You are on page 1of 25

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/331951459

Challenges and Reliability of Predictive Maintenance

Preprint · March 2019


DOI: 10.13140/RG.2.2.35379.89129

CITATION READS

1 9,920

1 author:

Vincent Meyer Zu Wickern


Hochschule Rhein-Waal
1 PUBLICATION   1 CITATION   

SEE PROFILE

All content following this page was uploaded by Vincent Meyer Zu Wickern on 22 March 2019.

The user has requested enhancement of the downloaded file.


RHEIN-WAAL UNIVERSITY OF APPLIED SCIENCES

Challenges and Reliability of Predictive


Maintenance

by
Vincent F. A. Meyer zu Wickern

in the
Faculty of Communication and Environment

March 2019
Declaration of Authorship

I, Vincent F. A. Meyer zu Wickern, declare that this thesis titled, ‘Challenges and Reliability of
Predictive Maintenance’ and the work presented in it are my own. I confirm that:

 This work was done wholly or mainly while in candidature for a research degree at this University.

 Where any part of this paper has previously been submitted for a degree or any other qualification
at this University or any other institution, this has been clearly stated.

 Where I have consulted the published work of others, this is always clearly attributed.

 Where I have quoted from the work of others, the source is always given. With the exception of
such quotations, this paper is entirely my own work.

 I have acknowledged all main sources of help.

 Where the thesis is based on work done by myself jointly with others, I have made clear exactly
what was done by others and what I have contributed myself.

Signed:

Date:

i
RHEIN-WAAL UNIVERSITY OF APPLIED SCIENCES

Abstract
Faculty of Communication and Environment

by Vincent F. A. Meyer zu Wickern

Predictive maintenance (PdM) is a concept, which is applied to optimize asset maintenance plans
through the prediction of asset failures with data driven techniques. The application of PdM may
be beneficial to companies through the reduction of downtimes and the increase of product quality
among other favorable effects, especially in manufacturing, which is the area of focus in this paper.
Nevertheless, the PdM methods face multiple challenges, which may limit or impede its use. It is
important to gather these challenges to be able to draw a statement on the reliabilty of the PdM
methods, which is aimed in this paper. Hence, the concept of PdM together with its benefits and
functionality is presented and challenges are discussed to form a current view on the reliability and
points of focus in predictive maintenance.

It is found that there are various factors impacting the reliability and applicability of PdM methods
in companies. While certain organizational and economic prerequisites are essential for the use of
PdM methods, technological challenges, which have yet to be resolved, limit the effectiveness of PdM
methods. In consequene, opposed to stating an universal decision on the reliability of PdM methods,
this paper evaluates PdM method’s reliability and success to be company and purpose dependent.
The influences and limitations of PdM methods consolidated in this paper may support to analyze
a specific company case. In future, the factors explained in this paper may be reviewed in PdM-
practicing companies to validate them in conjunction. High attention from the scientific community
towards the PdM challenges leads to an expectation of solutions for these challenges in the future.
Contents

Declaration of Authorship i

Abstract ii

Abbreviations iv

1 Introduction 1

2 Concept of Predictive Maintenance 2

3 Benefits of Predictive Maintenance 3

4 Data Sources 4

5 Predictive Maintenance Modelling 6

6 Challenges of Predictive Maintenance 8


6.1 Financial and organizational obstacles . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
6.2 Data source limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
6.3 Limits of machine repair activities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
6.4 Optimization narrowness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
6.5 Sentiment towards challenges and outlook . . . . . . . . . . . . . . . . . . . . . . . . . 12

7 Results and Discussion 13


7.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
7.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

8 Conclusion 15

Bibliography 16

iii
Abbreviations

ANN Artificial Neural Network

AUC Area under the curve

CBM Condition Based Monitoring

CRM Customer Relationship Management

ERP Enterprise Resource Planning

GMM Gaussian Mixture Model

HR Human Resources

ICS Industrial Control System

KPI Key Performance Indicator

MIL Multi-Instance Learning

PdM Predictive Maintenance

PHM Prognostic Health Management

PM Preventative Maintenance

PM-AUC Predictive Maintenance-based area under precision-recall

PLC Programmable Logic Controller

OEE Overall Equipment Effectiveness

OEM Original Equipment Manufacturer

RCM Remote Condition Monitoring


iv
Abbreviations v

RM Reactive Maintenance

ROI Return on Investment

RUL Residual Useful Life

SCADA Supervisory Control and Data Acquisition

SMEs Small and medium-sized enterprises

SOM Self-Organizing Map

SVM Support Vector Machine


Chapter 1

Introduction

Many companies face the problem of asset degradation, which is critical especially if this degradation
leads to a shutdown or unavailability of essential assets [1]. Therefore a concept called Preventative
Maintenance (PM) arose, which focuses on the prevention of asset failures and their surrounding
environments through the introduction of maintenance cycles. The concept of PM has developed into
Predictive Maintenance (PdM), which aims to forecast failures of assets by analyzing asset related
data to enable the choice of an optimal point in time for maintenance.

PdM methods are especially used in IoT devices [2]. Peon found PdM to be already applied to a high
extent in industry [3]. This paper focuses on the application area of manufacturing, as it is perceived as
a major application area due to its influence on all producing companies using continually degrading
assets. Despite its already frequent use, PdM is still a highly researched field with many papers
published in the last years, e.g. [4] [5] [6]. However, the question remains whether the PdM methods
are reliable and if they benefit the companies using them. Another question is, which challenges to
PdM methods may influence their reliability and if there are limitations. This paper will strive to
answer these questions and pursue a statement on the reliability of PdM methods in manufacturing.
This statement and the corresponding influences may support companies in their decision about an
introduction and application of PdM methods. Moreover, it may provide a foundation for the further
direction of scientific research in PdM, e.g. by collating research and solution gaps.

To be able to derive a statement about the current reliability of PdM methods, the concept, goals,
functionality, and, in particular, challenges of PdM have to be considered. Therefore, all these points
are researched in a current literature review and are presented in this paper. Opposed to the focus
on one specific challenge, as is the aim of many papers on PdM, e.g. [7] [2] [8], or the presentation
of PdM as part of a greater field, such as smart manufacturing or digital maintenance, e.g. in [6] [9],
this paper connects the PdM specific ideas and presents them in conjunction.

1
Chapter 2

Concept of Predictive Maintenance

In manufacturing companies, different maintenance strategies are used: Reactive Maintenance (RM),
PM and PdM [6]. The most traditional of these strategies is RM, which begins with a correction
of a failure after this failure has taken place. In contrast to this, PM aims to prepone the time of
preventative measures before the time of a potential asset failure.

An important shortcoming in PM strategies is that the current condition of an asset does not influ-
ence its maintenance schedule [10]. Condition Based Monitoring (CBM) as one implementation of
PdM, in contrast, includes measurements of the condition of the assets into its maintenance planning.
Frequently recorded characteristics are for example the temperature or vibration in a machine.

CBM related activities can be divided into three groups of steps: Data acquisition, data preprocessing
and definition of maintenance decisions [11]. In the first step, data acquisition, data is recorded and
collected, e.g. from sesnsors. In contrast, preventative maintenance decisions on schedules are based
on experience and intuition of the involved people as well as alerting systems, spreadsheets, operator
logs, and shift transfer discussions. Data preprocessing, the second step in the work of a CBM, adjusts
and interprets the data gathered in the first step, e.g. with noise reduction [11].

With the returned information from data preprocessing, maintenance policies and decisions can be
derived. Both diagnostics and prognostics are applied in this step. With diagnostics, prior failures can
be singled out and measured. In the course of diagnostics, failures have to be recorded, distinguished
and identified [11]. Prognostics are executed to predict failures, which may be foreseen in the future.
With prognostics, the Residual Useful Life (RUL), i.e. the remaining time before an asset runs into
a failure [12], and the confidence interval can be estimated. The introduced asset condition is also
referred to as the degradation signal of an asset and is the essential indicator and calculation element
for predicting the RUL [13]. The invention of sensors and other means of information recording in
a production environment led to more attention to Prognostic Health Management (PHM), a term
comprising prognostic methods to measure and use the asset health statuses such as PdM methods.
2
Chapter 3

Benefits of Predictive Maintenance

One important Key Performance Indicator (KPI) for many producing companies is the Overall Equip-
ment Effectiveness (OEE) [14]. The OEE indicates the level of availability and performance of produc-
tion assets and their output quality. By including downtime into the OEE measurement, unplanned
outages of assets result in a negative impact on its value. In its aim to reduce unplanned outages, PdM
may have a positive effect on the OEE. Other metrics affected by the application of PdM include the
reduction of scrap and enhanced output quality [6].

Product quality does not only depend on the degradation status of production machines but among
others also on the components of a production machine [15] [12]. The planning of the necessity on
the amount of repair parts at a certain time is difficult for companies and a higher inventory with
many spare parts leads to higher inventory costs. Therefore it may be reasonable to employ PdM
methods for planning repair parts inventory levels and keep at least as much inventory as is predicted
necessary. In practice, PdM methods often lack a joint optimization of inventory, however, models
have been created, how stock management may be included in PdM optimization, e.g. by Soltani [12].

Another benefit of PdM may be achieved in the field of remote sensing. CBM as a part of PdM may
be of extended benefit under conditions, under which it is installed with remote sensing technologies
allowing for Remote Condition Monitoring (RCM) and therewith enabling the review of conditions in
an asset, where regular maintenance procedures would not be possible or safe [16], e.g. measurements
of oil temperatures in a running engine.

Another possibility, partly enabled by PdM, is the servitization of manufacturing as a new business
model [17]. In this business model, an Original Equipment Manufacturer (OEM) retains an enduring
relationship with the users of their products performing maintenance as a service to them. With
an application of PdM methods, an OEM can continuously collect data about the use, failures, and
degradation of their products to consider these aspects for product improvement [18].

3
Chapter 4

Data Sources

An inevitable necessity for the application of predictive maintenance is data, from which to analyze
the past of asset usage and extrapolate to the future. Data may come from different sources, some of
which are explained in the following.

An essential part of the necessary data is failure data that indicates when and under which circum-
stances a system broke down [11]. For extracting this information, fault detection is used comprising
the techniques of fault isolation and fault identification. With fault isolation, a failure point is distin-
guished and separated in the data set. With fault identification, the failure’s describing characteristics
and extent are derived. Fault detection generates one of two central PdM data categories called event
data, which stores information of failures and of corresponding corrective actions. The other central
component, which is extracted with PdM data acquisition, is condition monitoring data, in which
information is gathered that could bear useful indicators on the reasons for a failure.

To gain information about current circumstances of an active asset, sensors may be used, which deliver
data in form of signals about measurable attributes of the asset, e.g. temperature and voltage [1].
Sensors can send signals in real-time and allow a direct reaction on their values. In extreme cases,
sensors can send alerts, if a critical value is measured. Despite these advantages, sensors have the
drawback that it is difficult to install them into assets, which are currently in use, since they require
laborious hardware changes and consequently long downtimes. The overall costs and efforts installing
them represent another challenge, especially for cost-sensitive environments. In some cases, their use
may further be restricted by regulations.

As a source of real-time information, sensors fall into the group of streaming, together with for example
satellites, weather measurements and Programmable Logic Controller (PLC) units [1]. PLCs are digital
peripheral devices, which are applied to lead and automate electromechanical processes and assets,
e.g. the steering of production machinery. PLCs are commonly used together with other PLCs and
logic controllers, which are all controlled by a Supervisory Control and Data Acquisition (SCADA)
4
Introduction 5

system [19]. SCADA systems are applied to control distributed systems, which may stretch over a
wide area. Together with other control systems, SCADA and PLCs can be subdivided under the term
Industrial Control System (ICS), which can be directly used as a data source for PdM.

If the installation of sensors is impeded by the challenge to take assets off production or other reasons,
the use of logs represents one alternative [20]. Frequently, the created output of logs has been designed
by programmers to present warnings, error messages, exceptions, condition information and other
pieces of information. This data may as well be relevant to reproduce the condition of assets or
the time and cause of a failure. However, since logs are seldom built with the intent of using it for
predictive maintenance, some challenges are faced that result from this diverted use. Logs frequently
do not indicate failures and do not follow a machine-readable structure, but unstructured text or
symbols. Moreover, much of the log’s output is not useful for PdM, so that it first has to be cut
down to the relevant part. Especially service information about repair activities with descriptions of
incidents, causes, and corrective activities may be useful.

Similar to ICS or logs, Business Information systems, such as Customer Relationship Management
(CRM), Enterprise Resource Planning (ERP), Human Resources (HR), financial systems, and other
information systems may be diverted from their intended use and additionally be exploited for data
[1]. This data may as well be integrated into a data warehouse or common relational databases.
Spreadsheets are often used in a company environment and also may bear useful information to PdM.
Chapter 5

Predictive Maintenance Modelling

If analyzable asset data, as described in the section Data Sources, is available, a statistical model for
PdM, which is necessary for the prediction of machine failures, can be created. Models may comprise
techniques from Data Mining, Machine Learning, and Simulation [1]. Data Mining supports in the
study of large amounts of data to detect patterns and relations. Machine Learning then uses this data
and patterns to learn how the variables influence each other. Simulation can finally present another
outcome in the same model given different attribute values in the data. All these techniques typically
model and work with the degradation of assets in the PdM environment, which is the dominant
indicator of RUL and hence the optimal maintenance time points. A measure of degradation can be
included in a prediction model that can predict a machine’s condition. This prediction model may
then be evaluated against to the model goals. This shortly outlined concept is explained further in
this section. However, the described techniques are not claimed to be complete.

Similar or equal assets may have similar degradation developments leading to the RUL depending on
time and conditions that for example show a linear or exponential deterioration [10]. These devel-
opments can be categorized into proportional hazard models and degradation models. Proportional
hazard models are explained by external running conditions that are fed into a benchmark hazard
function. Major application areas are engineering uses, e.g. aircraft and mechanical apparatus. Degra-
dation models, on the other hand, are fed with condition-based sensor data, which are also known as
degradation signals and which are directly sent by assets, in the deteriorated version. A limitation to
degradation modeling is that not all device technologies may be described by a proportional hazard
model or a degradation model calculating the RUL [21]. In practice, it is not always possible to model
an asset’s degradation with existing models.

Each PdM model has to be evaluated for its ability to predict asset failures and to not deliver false
alarms predicting failures that would not occur [2]. The most common measures for model quality are
precision, recall, and the F1 score. The precision can be calculated by dividing the true positives of the

6
Introduction 7

predictions by the sum of true positives and false positives [20]. In this way, precision calculates how
many predictions, which have been proposed, were correct. The recall is the result of a division of true
positives by all failures. Hence, the recall figure states how many of the failures have been detected.
The F1-score is the harmonic mean across the precision and recall, so it calculates the balance between
these figures [2].

PdM models are dependent on the underlying maintenance strategy, which is dependent on multiple
influences, such as repair costs and failure consequences [20]. Taking into account these influences,
precision and recall are strived to be optimized in balance. The quality of balance between precision and
recall can be calculated with a Predictive Maintenance-based area under precision-recall (PM-AUC),
which is calculated like a regular Area under the curve (AUC), but with the described specifications
of precision and recall.

The set of data mining algorithms applied in PdM is diverse. Anomaly detection may be used to find
data points or structures, which are very different from the rest of the data [1]. This may especially be
helpful in the failure isolation. Failure isolation may also be aided by Multi-Instance Learning (MIL),
which is a technique, in which the data set to be analyzed consists of bags of data that are either
classified as positive or negative [20]. Particularly in log analysis, analyzing these bags may support to
analyze many daily logs or weekly logs that are labeled negative, if no failures occurred, and positive,
if at least one failure occurred.

After failure detection, other techniques are used to link these failures to conditions in the asset [1].
The identifications of connections and ordering in the data are referred to as association rules, which
help to detect causality in the data, e.g. with association rules it may be shown that temperatures
higher than a certain threshold and time after a certain hour per day lead to a higher risk of running
into a failure on the same day. Regression also presents a model based relation between multiple
variables resulting in a numerical value of the depending variable, commonly being the RUL. Another
applied field of techniques is classification, in which a group of attribute values is categorized into a
class that may provide further information by finding similarities in RULs of the same class.

One technique within the group of applied classifiers is the random forest model. Random forest models
are beneficial in PdM since they are based on a little number of variables and therefore need relatively
less training data than other classifiers [2]. Random Forests are able to cope with unbalanced data
sets, are simple to understand and enable feature ranking. Imbalance in the data refers to the fact that
most often only a few failures can be identified, while there are many records containing conditions
[20]. Other classifiers used for PdM may also implementations of a Support Vector Machine (SVM) or
an Artificial Neural Network (ANN). Apart from this, unsupervised learning classication, referred to
as clusting, can be used, to enable a machine to find groups in the data set on its own [17]. Clustering
models in use are for example the Self-Organizing Map (SOM) or the Gaussian Mixture Model (GMM).
Chapter 6

Challenges of Predictive Maintenance

In this chapter, challenges and limitations to predictive maintenance are explained. In the list of
challenges found in scientific literature, four groups can be made out, which will be taken into account:
Financial and organizational obstacles, data source limitations, limits of machine repair activities and
optimization narrowness. In consequence, the current sentiment towards the challenges and an outlook
is given.

6.1 Financial and organizational obstacles

Profit targeting companies inevitably consider expected costs on any new investment. PdM efforts,
such as the installation of sensors, extraction of information, preparation and maintenance of models
and maintenance activities, generate costs for companies, in which PdM methods are introduced.
These costs may vary according to multiple factors, e.g. the type and complexity of assets and
corresponding sensors, cost of consulting, installation and knowledge extraction according to whether
the necessary skills can be sourced from in-house or external people. A method to evaluate, whether
a PdM introduction may be of benefit, is the educated creation of a projected Return on Investment
(ROI) [1]. The projection of the ROI has to consider the value of PdM results, the amortization time
and the described costs. The financial reasoning of PdM usage and applicability is further dependent on
the size and type of company, in which it is introduced [6]. While small and medium-sized company are
generally more limited in their technological apparatus, larger companies may also be less endangered
by the financial risks of the PdM investment. Technology providers usually adapt to the choice of
their customers and the market need.

Another source of cost may arise through an extended effort of extracting insights from PdM data.
While continuous monitoring with current information on asset conditions is available, the produced
information and visualizations are difficult to understand with current PdM solutions [11]. Directly

8
Introduction 9

accessing the model output is complex and as stated by Efthymiou et al, the created visualizations
are held too simple and miss an insightful as well as a user-friendly view. Bokrantz et al propose that
a PdM system should be easily understandable and automated in the future, particularly considering
that the complexities of models increase further with big data [5].

As insights may be difficult to extract, knowledge management is one factor to make the gathered
knowledge available to an extended group of people and retain this knowledge. Especially the area of
fault detection demands a high amount of prior technical and domain knowledge as well as specific
training [11]. A way to distribute fault detection capabilities is the introduction of fault detection
automation. In practice, it is stated that failure information from the past is difficult to isolate.
Failure detection clarity regularly decreases with the amount of time dated back. Taking rule-based
approaches in fault detection, the number of combinations exponentially grows with the count of
characteristics and may be too large to process. In the research of Jin et al, technology providers
reported that missing failure occasions in the given data present an obstacle in the creation of PdM
models [6]. This may lead to incorrect alarms in the derived model, as conditions during a failure
status may be unclear.

Knowledge Management does not only constitute a technological possibility, but also an organizational
method of collaboration. The organization, in which prognostic measures are introduced has been
recognized as an important factor for the benefits, which are achieved with the use of PdM methods
[6] [8]. The employees in Small and medium-sized enterprises (SMEs) saw the extension of available
information to an employee on the production floor more critical than larger companies in a survey by
Jin et al [6]. In PHM sceptical companies, such measures were not able to build confidence with the
employees using them or the systems were perceived with rivalry, as they would take over elements
of the employees’ work. Johnsson et al found organizational issues to surpass human factors in their
relevance to PdM and defined recommendations on organizational and human factor related PdM
preconditions and PdM execution [8]. Among other points, they recommended employees in high
positions to favor PdM openly to raise acceptance and awareness. Another focus was especially set
on training and knowledge of employees, especially with regards to information technology and the
factors and inferences that influence a PdM system.

6.2 Data source limitations

Available relevant data is essential in the creation of a PdM model, however, companies are stated
to be seldom in possession of all useful data at the start of a PdM introduction [1]. After using the
already available data, gaps have to be specified and aimed to be solved. Additionally, the quality of
existing sources of information may not meet the required needs. If only a part of the data is affected
by unsatisfactory quality, this may be overcome in data preparation, as long as the quantity of data
Introduction 10

points is sufficient to reach a statistical significance and as long as fault detection can successfully
isolate critical points for the machine.

The PdM methods using company may then face challenges, when the necessary confidence in data
does not hold true, i.e. if sensors, controllers or other data sources, deliver imprecise or incorrect
measurements [17]. This may produce incorrect predictions and miss the urgency for maintenance or
send false alarms [11]. An additional challenge to sensor technology is that sensors currently tend to
work offline without contributing the data online. Moreover, sensors are subject to asset interruptions,
instrument degradation, and noise. Consequently, the improvement of stability, precision, and integrity
of source data is an essential challenge, which is recognized in the field of PHM [21].

A deterioration over time is additionally true for PdM prediction models, if they are not retrained
or automatically updated with new data [17]. According to Lee et al, in practice, many PHM based
systems are based on the data that was initially extracted when developing introducing the system,
but as an update with new data has to be performed by experts, it is often not applied. However,
newer data is more relevant for the prediction of the current machine and should be able to be added
to the system flexibly.

6.3 Limits of machine repair activities

Being able to predict an asset’s RUL, maintenance times can be determined, but the actual mainte-
nance of an asset still faces challenges related to a dependency on human interactions and missing
self-maintenance, which are explained in the following.

As assets are currently dependent on human operators for control and maintenance, the effectiveness of
maintenance is dependent on the quality of human management and skills [17]. Particularly machine
assets commonly work in a reactionary execution of commands and do not question the plan designed
for them. However, human task planning is based on data and experience, which the machine could
also be able to retrieve. In this way, an intelligent asset could propose or even autonomously start
beneficial actions for system health, asset throughput or product quality.

A further step into asset autonomy builds on asset self-awareness and self-maintenance [17]. A self-
aware asset can evaluate its conditions according to the data that is currently extracted to and stored
in a PdM system and may recognize critical conditions up to the autonomous definition of maintenance
decisions. Opposed to a central system controlling one or more assets, all necessary information to take
PdM decisions, as well as the degradation and prediction model, would be distributed and available
on an asset level. Machines may then plan maintenance schedules for themselves or for a group of
assets. However, currently, assets lack this level of self-awareness and self-maintenance.
Introduction 11

6.4 Optimization narrowness

An adequate decision on the most appropriate point in time for the maintenance of an asset depends
on the goal and the figures that are strived to be optimized. In literature, some authors question the
current aspirations in optimization as potentially being too narrow and state missing relevant figures
and interdependencies in their work, some of which are elaborated in the following.

A related figure to the health of an asset, especially in a production environment, is the consumption
of energy and environmental footprint that an asset creates [11]. Since the use of fossil fuels is
predominant in use for manufacturing companies, pollution created by the company directly correlates
to how much energy is consumed. Not only energy but also the use of raw materials and the type
of raw materials contribute to environmental contamination. Scrap as a byproduct of production is
interlinked with the quality and degradation status of an asset. Environmental impact is currently
not taken into account as a criterion in PdM decisions, apart from illustrations of environmental
footprint in production. However, illustrations for PdM models have a limited applicability due to
the complexity and high dimensionality of PdM models. Schaltegger and Synnestvedt stated in their
work on relations between environmental success and economic success that environmental efficiency
has most often an indirect economic influence, although the link is dependent on many factors, such
as regulations, customers’ consciousness to economic impact and readiness to accept higher costs for
environmental friendly produced products [22]. Hence, on the one hand, it is difficult for a company to
find its own importance of environmental success, which is a precondition to inserting it into the PdM
model. On the other hand, it may have an economic impact and may be beneficial to the company’s
success.

Another factor on a company’s success is how well a company manages its level of stocks. As stated in
the chapter Benefits of Predictive Maintenance, PdM methods often lack a simultaneous improvement
of the inventory level, which may lead to cases of imperfect preventative maintenance or inadequate
levels of stocks with potentially too many spare parts resulting in high costs [12].

A further gap in PdM models is frequently the inclusion of actual costs incurred with recommended
maintenance strategies [2]. Measures of quality for PdM are most often precision, recall, and the F1
score to quantify the divergence of predictions to occurrences in reality. In these figures, however,
maintenance strategy costs are not included. Exemplary costs given by a maintenance strategy are
incident observation and analysis, travel to the asset, labor costs for repair, spare parts, and downtime
of an asset. Spiegel et al propose a deeper analysis of company processes to derive a sharper view of
the cost structures and influences to costs. These can then be included in the PdM model design and
predictions.

An analysis view on maintenance and asset degradation, which multiple authors consider missing, is
the integration of multiple assets into one PdM system [17] [11]. Currently, the majority of predictive
Introduction 12

analyses concentrate on one or very few assets and circumstances. Although in manufacturing fre-
quently multiple machines, which belong to one company, are used in different environments and on
different activities, the data from those multiple machines is not used to achieve a clearer picture of
each machine and the fleet of machines is not analyzed as such. According to Efthymiou et al, already
some steps have been done to combine the data from different assets into a combined platform [11].

6.5 Sentiment towards challenges and outlook

Challenges from all four groups explained above may be considered when analyzing results from a
survey created by Jin et al across fifteen manufacturing companies and eight technology or consulting
firms of different size, which were asked about their current practice of PdM methods [6]. They found
that only a small amount of those companies applied PdM practices. Main barriers perceived by them
were costs, employees and their expertise, the readiness of the organization and technology as well as
the complexity of introduction. Further, regarding the technical reliability of PdM systems, multiple
manufacturers in their surveys complained about too many false positives in the analysis and alerts. In
spite of the lack of achievements with PdM uses so far, many PdM projects were at that time running
in these companies and the majority of company representatives uttered an optimistic opinion about
the future of PHM in their companies.

McKinsey & Company found in 2015 that the use of PdM methods commonly would reduce times,
in which machines are not running due to failures or maintenance, by 30% to 50% and would raise
machine life by 20% to 40% [4] [5]. As Bokrantz et al found numbers such as these to be reflected
in the expectations of companies and as they found a lack of research between the expectations in
digital production and the outlook of maintenance in the future, they included PdM as a part of their
research on the future of digitized maintenance. In this research they presented scientific expectations
for stable and efficient PdM systems in the future. To meet manufacturing companies’ expectations,
Bokrantz et al found various obstacles to that would need to be overcome. Necessary managerial
improvements in PdM systems would need to be achieved in security, data integrity, statutory rules,
network connectivity, human-machine cooperation, and standardization, e.g. in generally accepted
data interfaces and protocols [5]. Technological achievements would have to be made in the interop-
erability within machines, data science and the maturity of sensors and additive manufacturing. In
their organizational and financial situation analysis, they found the resistance to change as well as
restrictions in PdM budgets to be hindering factors which would need to be transformed [23] [9] [24]
[25] [5].
Chapter 7

Results and Discussion

7.1 Results

In this chapter, the elaborated challenges of PdM methods from section Challenges of Predictive
Maintenance are reflected with regards to the underlying questions of this paper, namely how reliable
predictive maintenance is and which challenges it faces.

The findings, whether PdM methods meet their expectations deviate across different studies. While
McKinsey & Company found PdM methods to achieve strong improvements in the reduction of down-
times and extension of asset lifetime, other authors such as Jin et al and Bokratz et al identified
dissatisfaction with the results of installed PdM solutions [4] [5] [6]. Particularly the number of false
positives referring to the amount of unnecessary corrective actions that had been recommended by the
PdM system were stated to be too high [6].

The factors and barriers, which currently prevent companies from reaching their PdM goals, origin
in the groups described throughout this paper. Some of these challenges, especially challenges of
non-technological nature, can be overcome by companies already at the current time, such as a higher
budget for prognostic measures or an intensified preprocessing of source data to meet the demands of
a sophisticated PdM system. Other challenges require extended research, e.g. improved visualizations
of complex systems or the technological ability to control a fleet of assets with one PdM system. Apart
from the challenges that have yet to be solved and the challenges that a company can overcome itself,
a company also needs to answer, whether their organization is suitable for the introduction of prog-
nostic measures. Johnsson et al described prerequisites and organizational recommendations, which a
company may consider. If they do not meet these factors, they would need to contemplate, whether
they aspire a change towards a PdM friendly organization and evaluate this change as achievable.

13
Discussion 14

Connecting the different factors for the reliability of PdM predictions, no universal statement about
the reliability of PdM methods can be given. Rather, reliability depends on the individual company
together with their way of application and their introduction measures. There are already companies
using PdM solutions satisfying their demands, but this does not apply to all [4] [5].

7.2 Discussion

The surveys taken in the referenced papers were limited to only a fraction of the general number of
companies, from which the experiences with PdM in companies have to be generalized. If there are
challenges not observed by the included companies, these are not known by the scientific community
and not be presented in this paper. Moreover, as effective PdM methods may create a competitive
advantage over competitors, not all companies may be inclined to present their solution to current
challenges to retain their competitive advantage. Therefore, this paper does not claim to be complete
in its presentation of PdM challenges.

As the expectations of PdM methods are set high by potential users, much research has already
been conducted on PdM [5]. By connecting various papers, particularly on their view of challenges
in PdM, a holistic view on PdM methods could be presented. No paper rejected the concept and
application of PdM entirely, but many challenges were raised. While various challenges, such as
the lack of appropriate visual presentation of PdM models [11][5] and lack of environmental impact
measurement [21] [11] have been raised in multiple sources of scientific literature for PdM, none of the
challenges mentioned in this paper have been already rejected or reported as entirely solved. However,
proposals for inclusions of new developments into future PdM solutions have been made, such as a cost
effectiveness method [2] or a joint optimization of inventory management with PdM [12]. Ultimately,
no paper has found PdM methods to be universally applicable. Nevertheless, considering a PdM
solution introduction, a company may review the challenges and references collated in this paper as a
decision basis. In the future, the connected factors for PdM solution reliability presented in this paper
may be validated in PdM practicing companies. With results from such a review, the single factors
may be weighted, until a guideline for a successful PdM practice at the current time can be derived.

As the current challenges of PdM, which are described in this paper, have been defined by the scientific
community, as research on the solution to these challenges has been proposed by multiple authors and
as there is a business demand from the optimistic companies using PdM, it may be expected that the
challenges presented in this paper will in the future be solved and the application use will increase for
manufacturing companies.
Chapter 8

Conclusion

In this paper, challenges and limitations to PdM methods have been presented and an outlook to the
further development of these challenges has been given. There are several challenges potentially limit-
ing the reliability and success of PdM methods in manufacturing companies. These challenges result
from organizational, economic and technological factors, which have been described. In consequence,
the challenges may limit PdM before its general introduction due to organizational and financial rea-
sons, in the use of source data, by the limits of machine repair activities and by potentially being too
narrow in their optimization focus.

It has been found that no universal statement about the reliability of PdM methods is applicable to
all companies. Company factors, such as size, organizational environment, technological readiness,
asset landscape and willingness to accept the necessary costs, influence the benefits that a company
may achieve with a PdM solution. While some technological challenges have not yet been solved,
through a clearly identified research gap and attention, further solutions may be expected in the future.
Therefore, this paper presents a point in time view on the current challenges faced by companies in
the application of PdM methods and may be considered when searching for the individual factors
that decide upon a specific company’s outlook of PdM method reliability. It is proposed to validate
these connected factors as a group with PdM practicing companies. Understanding the influences of
this new and developing field of PdM may then support to progress further through this enhanced
understanding.

15
Bibliography

[1] S. Diamond and A. Marfatia. Predictive maintenance for dummies. hoboken, 2013. URL
http://cdn2.hubspot.net/hub/215445/file-534377792-pdf/Predictive_Maintenence_
forDummies-Operations.pdf.

[2] Stephan Spiegel, Fabian Mueller, Dorothea Weismann, and John Bird. Cost-sensitive learning for
predictive maintenance. arXiv preprint arXiv:1809.10979, 2018. URL https://arxiv.org/abs/
1809.10979v1.

[3] G. Peón-Hernández. Feasibility of predictive maintenance techniques for the infrastructure in


cv, el, fm and he. 3rd TS Workshop, Archamps, France, 27 - 29 May 2008:163–168. URL
http://cds.cern.ch/record/1119526.

[4] McKinsey Report. How to navigate digitization of the manufactur-


ing sector. 2015. URL
https://www.mckinsey.com/~/media/McKinsey/
Business%20Functions/Operations/Our%20Insights/Industry%2040%20How%
20to%20navigate%20digitization%20of%20the%20manufacturing%20sector/
Industry-40-How-to-navigate-digitization-of-the-manufacturing-sector.ashx.

[5] Jon Bokrantz, Anders Skoogh, Cecilia Berlin, and Johan Stahre. Maintenance in digitalised
manufacturing: Delphi-based scenarios for 2030. International Journal of Production Economics,
191:154–169, 2017. ISSN 0925-5273. URL https://doi.org/10.1016/j.ijpe.2017.06.010.

[6] Xiaoning Jin, Brian A. Weiss, David Siegel, and Jay Lee. Present status and future growth
of advanced maintenance technology and strategy in us manufacturing. International journal
of prognostics and health management, 7(Spec Iss on Smart Manufacturing PHM), 2016. URL
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5207222/.

[7] Xiaodong Yao, Xiaolan Xie, Michael C. Fu, and Steven I. Marcus. Optimal joint preventive
maintenance and production policies. Naval Research Logistics (NRL), 52(7):668–681, 2005. ISSN
0894-069X. doi: 10.1002/nav.20107. URL https://doi.org/10.1002/nav.20107.

16
Bibliography 17

[8] Katrin Jonsson, Jonny Holmström, and Per Levén. Organizational dimensions of e-maintenance:
a multi-contextual perspective. International Journal of System Assurance Engineering and Man-
agement, 1(3):210–218, 2010. ISSN 0975-6809. URL https://link.springer.com/article/10.
1007/s13198-011-0043-z.

[9] Klaus-Dieter Thoben, Stefan Wiesner, and Thorsten Wuest. Industrie 4.0 and smart
manufacturing–a review of research issues and application examples. Int. J. Autom. Technol,
11(1), 2017. URL https://www.researchgate.net/profile/Thorsten_Wuest/publication/
312069858_Industrie_40_and_Smart_Manufacturing_-_A_Review_of_Research_Issues_
and_Application_Examples/links/586e0bb108aebf17d3a73562.pdf.

[10] Kevin A. Kaiser and Nagi Z. Gebraeel. Predictive maintenance management using sensor-based
degradation models. IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems
and Humans, 39(4):840–849, 2009. ISSN 1083-4427. URL https://www.researchgate.net/
profile/Nagi_Gebraeel/publication/224442007_Predictive_Maintenance_Management_
Using_Sensor-Based_Degradation_Models/links/56479e3108ae54697fbbe7e7.pdf.

[11] K. Efthymiou, N. Papakostas, D. Mourtzis, and G. Chryssolouris. On a predictive maintenance


platform for production systems. Procedia CIRP, 3:221–226, 2012. ISSN 2212-8271. doi: 10.
1016/j.procir.2012.07.039. URL https://doi.org/10.1016/j.procir.2012.07.039.

[12] Morteza Soltani. Joint optimization of opportunistic predictive maintenance and multi-location
spare part inventories for a deteriorating system considering imperfect actions. arXiv preprint
arXiv:1810.06315, 2018. URL https://arxiv.org/abs/1810.06315.

[13] Nagi Z. Gebraeel, Mark A. Lawley, Rong Li, and Jennifer K. Ryan. Residual-life distributions
from component degradation signals: A bayesian approach. IIE Transactions, 37(6):543–557,
2005. ISSN 0740-817X. doi: 10.1080/07408170590929018. URL https://doi.org/10.1080/
07408170590929018.

[14] J. Venkatesh. An introduction to total productive maintenance (tpm). The plant maintenance
resource center, pages 3–20, 2007. URL http://faculty.nps.edu/dl/sysengineering/se3302/
pdf/anintroductiontototalproductivemaintenance.pdf.

[15] Seyed Iravani and Izak Duenyas. Integrated maintenance and production control of a deteriorating
production system. IIE Transactions, 34(5):423–435, 2002. ISSN 0740-817X. URL https:
//link.springer.com/article/10.1023/A:1013596731865.

[16] David Golightly, Genovefa Kefalidou, and Sarah Sharples. A cross-sector analysis of human
and organisational factors in the deployment of data-driven predictive maintenance. Information
Systems and e-Business Management, pages 1–22, 2017. ISSN 1617-9846. URL https://link.
springer.com/article/10.1007/s10257-017-0343-1.
Bibliography 18

[17] Jay Lee, Hung-An Kao, and Shanhu Yang. Service innovation and smart analytics for industry
4.0 and big data environment. Procedia CIRP, 16:3–8, 2014. ISSN 2212-8271. doi: 10.1016/j.
procir.2014.02.001. URL https://doi.org/10.1016/j.procir.2014.02.001.

[18] S-H Suh, S-J Shin, J-S Yoon, and J-M Um. Ubidm: A new paradigm for product design and
manufacturing via ubiquitous computing technology. International Journal of Computer Inte-
grated Manufacturing, 21(5):540–549, 2008. ISSN 0951-192X. URL https://doi.org/10.1080/
09511920802023012.

[19] Keith Stouffer and Joe Falco. Guide to supervisory control and data acquisition (SCADA) and
industrial control systems security. National institute of standards and technology, 2006. URL
https://sm.asisonline.org/ASIS%20SM%20Documents/nist_scada0107.pdf.

[20] Ruben Sipos, Dmitriy Fradkin, Fabian Moerchen, and Zhuang Wang. Log-based predictive main-
tenance. 2014. URL https://doi.org/10.1145/2623330.2623340.

[21] Rajkumar Roy, Rainer Stark, Kirsten Tracht, Shozo Takata, and Masahiko Mori. Continuous
maintenance and the future–foundations and technological challenges. CIRP Annals, 65(2):667–
688, 2016. ISSN 0007-8506. URL https://doi.org/10.1016/j.cirp.2016.06.006.

[22] Stefan Schaltegger and Terje Synnestvedt. The link between ‘green’and economic success: en-
vironmental management as the crucial trigger between environmental and economic perfor-
mance. Journal of environmental management, 65(4):339–346, 2002. ISSN 0301-4797. URL
https://doi.org/10.1006/jema.2002.0555.

[23] Mario Hermann, Tobias Pente, and Boris Otto. Design principles for industrie 4.0 scenarios. 2016.
URL https://doi.org/10.1109/HICSS.2016.488.

[24] László Monostori, Botond Kádár, T. Bauernhansl, S. Kondoh, S. Kumara, G. Reinhart, O. Sauer,
G. Schuh, W. Sihn, and K. Ueda. Cyber-physical systems in manufacturing. CIRP Annals, 65
(2):621–641, 2016. ISSN 0007-8506. URL https://doi.org/10.1016/j.cirp.2016.06.005.

[25] Hyoung Seok Kang, Ju Yeon Lee, SangSu Choi, Hyun Kim, Jun Hee Park, Ji Yeon Son, Bo Hyun
Kim, and Sang Do Noh. Smart manufacturing: Past research, present findings, and future
directions. International Journal of Precision Engineering and Manufacturing-Green Technology,
3(1):111–128, 2016. ISSN 2288-6206. URL https://link.springer.com/article/10.1007/
s40684-016-0015-5.

View publication stats

You might also like