfully implemented predictive maintenance [15].
6. Conclusion
The present survey represents a comprehensive anal-
ysis of the topic of predictive maintenance. Predic-
tive maintenance distinguishes itself from conventional
maintenance policies by the attempt to not solely de-
tect an anomaly of a machine or component, but to pre-
dict when the failure might occur in order to efficiently
schedule maintenance actions in advance. Researchers
are severely interested in the concept of predictive main-
tenance since hundreds of papers with regard to the
topic are published every year. Minimized downtime,
prolonged machine life, increased productivity and re-
duced costs are merely a few promising prospects of
predictive maintenance. The objective of the present
survey is to detect and categorize a variety of aspects
with regard to the comprehensive topic of predictive
maintenance. For the purpose of the survey 150 papers
are analyzed and categorized. The result is a data grid
with 69 attributes which are clustered into the follow-
ing 10 categories: Goals, Condition Monitoring, Main-
tenance Scope, Degradation Process, Fault Detection,
System Size, Scheduling, Prognostic Techniques, Data
Handling and Evaluation. These 10 categories together
with their corresponding attributes built a framework for
predictive maintenance.
6.1. Reflection of the Procedure
For the present survey a structured literature review
(SLR) is conducted. This approach is beneficial in con-
trast to the berry picking method as it captures all facets
of the topic. Since the objective of the present survey
is to categorize and structure a comprehensive scope of
13
the research field of predictive maintenance, the SLR is
the more appropriate choice. In order to construct the
framework for predictive maintenance, the 150 papers
are analyzed and categorized which results in 10 cate-
gories and 69 attributes each belonging to one category.
Although the number of papers in the research field of
predictive maintenance ranges into the thousands, the
number of papers in this survey represents an informa-
tive cross subsection of the topic. Hence, the frame-
work provides a comprehensive overview about predic-
tive maintenance. The concept of deducing attributes,
grouping them into categories and building a framework
related to a specific topic is very beneficial in order to
get a structured and deeper understanding of a subject
and to reveal potential gaps in the existing literature. Es-
pecially when analyzing a considerable number of sci-
entific publications.
Considering that there haven been thousands of pa-
pers published over decades with regard to predictive
maintenance raised the question in which way the exist-
ing research can be classified in order to structure it and
to reveal potential gaps for further research. The frame-
work structures and classifies papers published between
1993 and 2018. The framework makes no claim of com-
pleteness, neither are the attributes within each category.
Nevertheless, the framework of the present survey cov-
ers relevant aspects and facets of the topic of predictive
maintenance and builds a comprehensive introduction
into the research field of predictive maintenance.
The framework of the present survey shows that fur-
ther research might be appropriate in certain directions
of the topic. First, the effect of different types of ma-
chine and component dependencies in the context of
multi-component systems. Second, remote data storage
and access using IoT technology for remote predictive
maintenance approaches. Third, the relevance of data
fusion and data filtering due to big data issues as a result
of an increasing number of monitoring sensors. Finally,
the possibilities for a less complex and more affordable
implementation of predictive maintenance systems even
for small and medium sized companies. Note that these
research gaps are presumed based on the result of the
present survey which does not cover all the existing aca-
demic literature.
References
[1] M. Y. You, F. Liu, W. Wang, G. Meng, Statistically planned
and individually improved predictive maintenance management
for continuously monitored degrading systems, IEEE Transac-
tions on Reliability 59 (4) (2010) 744–753. doi:10.1109/TR.
2010.2085572.
[2] R. K. Mobley, An introduction to predictive maintenanc (2002)
1–6.
[3] A. Grall, L. Dieulle, C. Berenguer, M. Roussignol, Continuous-
time predictive-maintenance scheduling for a deteriorating sys-
tem, IEEE Transactions on Reliability 51 (2) (2002) 141–
150. arXiv:arXiv:1011.1669v3,doi:10.1109/TR.2002.
1011518.
[4] K. A. Nguyen, P. Do, A. Grall, Multi-level predictive main-
tenance for multi-component systems, Reliability Engineering
and System Safety 144 (2015) 83–94. doi:10.1016/j.ress.
2015.07.017.
[5] R. C. Yam, P. W. Tse, L. Li, P. Tu, Intelligent predictive decision
support system for condition-based maintenance, International
Journal of Advanced Manufacturing Technology 17 (5) (2001)
383–391. doi:10.1007/s001700170173.
[6] A. K. Jardine, D. Lin, D. Banjevic, A review on ma-
chinery diagnostics and prognostics implementing condition-
based maintenance, Mechanical Systems and Signal Processing
20 (7) (2006) 1483–1510. arXiv:0208024,doi:10.1016/j.
ymssp.2005.09.012.
[7] B. Schmidt, L. Wang, Cloud-enhanced predictive mainte-
nance, The International Journal of Advanced Manufactur-
ing Technology 99 (1-4) (2018) 5–13. doi:10.1007/
s00170-016- 8983-8.
[8] X. Zhou, L. Xi, J. Lee, Reliability-centered predictive main-
tenance scheduling for a continuously monitored system sub-
ject to degradation, Reliability Engineering and System Safety
92 (4) (2007) 530–534. doi:10.1016/j.ress.2006.01.
006.
[9] J. Lee, J. Ni, D. Djurdjanovic, H. Qiu, H. Liao, Intelligent
prognostics tools and e-maintenance, Computers in Industry
57 (6) (2006) 476–489. arXiv:arXiv:1011.1669v3,doi:
10.1016/j.compind.2006.02.014.
[10] A. Muller, A. Crespo Marquez, B. Iung, On the concept of e-
maintenance: Review and current research, Reliability Engi-
neering and System Safety 93 (8) (2008) 1165–1187. doi:
10.1016/j.ress.2007.08.006.
[11] J. Yan, Y. Meng, L. Lu, L. Li, Industrial Big Data in an Industry
4.0 Environment: Challenges, Schemes, and Applications for
Predictive Maintenance, IEEE Access 5 (2017) 23484–23491.
doi:10.1109/ACCESS.2017.2765544.
[12] A. Kofod-Petersen, How to do a structured literature review in
computer science, Researchgate (2014) 1–7.
[13] A. Booth, Using evidence in practice. unpacking your literature
search toolbox: on search styles and tactics, Health Information
and Libraries Journal 25 (2008) 313–317.
[14] H. De Faria, J. G. S. Costa, J. L. M. Olivas, A review of monitor-
ing methods for predictive maintenance of electric power trans-
formers based on dissolved gas analysis, Renewable and Sus-
tainable Energy Reviews 46 (2015) 201–209. doi:10.1016/
j.rser.2015.02.052.
[15] K. Efthymiou, N. Papakostas, D. Mourtzis, G. Chryssolouris,
On a predictive maintenance platform for production systems,
in: 45th CIRP Conference on Manufacturing Systems, Vol. 3,
2012, pp. 221–226. doi:10.1016/j.procir.2012.07.039.
[16] D. J. Edwards, G. D. Holt, F. C. Harris, Predictive maintenance
techniques and their relevance to construction plant, Journal of
Quality in Maintenance Engineering 4 (1) (1998) 25–37. doi:
10.1108/13552519810369057.
[17] A. Heng, S. Zhang, A. C. Tan, J. Mathew, Rotating machinery
prognostics: State of the art, challenges and opportunities, Me-
chanical Systems and Signal Processing 23 (3) (2009) 724–739.
arXiv:0208024,doi:10.1016/j.ymssp.2008.06.009.
[18] S. T. Kandukuri, A. Klausen, H. R. Karimi, K. G. Robbersmyr,
A review of diagnostics and prognostics of low-speed machinery
14
towards wind turbine farm-level health management, Renewable
and Sustainable Energy Reviews 53 (2016) 697–708. doi:10.
1016/j.rser.2015.08.061.
[19] B. Lu, D. B. Durocher, P. Stemper, Predictive maintenance tech-
niques, IEEE Industry Applications Magazine 15 (6) (2009) 52–
60. doi:10.1109/MIAS.2009.934444.
[20] S. Munirathinam, B. Ramadoss, Big data predictive analtyics
for proactive semiconductor equipment maintenance, in: 2014
IEEE International Conference on Big Data, 2014, pp. 893–902.
doi:10.1109/BigData.2014.7004320.
[21] S.-j. Wu, N. Gebraeel, M. A. Lawley, Y. Yih, A Neural Net-
work Integrated Decision Support System for Condition-Based
Optimal Predictive Maintenance Policy, IEEE Transactions on
Systems, Man, and Cybernetics - Part A: Systems and Humans
37 (2) (2007) 226–236. doi:10.1109/TSMCA.2006.886368.
[22] R. Baidya, S. K. Ghosh, Model for a predictive maintenance sys-
tem effectiveness using the analytical hierarchy process as ana-
lytical tool, in: IFAC-PapersOnLine, Vol. 28, 2015, pp. 1463–
1468. doi:10.1016/j.ifacol.2015.06.293.
[23] E. Deloux, B. Castanier, C. Berenguer, Predictive maintenance
policy for a gradually deteriorating system subject to stress, Re-
liability Engineering and System Safety 94 (2) (2009) 418–431.
doi:10.1016/j.ress.2008.04.002.
[24] K. A. Kaiser, N. Z. Gebraeel, Predictive maintenance manage-
ment using sensor-based degradation models, IEEE Transac-
tions on Systems, Man, and Cybernetics Part A:Systems and
Humans 39 (4) (2009) 840–849. doi:10.1109/TSMCA.2009.
2016429.
[25] Z. Yang, D. Djurdjanovic, J. Ni, Maintenance scheduling in
manufacturing systems based on predicted machine degrada-
tion, Journal of Intelligent Manufacturing 19 (1) (2008) 87–98.
doi:10.1007/s10845-007- 0047-3.
[26] A. Raza, V. Ulansky, Modelling of Predictive Maintenance for a
Periodically Inspected System, in: The 5th International Confer-
ence on Through-life Engineering Services, Vol. 59, 2017, pp.
95–101. doi:10.1016/j.procir.2016.09.032.
[27] C. Okoh, R. Roy, J. Mehnen, Predictive Maintenance Mod-
elling for Through-Life Engineering Services, in: The 5th In-
ternational Conference on Through-life Engineering Services,
Vol. 59, The Author(s), 2017, pp. 196–201. doi:10.1016/j.
procir.2016.09.033.
[28] M. C. Carnero, An evaluation system of the setting up of pre-
dictive maintenance programmes, Reliability Engineering and
System Safety 91 (8) (2006) 945–963. doi:10.1016/j.ress.
2005.09.003.
[29] N. Z. Gebraeel, M. A. Lawley, R. Li, J. K. Ryan, Residual-life
distributions from component degradation signals: A Bayesian
approach, IIE Transactions (Institute of Industrial Engineers)
37 (6) (2005) 543–557. arXiv:arXiv:1011.1669v3,doi:
10.1080/07408170590929018.
[30] S. Orhan, N. Akturk, V. Celik, Vibration monitoring for defect
diagnosis of rolling element bearings as a predictive mainte-
nance tool: Comprehensive case studies, NDT&E International
39 (4) (2006) 293–298. doi:10.1016/j.ndteint.2005.08.
008.
[31] M. Traore, A. Chammas, E. Duviella, Supervision and prog-
nosis architecture based on dynamical classification method for
the predictive maintenance of dynamical evolving systems, Re-
liability Engineering and System Safety 136 (2015) 120–131.
doi:10.1016/j.ress.2014.12.005.
[32] J. Lindstrom, H. Larsson, M. Jonsson, E. Lejon, Towards Intel-
ligent and Sustainable Production: Combining and Integrating
Online Predictive Maintenance and Continuous Quality Con-
trol, in: The 50th CIRP Conference on Manufacturing Systems,
Vol. 63, The Author(s), 2017, pp. 443–448. doi:10.1016/j.
procir.2017.03.099.
[33] L. Dieulle, C. Berenguer, A. Grall, M. Roussignol, Continuous
time predictive maintenance scheduling for a deteriorating sys-
tem, Reliability and Maintainability Symposium Proceedings
(2001) 150–155doi:10.1109/RAMS.2001.902458.
[34] C. M. Tan, N. Raghavan, Imperfect predictive maintenance
model for multi-state systems with multiple failure modes and
element failure dependency, in: 2010 Prognostics and Sys-
tem Health Management Conference, PHM ’10, 2010. doi:
10.1109/PHM.2010.5414594.
[35] A. Ladj, C. Varnier, F. B. S. Tayeb, IPro-GA: an integrated prog-
nostic based GA for scheduling jobs and predictive maintenance
in a single multifunctional machine, in: IFAC-PapersOnLine,
Vol. 49, 2016, pp. 1821–1826. doi:10.1016/j.ifacol.
2016.07.847.
[36] K. A. Nguyen, P. Do, A. Grall, Joint predictive maintenance
and inventory strategy for multi-component systems using Birn-
baum’s structural importance, Reliability Engineering and Sys-
tem Safety 168 (May) (2017) 249–261. doi:10.1016/j.
ress.2017.05.034.
[37] N. Gebraeel, Sensory-updated residual life distributions for
components with exponential degradation patterns, IEEE Trans-
actions on Automation Science and Engineering 3 (4) (2006)
382–393. doi:10.1109/TASE.2006.876609.
[38] A. H. Elwany, N. Z. Gebraeel, Sensor-driven prognostic models
for equipment replacement and spare parts inventory, IIE Trans-
actions (Institute of Industrial Engineers) 40 (7) (2008) 629–
639. doi:10.1080/07408170701730818.
[39] H. M. Hashemian, W. C. Bean, State-of-the-art predictive main-
tenance techniques, IEEE Transactions on Instrumentation and
Measurement 60 (10) (2011) 3480–3492. doi:10.1109/TIM.
2009.2036347.
[40] H. M. Hashemian, Wireless sensors for predictive maintenance
of rotating equipment in research reactors, Annals of Nuclear
Energy 38 (2-3) (2011) 665–680. doi:10.1016/j.anucene.
2010.09.012.
[41] A. Van Horenbeek, L. Pintelon, A dynamic predictive mainte-
nance policy for complex multi-component systems, Reliabil-
ity Engineering and System Safety 120 (2013) 39–50. doi:
10.1016/j.ress.2013.02.029.
[42] Y. Yuan, X. Jiang, X. Liu, Predictive maintenance of shield tun-
nels, Tunnelling and Underground Space Technology 38 (2013)
69–86. doi:10.1016/j.tust.2013.05.004.
[43] C.-H. H. D.-H. Z. Xiao-Sheng Si, Wenbin Wang, Remaining
useful life estimation - a review on the statistical data driven
approaches, European Journal of Operational Research 213 (1)
(2011) 1–14.
[44] R. B. S. Agatonovic-Kustrin, Basic concepts of artificial neural
network (ann) modeling and its application in pharmaceutical
research, Journal of Pharmaceutical and Biomedical Analysis
22 (5) (2000) 717–727.
[45] K. S. R. K. P. Sudheer, A. K. Gosain, A data-driven algorithm
for constructing artificial neural network rainfall-runoffmodels,
Hydrological Processes 16 (6) (2002) 1325–1330.
[46] Y. Yamato, Y. Fukumoto, H. Kumazaki, Predictive Mainte-
nance Platform with Sound Stream Analysis in Edges, Journal
of Information Processing 25 (2017) 317–320. doi:10.2197/
ipsjjip.25.317.
[47] K. Wang, Intelligent Predictive Maintenance (IPdM) system, In-
dustry 4.0 scenario, WIT Transactions on Engineering Sciences
113 (2016) 259 – 268. doi:10.2495/IWAMA150301.
[48] A. Schirru, S. Pampuri, G. De Nicolao, Particle filtering of
hidden gamma processes for robust Predictive Maintenance
in semiconductor manufacturing, in: 2010 IEEE Interna-
tional Conference on Automation Science and Engineering,
15
CASE 2010, 2010, pp. 51–56. doi:10.1109/COASE.2010.
5584518.
[49] Y. C. Chiu, F. T. Cheng, H. C. Huang, Developing a factory-wide
intelligent predictive maintenance system based on Industry 4.0,
Journal of the Chinese Institute of Engineers 40 (7) (2017) 562–
571. doi:10.1080/02533839.2017.1362357.
[50] B. De Saporta, F. Dufour, H. Zhang, C. Elegbede, Opti-
mal stopping for the predictive maintenance of a structure
subject to corrosion, Proceedings of the Institution of Me-
chanical Engineers, Part O: Journal of Risk and Reliability
226 (2) (2012) 169–181. arXiv:arXiv:1101.1740v2,doi:
10.1177/1748006X11413681.
[51] X. Lei, P. A. Sandborn, N. Goudarzi, M. A. Bruck, PHM Based
Predictive Maintenance Option Model for Offshore Wind Farm
O&M Optimization, in: Proceedings of the Annual Conference
of the Prognostics and Health Management Society 2015, 2015,
pp. 1–10.
[52] B. Schmidt, L. Wang, D. Galar, Semantic Framework for Pre-
dictive Maintenance in a Cloud Environment, in: 10th CIRP
Conference on Intelligent Computation in Manufacturing En-
gineering - CIRP ICME ’16, Vol. 62, 2017, pp. 583–588. doi:
10.1016/j.procir.2016.06.047.
[53] S. Yang, An experiment of state estimation for predictive main-
tenance using Kalman filter on a DC motor, Reliability Engi-
neering & System Safety 75 (1) (2002) 103–111. doi:10.
1016/S0951-8320(01)00107- 7.
16