This is an Open Access document downloaded from ORCA, Cardiff University's institutional

repository: http://orca.cf.ac.uk/102384/

This is the author’s version of a work that was submitted to / accepted for publication.

Citation for final published version:

Oliff, Harley and Liu, Ying 2017. Towards industry 4.0 utilizing data-mining techniques: a case
study on quality improvement. Procedia CIRP 63 , p. 167. 10.1016/j.procir.2017.03.311 file

Publishers page: http://dx.doi.org/10.1016/j.procir.2017.03.311

<http://dx.doi.org/10.1016/j.procir.2017.03.311>

Please note:
Changes made as a result of publishing processes such as copy-editing, formatting and page
numbers may not be reflected in this version. For the definitive version of this publication, please
refer to the published source. You are advised to consult the publisher’s version if you wish to cite
this paper.

This version is being made available in accordance with publisher policies. See
http://orca.cf.ac.uk/policies.html for usage policies. Copyright and moral rights for publications
made available in ORCA are retained by the copyright holders.
 Available online at www.sciencedirect.com

ScienceDirect
Procedia CIRP 00 (2017) 000–000
www.elsevier.com/locate/procedia

The 50th CIRP Conference on Manufacturing Systems

Towards Industry 4.0 Utilizing Data-Mining Techniques: a Case Study on Quality Improvement
Harley Oliffa*, Ying Liua.
a
Mechanical and Manufacturing Engineering, Cardiff University, Cardiff, UK, CF22 3AA

* Corresponding author. Tel.: +44(0)7446 103980; E-mail address: [email protected]

Abstract

The use of data-mining as an analytical tool has been increasing in recent years; and the emergence of new manufacturing paradigms such as the
Industry 4.0 initiative have led many smaller manufacturers to look at utilizing these powerful techniques; however, practical applications are
still in their infancy, and remain out of reach for many of these small manufacturing enterprises (SME’s). This paper focuses on methods to
integrate these emerging paradigms into existing manufacturing processes, specifically, how data-mining principles may be used to begin to
explore the concept of Intelligent Manufacturing under Industry 4.0; with a focus on improving product and process quality.
In collaboration with an industrial partner; a respected manufacturer of household electronic appliances, techniques were developed using open-
source and freely-available software, running on readily available hardware and using only existing data-collection points, that were able to
provide actionable feedback which could be used to make improvements to the manufacturing operations; and to increase product quality. This
paper serves as evidence that the ability to utilise these techniques is now within reach of numerous smaller manufacturing operations, and
provides a further understanding of how moves towards fully Industry 4.0 ready factories may be made in the years to come.

© 2017 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of the scientific committee of The 50th CIRP Conference on Manufacturing Systems.

Keywords: Industry 4.0; Data-Mining; Intelligent Manufacturing.

1. Introduction networks of distributed control, and to self-adjust and self-

correct should problems arise.
The principles of data and text mining are long established The objectives have been prompted, in part, by the mounting
and well understood. However, the resurgence in popularity of pressures and challenges facing manufacturing industries in the
the field -due to recent successes in the field of machine new era. There is a massively increased demand for high
learning algorithms [31] has paved the way for this powerful quality, bespoke products [33], developed using sustainable
new tool to be adopted by industry. This forward thinking, and efficient methodologies. To meet this demand, intelligent,
combined with the development of the relevant technologies reconfigurable systems need to be developed. Estimates by
[28] and the entry into the age of ‘big-data’ [27]; has shifted the government agencies put the potential gains in efficiency of
ability to make use of these powerful methodologies beyond such processes as high as 30%.
the bounds of academic institutions. Initial steps have been taken in the implementation of
The ‘Industry 4.0’ methodology laid out at the 2013 intelligent systems, and many companies with large
Hamburg world fair [42] has developed into a focused and manufacturing requirements have begun to explore the
unique objective. The term, coined by the German government potential of this area [4], however, the vast investments needed
to describe their ongoing vision for manufacturing, whilst this [26] both in capital and skills present a significant obstacle.
concept is multi-faceted and often open to interpretation, there Through this research, we outline a methodology to adopt the
exists a clear theme of intelligent manufacturing; which makes principles of data-mining and utilize them to support decision
use of advanced computational technologies, and the making with respect to quality, at both the component and
advancements in digital systems and machine learning control levels.
processes to support decision making, run self-sufficiently via

2212-8271 © 2017 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of the scientific committee of The 50th CIRP Conference on Manufacturing Systems.
 Harley Oliff & Ying Liu/ Procedia CIRP 00 (2017) 000–000 2

2. Literature Review of’ Industry 4.0, and both Theoretical and Technological
advancements are being seen at an ever increasing rate. Main
2.1. Big Data & Industry 4.0 research focuses include: novel automation control systems,
with a focus on, decentralization, virtualization,
Since its introduction, the Industry 4.0 initiative has been reconfiguration, and adaptability [29, 9, 18, 35]; the
widely discussed and in the author’s opinion has migrated into development and application of machine learning and artificial
the role of a popular science buzzword to simply relate intelligences [32]; and virtual and augmented reality systems,
emerging digital and advanced manufacturing technologies. which are being used to bridge gaps in geography, knowledge
The initiative is broad, and a significant volume of focused and skill level [24]. In addition, other enabling technologies
research remains until such a paradigm can be realized if indeed and associated fields of research have also seen renewed
it is possible to develop a system in line with the Industry 4.0 interest and novel ideas. Algorithm development and software
initiative in its current form at all. engineering [36] have both seen a variety of successful
Much of the shift towards Industry 4.0 has been driven by
advances in previous years.
the emergence of ‘Big-Data’, and the issues associated with the
Current implementations have demonstrated adaptive
way industrial operations collect, manage and interpret their
scheduling, real-time modelling of processes, and Decision
data remain prevalent [7]. The concept of big data and
considerations of how to deal with such large datasets is an Support Systems, that have been used to refine processes and
intrinsic challenge of any system operating in an Industry 4.0 component design. Indeed, significant studies in this area have
environment as it typically renders traditional statistical resulted in a variety of frameworks by which to classify and
processing methods useless due to its complexity and sheer evaluate such emerging systems [30]. The 5C’s architecture,
size. proposed by Lee [17], outlines the different intelligence levels
Hilbert [13] outlines five main characteristics with which to achievable, and their associated technologies and capabilities.
describe big data: Volume, the quantity of generated and stored This architecture is illustrated in Figure.1.
data. Variety, the type, and nature of the data. Velocity, the
speed at which the data is generated and processed. Variability,
The consistency of the data. Veracity, the quality of the
captured data, which can vary massively between devices or
even individual sensors. These five features of big-data present
substantial challenges [5, 39], but are the source of its massive
potential.
It is well known that the rate of data generation, capture and
storage is continually increasing [14], and soon the volume of
data generation in this field will require the consideration of
potentially unbound datasets and continuous data streams [21].
However, despite the vast amounts of information that is being
generated, relatively few companies involved in the
manufacturing sector are utilizing this data [16]. Figure.1. The 5C’s architecture. Figure reproduced from [17]
Current research efforts [2, 26] have attempted to provide
comprehensive definitions of the necessary ‘criteria’ that need Many of the aforementioned technologies still require
to be met, across all areas of the business. Others, have significant development before they become realistic to use on
attempted to illustrate how the paradigm will be implemented an industrial scale, and as such are of limited use to those
in the future [41, 23]. However, Industry 4.0 is a multi-faceted without the funding to conduct their own research. Indeed,
problem, and it is unlikely that all aspects of it will be many obstacles to the revolution will become apparent only
applicable to all businesses. once the research reaches a commercial level. Issues such as
standardization [37] and validation [10] of such novel
2.2. Intelligent Manufacturing architectures are likely to further impede
progress for those manufacturing facilities without the
Intelligent Manufacturing describes any manufacturing necessary resources; as are ethical concerns and political
processes which involve a degree of computational interventions [40].
intelligence. This can be via the use of embedded sensors as in Whilst the area of Intelligent Manufacturing is itself a multi-
the case of IoT technologies [3], and cover the use of analytical faceted problem, the recurring element that underpins much of
techniques on historical process data to provide Knowledge this revolution is the collection, utilization and understanding
Discovery and support decision making within manufacturing of data, or the study of ‘Informatics'; almost all of the areas
systems [15, 19] or, ultimately, the development and linked with the intelligent manufacturing research area rely on
implementation of full Cyber-Physical-Systems [20], a the capture and analysis of data in some way. To this end the
synthesis of physical and digital technologies across the entire use of advanced data analytics and machine learning is a key
manufacturing system; and necessary associated technologies technology to develop to further these other technologies; and
and frameworks. the next step in this chain lies in utilizing the vast reserves of
Intelligent manufacturing itself encompasses many data through data mining and knowledge discovery, to better
emerging technologies and processes that are considered ‘part understand these manufacturing processes.
 Harley Oliff & Ying Liu/ Procedia CIRP 00 (2017) 000–000 3

2.3. Industry 4.0 in SME’s applicable and implementable to many manufacturing

processes.
Initial steps have been taken in the implementation of
intelligent systems [12], and their appearance in manufacturing Table 1. Attribute descriptions for the provided dataset.
operations globally is ever increasing. Attribute Data Type Description
Small-scale implementations have achieved success, but Line Nominal The production line used for
tend to focus on specific tasks, such as control of motors and manufacture
actuators to maintain process parameters. With a lack of Model Nominal The identifying model code
understanding at the process scale. Date Date Datestamp of each instance
The reality is that few companies have the necessary
Fault Group Text The Group of faults into which the
systems and capital in place to make leaps such as these in their specific fault falls, typical values:
operational processes, and find themselves presented with scratch/damage, electrical, fit, etc.
substantial barriers with respect to access. Due to the vast scope Fault Text Details of the specific nature of the
of the technologies and methodologies, and substantial costs fault
involved and lack of understanding and competence with Remedy Text Details of corrective action/disposal
advanced manufacturing techniques, at the employee level [1].
Remedy Detail Text Additional notes on corrective action
The current literature highlights a gap in the application of
Serial Number Nominal The unique serial number of the
these technologies. The rate of technological advancement in
affected product
this area is outpacing its adoption in the manufacturing sector,
Surname Nominal The surname of the quality engineer
as the challenges associated with practical use prevent many of
entering the data
the smaller operations from utilizing these advancements. With
this in mind, the following process was developed and
validated using a case study, to seek to overcome many of the The process focuses on building an analytic model to
common barriers to access. produce a set of rules to be used as a decision support system,
thus targeting the Cognition level in Lee’s 5C’s architecture. A
3. Process Development flowchart illustrating the steps of the proposed process can be
seen in Figure.2.
This section outlines the approach taken by the authors to
develop a system within the confines of the existing system to
implement data-mining to focus on the discovery of patterns
and knowledge with which to provide a decision support
system to the production engineers, with a focus on improved
quality. A system model that provides insights to support
decision making meets the necessary criteria of the cognition
level of the 5C’s architecture, demonstrating a level of
intelligence. The research was conducted with the support of
an industrial partner, a small manufacturing enterprise that
produces washing machines and tumble-dryers, in a range of
models. Discussions with our Industrial partner led to a list of
Figure.2. Flowchart illustrating the proposed system.
criteria to meet as follows: The process must be built around
the use of archived data; as automated digital collection of the
data would require significant investment; The process must be The manual data-collection methods resulted in data
developed to utilize readily available tools; The process must delivery in a discrete time period, typically daily. As such,
run without interference on established computing hardware software with which to perform this analysis was chosen on the
within the facility. basis of its ability to handle datasets rather than continuous data
Following discussions, a dataset was provided which streams.
consisted of a collection of re-work records, consisting of brief, The process will use the WEKA (Waikato Environment for
textual descriptions of observed faults, and the actions taken to Knowledge Acquisition) data-mining software. WEKA is a
correct these faults. This was supported by supplementary, powerful, java based, analytical tool focused on the application
nominal attributes; such as the model number, the date etc. A of data-mining techniques to datasets. Whilst many
full list of the attributes can be seen in Table.1. professional and supported data-mining software packages
Each instance in the dataset was representative of a single exist, WEKA offers distinct advantages to SME’s in that it is
fault, and the data could be provided at a daily rate or any quick and easy to implement and access; easy to use; and
specified combination thereof, and with significant historical requires zero financial investment.
archived data to support and train. Based on the nature of the
available dataset, a process was hypothesised that would enable 3.1. Case Study
the necessary preparation and knowledge extraction of the data.
Once validated through the case-study, the process would be Pre-processing of collected data is frequently necessary to
improve accuracy and reliability of predictions. No standard
data-collection methodologies exist, and the approach was
 Harley Oliff & Ying Liu/ Procedia CIRP 00 (2017) 000–000 4

necessarily heuristic. It was decided after careful examination considered preliminarily for this research, due to their ease of
of the dataset that as the dataset was only concerned with construction and interpretation.
failures and problems that had occurred, there was little Two main variations were tested: the PART algorithm, an
variation in the sentiment and style of the textual information. implementation within WEKA of the C4.5 algorithm [25]
The vast majority of instances all described some form of which uses a divide-and-conquer approach to build a decision
failure, the variance being the cause, and they were typically tree, before ‘pruning’ the unnecessary structures within the
written by the same operators describing the same issues, and tree; and the JRip or RIPPER (Repeated Incremental Pruning
hence were syntactically and stylistically similar. To overcome to Produce Error Reduction) algorithm [8].
this, it was decided to treat each instance of textual information An initial sample of the dataset containing 1000 instances
nominally; that is, instead of separating each instance out and of quality control entries had been used to validate the pre-
using a bag of words approach, the text contained would act in processing techniques, however, it was necessary to determine
the same way as any other nominal value. The rule-based the optimum dataset size, as both too many instances and too
learning algorithms would then be able to build a classification few could lead to inaccurate models. A 6000 instance dataset
model based on the frequency of the attribute values; attribute was prepared using the relevant pre-processing techniques, the
values that occur frequently within the same instance indicate PART and JRip algorithms were then run and evaluated using
a relationship between the attributes. a 10-fold-cross-validation, and the number of instances
The Fault Group attribute was selected to act as our class reduced between iterations. The percentage of correctly
attribute, and describes the category of fault recorded. Typical classified instances, when evaluated, is plotted for both
values include: ‘fit’, for faults involving assembly failures, and algorithms in Figure.3a.
‘scratch/damage’, where components are damaged and The results of this preliminary assessment indicated that a
unusable. Multiple factors may contribute to the occurrence of dataset size exceeding 5000 instances leads to negligible gains
these faults, and an accurate model would produce a set of in model accuracy for both algorithms. This corresponds to
rules, indicating how the different attribute values influence the approximately 10 days; an approximate working fortnight's
faults observed to be occurring. This rule set can then be used worth of records.
as a decision support system, supplying information to the
process engineers about the observed processes, through the
construction of an Ishikawa diagram, an established quality and
process control technique. The presentation of the analysis in
this way will enable multiple rules to be visualized as the causal
factors of each category branch on the diagram.
It was necessary to convert the textual data to the
lowercase and remove all spaces, to prevent the algorithm
distinguishing between different capitalizations or descriptions
of the same problem. The WEKA software considers the same
value written in the upper case a different value.
Consideration must also be given to outlier detection;
infrequently occurring events that may lead to inaccuracies in
the model. Using the RemoveFrequentValues filter, instances
with unique attribute values that occur only once in the dataset
can be removed. The InterquartileRange filter was then applied
to isolate and remove any other infrequent instances. Figure.3. Percentage of correctly classified instances against the number of
Feature Selection in this instance was deemed unnecessary
instances. a) Evaluated using 10-fold-cross-validation. b) Evaluated using
due to the limited number of attributes within the dataset.
isolated Test Set.
However, consideration was given to the attributes that would
be used to extract information. Several attributes exhibited little
By logically partitioning the dataset into a decided time
variation, and others, it was clear, had little information to offer
step, both short-term and long term patterns may be
in terms of assembly faults. As such these attributes should be
discovered, by considering the duration over which the data to
removed to minimise noise.
be analysed was collected. The process also aims to be self-
validating: As rules are uncovered and used to make decisions
3.2. Algorithms
and take corrective actions, where the source of failure is
resolved, the rules will change to reflect different patterns, as
Extensive literature exists covering a wide range of the the prevalence of the resolved fault will decrease in the dataset.
techniques that may be utilised as a part of the data mining In order for this to remain true, each subsequent set of rules
process. A major family of algorithms are those which focus on produced via the method outlined must be considered to
Rule Based learning. As explained by [22], these algorithms supersede the previous set in terms of validity; the most recent
are the oldest; some of the simplest; and work by using set is the most accurate analysis of the current state.
mathematical relationships to determine a set of ‘rules' by One potential issue with using a cross-fold-validation
which to classify the data. As the computing power available technique for model evaluation is that of overfitting, models
continues to increase, these algorithms are becoming tested on the data used to train them often learn the patterns
increasingly complex. These types of algorithm were within that dataset, but perform worse when tested on data
collected at a different instance in time. To remove this factor,
 Harley Oliff & Ying Liu/ Procedia CIRP 00 (2017) 000–000 5

and isolated Test-Set consisting of 1000 instances was created patterns contained in any future datasets produced. To
and used to evaluate the PART and JRip algorithms in the same demonstrate this, a further dataset containing the entries from
manner as before. The results are shown in Figure.3b. 3 days of runtime, a Set of Rules was produced. Those with the
The results of the Test-Set evaluation supported the highest coverage (a metric which expresses the number of
conclusion that increases in accuracy are negligible when the correctly classified instances in comparison to the number
dataset size begins to exceed 5000 instances. The accuracy of incorrectly classified) are shown in Table.3. Rules with a high
the JRip and PART algorithms for the 5000 instance tests were coverage, are not only prevalent in the dataset (and hence occur
94.5% and 96.0% respectively. frequently), but are also those which the generated model is
able to predict with the highest degree of accuracy.
3.3. Improving Model Accuracy
Table 3. Rules Generated by the final model
Two methods of improving classifier accuracy are bagging and Rule Class (Fault Group) Coverage
boosting which are both methods that fall into the category of
Fault = Plinth Fit 15.7/0.0
ensemble learning. Standard classifiers build simple models of AND Model = 85969
the data, whereas, in ensemble learning, multiple base models Fault = Timer knob AND Fit 31.3/0.0
are combined to produce an amalgamated model. Model = 74628
Bagging involves the creation of new datasets for multiple Fault = Door Assy Fault Scratch/Damage 45.5/0.0
classifiers. In bagging, a dataset of N instances (where N is the AND Remedy = Change Part
size of the original dataset) is created by randomly drawing Fault = Drum Fit 175.0/9.0
with replacement. The replacement means that instances from AND Model = 74628
the original dataset may occur more than once, or not at all. Remedy = Retest Auto Auto Test 213.0/0.0
These models then utilise a voting system to fully develop the
Remedy = Fit Part Missing Part 13.2/0.0
final classifier.
AND Date = 13/11/2015
Boosting involves the creation of a series of classifiers,
Remedy = Fit Part Missing Part 9.0/0.0
where each in the series is given a different training set that is AND Date = 12/11/2015
based on the performance of the preceding classifiers, and their
Fault = Worktop Fit 34.9/0.0
prediction errors. Instances that were incorrectly predicted in AND Remedy = Refit
previous models are given a greater weighting than those
classified correctly and are more likely to be chosen for future
datasets. In this way, the classifier becomes iteratively better, From Table.3, several insights can be found directly, without
by focusing more heavily on the weaker areas of its learning at the use of additional quality tools. For instance, the model
successive stages. To implement these two ideas into WEKA, highlights that the Plinth component on the 85969 model
algorithms exist in the WEKA toolkit, the AdaBoostM1 produced by the company is a frequent source of failure,
algorithm to boost a classifier, and the bagging algorithm specifically relating to the fit of the component to the product.
[6,25].
The 5000 instance dataset was tested using bagged and 4. Case-Study Insights
boosted versions of the algorithms. Whilst running the
computations, the AdaBoostM1 algorithm, when used with the The aim of this case study was to establish how best to
PART algorithm, would cause WEKA to run out of heap utilise data mining to improve assembly and quality control
memory. Whilst it is possible to work around this issue, the processes; to allow them to be implemented into existing
concept of this case study was to produce a simplified and systems, with a minimal impact.
easily implementable procedure, and the necessary Validated results have been produced which can be easily
understanding of computation was judged to be excessive. The be interpreted and become actionable pieces of information. To
results of the bagging and boosting and their effects on the this end, the proposed system can be said to demonstrate an
accuracy of the developed models can be seen in Table.2. effective Decision Support System and qualifies at the
cognition level in the 5C’s architecture; demonstrating a level
Table 2. Accuracies of the Bagging and AdaBoostM1 meta-algorithms. of intelligence in-line with the Industry 4.0 initiative. In
Algorithm Bagging Accuracy AdaBoostM1 Accuracy addition, the case-study aims were fully met: the final system
is implementable, works with archived data, and has a low
JRip 95.4% 97.3%
computational requirement by design. The approach is
PART 96.9% - adaptable, and as long as suitable care is taken to correctly
partition the data and understand the effect that this partitioning
3.4. Decision Support Generation will have, the method can be used to determine a vast number
of different patterns depending on how the dataset is divided
A validated process now exists by which to produce a model, initially. Additional study of outlier detection and advanced
consisting of a set of rules, that can be used to support decisions algorithms could further refine the results, however the global
regarding product quality issues. The JRip algorithm and the model accuracy and high coverage of the drawn conclusions
Boosting technique (to improve accuracy), can be used to lead to considerable confidence that the results support real-
produce an accurate model, with a sufficient degree of world trends.
confidence. The model is trained using a 5000 instance dataset. Whilst the methods developed demonstrated the
This model can then be used to make predictions about the possibility of using data-mining in this way, they are by no
means ideal, and several challenges remain to be overcome.
 Harley Oliff & Ying Liu/ Procedia CIRP 00 (2017) 000–000 6

The process is very much a demonstration of supervised [15] Lee, J., Bagheri, B. and Jin, C., 2016. Introduction to cyber
manufacturing. Manufacturing Letters, 8, pp.11-15.
learning, and whilst valid, it requires significant human input,
[16] Lee, J., Kao, H.A. and Yang, S., 2014. Service innovation and smart
in terms of both processing, and interpretation. The next logical analytics for industry 4.0 and big data environment. Procedia CIRP, 16,
step in the evolution of this process would be an automated pp.3-8.
system, which would perform the necessary corrective actions, [17] Lee, J., Bagheri, B. and Kao, H.A., 2015. A cyber-physical systems
or notify quality and process engineers of emerging trends. architecture for industry 4.0-based manufacturing systems.
Manufacturing Letters, 3, pp.18-23.
5. Concluding Remarks [18] Leitão, P., 2009. Agent-based distributed manufacturing control: A
state-of-the-art survey. Engineering Applications of Artificial
Intelligence, 22(7), pp.979-991.
As explored previously, many manufacturing enterprises [19] Mi, M. and Zolotov, I., 2016, February. Comparison between multi-
are keen to adopt principles of intelligent manufacturing, but class classifiers and deep learning with focus on industry 4.0. In 2016
are presented with a barrier to doing so. This work presents Cybernetics & Informatics (K&I) (pp. 1-5). IEEE.
[20] Monostori, L., 2014. Cyber-physical production systems: roots,
evidence, that some of these barriers preventing such adoption, expectations and R&D challenges. Procedia CIRP, 17, pp.9-13.
may be overcome with considered use of freely available [21] Moran, K. 2012. Data analytics for manufacturing. Advanced
software and existing data. The industry 4.0 initiative places Manufacturing Technology. vol.5.
significant emphasis on the utilisation data to form intelligent [22] Nosofsky, R.M. et al. 1994. Comparing modes of rule-based
classification learning: A replication and extension of Shepard,
systems and processes, and by exploring the ways in which Hovland, and Jenkins (1961). Memory & cognition. Vol 22. pp.352-369.
companies may utilise their existing records, such an intelligent [23] Pan, M., Sikorski, J., Kastner, C.A., Akroyd, J., Mosbach, S., Lau, R.
system has been presented. Whilst in this instance, the and Kraft, M., 2015. Applying industry 4.0 to the Jurong Island eco-
methodology proved useful, countless variations in industrial park. Energy Procedia, 75, pp.1536-1541.
[24] Pirvu, B.C., Zamfirescu, C.B. and Gorecky, D., 2016. Engineering
manufacturing processes mean that such a problem is difficult insights from an anthropocentric cyber-physical system: A case study
to generalize to all processes, and significant further work is for an assembly station. Mechatronics, 34, pp.147-155.
required in this field to realise the full potential of intelligent [25] Quinlan, J.R., 1996, August. Bagging, boosting, and C4. 5.
manufacturing. In AAAI/IAAI, Vol. 1 (pp. 725-730).
[26] Rüßmann, M., Lorenz, M., Gerbert, P., Waldner, M., Justus, J., Engel,
P. and Harnisch, M., 2015. Industry 4.0: The Future of Productivity and
References Growth in Manufacturing Industries. Boston Consulting Group.
[27] Russom, P. 2011. Big Data Analytics. TDWI Best Practices Report.
[1] Adolph, S., Tisch, M. and Metternich, J., 2014. Challenges and Fourth Quarter.
approaches to competency development for future production. Journal [28] Sebastiani, F. 2001. Machine Learning in Automated Text
of International Scientific Publications–Educational Alternatives: 1001- Categorization. The National Research Council, Italy.
10. [29] Shafiq, S.I., Sanin, C., Szczerbicki, E. and Toro, C., 2016. Virtual
[2] Almada-Lobo, F., 2016. The Industry 4.0 revolution and the future of Engineering Factory: Creating Experience Base for Industry
Manufacturing Execution Systems (MES). Journal of Innovation 4.0. Cybernetics and Systems, 47(1-2), pp.32-47.
Management, 3(4), pp.16-21. [30] Shearer, C. 2000. The CRISP-DM model: the new blueprint for data
[3] Atzori, L. et al. 2010. The internet of things: A survey. In: Computer mining. Journal of Data Warehousing. Vol 5, pp. 22
networks. Vol 54. pp. 2787-2805. [31] Silver, D. and Hassabis, D., 2016. AlphaGo: Mastering the Ancient Game
[4] Auschitzky, E. et al. 2014. How Big Data can improve manufacturing of Go with Machine Learning. Research Blog.
[Online]. Available at: http://www.mckinsey.com/business- [32] Spezzano, G. and Vinci, A., 2015. Pattern Detection in Cyber-Physical
functions/operations/our-insights/how-big-data-can-improve- Systems. Procedia Computer Science, 52, pp.1016-1021.
manufacturing[Accessed: 16/10/2016] [33] Tao, F. et al. 2011. Cloud manufacturing: a computing and service-
[5] Babcock, B., et al. 2002. Models and issues in data stream systems. oriented manufacturing model. Proceedings of the Institution of
In Proceedings of the twenty-first ACM SIGMOD-SIGACT-SIGART Mechanical Engineers. Part B. Journal of engineering manufacture.
symposium on Principles of database systems pp. 1-16. Vol 225. pp.1969-1976.
[6] Breiman, L. 1996. Bagging predictors. Machine learning. Vol 24. [34] Theorin, A., Bengtsson, K., Provost, J., Lieder, M., Johnsson, C.,
pp.123-140. Lundholm, T. and Lennartson, B., 2016. An event-driven manufacturing
[7] Chen, F., Deng, P., Wan, J., Zhang, D., Vasilakos, A.V. and Rong, X., information system architecture for Industry 4.0. International Journal
2015. Data mining for the internet of things: literature review and of Production Research, pp.1-15.
[35] Valckenaers, P. and Van Brussel, H., 2005. Holonic manufacturing
challenges. International Journal of Distributed Sensor Networks, 2015,
execution systems. CIRP Annals-Manufacturing Technology, 54(1),
p.12.
pp.427-432.
[8] Cohen, W.W. 1995. Fast effective rule induction. Proceedings of the
[36] Vyatkin, V., 2013. Software engineering in industrial automation: State-
twelfth international conference on machine learning. pp. 115-123.
of-the-art review. IEEE Transactions on Industrial Informatics, 9(3),
[9] da Silva, R.M., Benítez-Pina, I.F., Blos, M.F., Santos Filho, D.J. and pp.1234-1249.
Miyagi, P.E., 2015. Modelling of reconfigurable distributed
[37] Weyer, S., Schmitt, M., Ohmer, M. and Gorecky, D., 2015. Towards
manufacturing control systems. IFAC-PapersOnLine, 48(3), pp.1284-
Industry 4.0-Standardization as the crucial challenge for highly
1289. modular, multi-vendor production systems. IFAC-PapersOnLine, 48(3),
[10] Feth, P., Bauer, T. and Kuhn, T., 2015. Virtual Validation of Cyber
pp.579-584.
Physical Systems. In Software Engineering & Management (pp. 201-
[38] Witten, I.H. et al. 2011. Data Mining: Practical Machine Learning
206). Tools and Techniques. 3rd Ed. Elsevier.
[11] Freund, Y. and Schapire, R.E. 1997. A decision-theoretic generalization
[39] Wu, X. et al. 2014. Data Mining with Big Data. IEEE Transactions of
of on-line learning and an application to boosting. Journal of computer
Knowledge and Data Engineering. Vol 26. Pp 97-107.
and system sciences. Vol 55. pp.119-139. [40] Zhou, K., Liu, T. and Zhou, L., 2015, August. Industry 4.0: Towards
[12] Hansson, K., Yella, S., Dougherty, M. and Fleyeh, H., 2016. Machine
future industrial opportunities and challenges. In Fuzzy Systems and
Learning Algorithms in Heavy Process Manufacturing. American
Knowledge Discovery (FSKD), 2015 12th International Conference
Journal of Intelligent Systems, 6(1), pp.1-13. on (pp. 2147-2152). IEEE.
[13] Hilbert, M. 2016. Big Data for Development: A Review of Promises and
[41] Zuehlke, D., 2010. SmartFactory—Towards a factory-of-things. Annual
Challenges. In: Development Policy Review. Vol 34. pp.135-174.
Reviews in Control, 34(1), pp.129-138.
[14] J. Gantz and D. Remsel, ‘‘The digital universe in 2020: Big data, bigger [42] Zühlke, D. 2014. Industrie 4.0: From Vision to Reality. Tech News. Vol
digital shadows and biggest growth in the far east,’’ in Proc. IDC iView,
26. pp 10-11.
IDC Anal. Future, 2012.