Power Plants Failure Reports Analysis for Predictive Maintenance
Vincenza Carchiolo
1
, Alessandro Longheu
2
, Vincenzo di Martino
3
and Niccolo Consoli
2
1
Dip. di Matematica e Informatica, Universita’ di Catania, Viale Andrea Doria 6, Catania, Italy
2
Dip. di Ingegneria Elettrica, Elettronica e Informatica, Universita’ di Catania, Viale Andrea Doria 6, Catania, Italy
3
BaxEnergy, Catania, Italy
Keywords:
Predictive Maintenance, Natural Language Processing, Ontologies, Wind Turbines, Renewable Energy.
Abstract:
The shifting from reactive to predictive maintenance heavily improves the assets management, especially for
complex systems with high business value. This occurs in particular in power plants, whose functioning is
a mission-critical task. In this work, an NLP-based analysis of failure reports in power plants is presented,
showing how they can be effectively used to implement a predictive maintenance aiming to reduce unplanned
downtime and repair time, thus increasing operational efficiency while reducing costs.
1 INTRODUCTION
It can be nowadays considered as a matter of fact
that climate changes and the increasing energy de-
mand are competing factors that push for deep ex-
ploitation of renewable energies (IEA, b); in partic-
ular their market share is expected to grow up to the
40% the next five years (IEA, a), overcoming the con-
tribution of coal and gas.
Among all sources as hydropower, geothermal,
solar and others, the forecasts for wind as renewable
energy endorse its position as one of the most rele-
vant (estimated in 2018 as 43% higher with respect
to 2015 (Council, 2019)). Wind power comes mainly
from wind turbines, that are mechatronic devices us-
ing blades and rotor to convert wind into mechanical
energy and shafts and generator to transform motion
into electrical energy.
Wind turbines (WT) industry is growing faster and
faster and larger devices are being developed, though
WTs are continuously exposed to (possibly) extreme
weather conditions, resulting in significant mechani-
cal stress and moreover both on-shore, as well as off-
shore placement, is often characterized by restricted
accessibility. This scenario can heavily impact on
WTs reliability, often leading its components to fail
during WTs’ lifetime (Guolin et al., 2016), therefore
making the Operation and Maintenance (O&M) a crit-
ical activity that actually impacts for up to 30% of the
WT life cycle (Fischer et al., 2012).
In particular, since any technical intervention (e.g.
replacing a WT damaged part) can be very expensive
and, moreover, the lack of power generation during
downtimes also impacts on revenues (Herbert et al.,
2010), several strategies can be adopted to limit these
expenses (Abichou et al., 2014), from condition mon-
itoring (CM) systems that falls into the so-called sig-
nal processing approach (signals can be WT blades
vibration, or acoustic emissions and/or thermography
measurements of WT internal components), to nu-
merical models of WTs, or data-driven strategies e.g.
based on SCADA (supervisory control and data ac-
quisition) (M
´
arquez et al., 2012) (Nabati and Thoben,
2017).
All these strategies can be effectively exploited to
endorse the Predictive Maintenance, which focuses
on identifying the optimal time to perform mainte-
nance, in particular after some WT working condi-
tion starts to decline and performance to decrease,
but before failure occurs. It, therefore, aims to es-
tablish a trade-off between preventative maintenance,
which uses strict time-based scheduling and may oc-
cur (and cost) too frequently, and reactive or run-to-
failure maintenance, where components are repaired
only after they have already failed (Selcuk, 2017).
Current studies on WT focused on predictive
maintenance because good wind turbine reliability,
together with predictable wind turbine maintenance
schedules, will result in reduced cost of energy
and then wind farm project success. This is even
more important for offshore wind farms because of
their higher initial capital cost and limited acces-
sibility causing higher operational and maintenance
costs (Qiu et al., 2012).
A traditional approach to predictive maintenance
is based on the use of SCADA systems, that provide
rich information about the plant itself giving signal
information and component information.
404
Carchiolo, V., Longheu, A., di Martino, V. and Consoli, N.
Power Plants Failure Reports Analysis for Predictive Maintenance.
DOI: 10.5220/0008388204040410
In Proceedings of the 15th International Conference on Web Information Systems and Technologies (WEBIST 2019), pages 404-410
ISBN: 978-989-758-386-5
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
There are a lot of research that, focusing on these
systems as a primary source of entry, has achieved
a good success in reporting failures and problems
within the plant through power-curve and temperature
analysis (Nabati and Thoben, 2017).
Some of these research outcomes have been rec-
ognized by industry and turned into applications. The
main advance in this approach is declined in the use
of artificial intelligence to analyze the collected infor-
mation. For example (Huuhtanen and Jung, 2018) use
convolutional neural networks (CNN) for monitoring
the operation of photovoltaic panels. The predictive
maintenance in this approach is activated when it is
observed a large deviation between predicted and ac-
tual (observed) power curve, whereas (Helbing and
Ritter, 2018), (de Azevedo et al., 2016) and (Romero
et al., 2018) show example of the analysis of SCADA
data through the use of deep learning for fault detec-
tion in wind turbines.
Improving wind turbine reliability requires to re-
duce downtime and increase availability by optimiz-
ing its design and imposing a well-organized main-
tenance schedule. This requires a full understand-
ing of the system and a detailed analysis of its fail-
ure mechanisms and cases, therefore in addition to
SCADA systems, a good approach is the exploita-
tion of maintenance report, whose content can be
effectively analyzed to extract relevant information
about WT failure components, endorsing the predic-
tive maintenance approach.
A maintenance report is a document in which
there are important data about the WT and in which
there probably is the main cause that led to the cor-
respondent status of the plant at that time. Our goal
is to find a correlation between the data shown in the
maintenance reports and the possible causes of fail-
ure; to this purpose, we use Natural Language Pro-
cessing (NLP).
NLP was formerly known as Computational Lin-
guistics in the 60s (Wagner, 2016) and it uses compu-
tational techniques and artificial intelligence to under-
stand, learn and synthesize human language content.
The foundations of NLP lie in several disciplines as
computer and information sciences, linguistics, arti-
ficial intelligence and robotics, psychology, philoso-
phy, logic and mathematics, electrical and electronic
engineering. Applications of NLP include machine
translation, natural language text processing and sum-
marization, user interfaces, multilingual and cross-
language information retrieval, speech recognition,
artificial intelligence, and expert systems, and many
others. Recently, in particular in the last 20 years, this
affected not only scientific research but also practi-
cal applications that have been successfully integrated
into consumer products as the Apple Siri (Apple, ) or
Microsoft Cortana (Microsoft, ) personal assistant, or
even in more specific context, ranging from personal
medical records data gathering, feeding and analy-
sis (Carchiolo et al., 2015) (Carchiolo et al., 2015)
to Twitter and/or Web data discovery for several pur-
poses (Carchiolo et al., 2015) (Longheu et al., 2016).
Here we adopt a NLP-based approach together
with Ontology-based information extraction for cap-
turing syntactic and semantic relations within words,
allowing to leverage maintenance phrases discovering
what went wrong and determined the failure.
The rest of paper is organized as follows: In sec-
tion 2 we address the question of analyzing mainte-
nance report, whereas in section 3 the overall project
that leverage information coming from reports is dis-
cussed; section 4 briefly concludes our work and show
future directions.
2 MAINTENANCE REPORT
NLP-BASED ANALYSIS
This section presents an analysis and methodology
applicable to the data that have been extracted from
maintenance reports of WTs.
In pursuing this goal, the first problem is to dis-
cover the meaningful information hidden in mainte-
nance reports and to collect all useful data from.
The main problem is to discover all the informa-
tion hidden in the maze of a maintenance report that
it can, by its nature, be very varied in form and infor-
mation. The information presents in a maintenance
report can have different view:
structured versus unstructured information
fault reporting information vs. repairing operation
measurable and non-measurable information
For example, the fault reporting can be described
by either a well-coded data (almost structured) as the
alarm code provided by a software tool or a qualitative
description, via a natural language, of the unexpected
behavior.
In the same way, the repairing operation descrip-
tion can contain the list of the WT replaced parts (each
with numbers and costs), the operations carried out
for the repair, the work hours but also information
about the success of the technical intervention; the lat-
ter, probably, hidden in a descriptive sentence written
in some natural language.
Fig 1 shows some examples of wind turbine main-
tenance report and it is notable how they contain a
lot of information, with different spatial organization,
Power Plants Failure Reports Analysis for Predictive Maintenance
405
Figure 1: Maintenance Report Examples.
different degree of structuring (free test, lines, num-
ber, error code, and so on), several forms and some-
times using also different languages. In fact, each ev-
ery wind turbine manufacturer use a customized form.
It is clear that such data require different techniques
due to their heterogeneity, e.g. data mining, sentiment
analysis or similar.
Limited research activities have been conducted
for detecting failures related to WTs. Sometimes,
NLP techniques are applied for the identification of
technology trends (Lee and Lee, 2013); they are
though applied with the aid of large available input
repositories. Those inputs consist of a lot of patent
documents from which the analysis can start. The
work uses those techniques to identify new innova-
tion patterns in the field of energy technologies.
There are other similar works that focus on a large
dataset of unstructured data and apply those tech-
niques for knowledge discovery of accident informa-
tion in the text. The involved text mining pipeline fo-
cuses on those inputs to recognize the main risk fac-
tors, which can be similar to the case of WTs main-
tenance report discussed in this paper. Those acci-
dent dataset has been collected over a 12-month pe-
riod, through collecting data from different available
accident datasets on the network. A smaller portion
of those documents has been recognized by an expert
analysis team and has been used to build the ontology
terms in the model proposed in (Ertek et al., 2017);
however, this study cannot be used in our use case
since there are not equivalent failure dataset available
in the network.
There are other researches on the text mining
domain that focus on a different technique named
FMECA (Failure Mode, Effects and Criticality Anal-
ysis), that is recognized as one of the earliest tech-
niques for failure analysis (Bouti and Kadi, 1994).
FMECA is used to analyze the reliability and safety
of equipment with inductive logic. This work is pretty
similar to our use case since it will involve the de-
velopment of an ontology-based model with detailed
concepts. Based on the analysis of the main re-
search contents of fault diagnosis coming from ma-
chine tools there are four main concepts that has been
recognized by this work. They are known as fault phe-
nomenon, fault maintenance, fault cause and fault lo-
cation and they are seen as fault events (Zhou et al.,
2017). Those concepts will be used to retrieve text
information inside the text with the use of clustering
techniques. However, this approach cannot be use-
ful to our domain case because it reflects a superset
of our failure maintenance reports sentences and be-
cause the developed ontology has been build with a
bigger dataset and with specific domain information
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
406
like frequencies of failures, severities, locations, mo-
tivations that are not available in our use case.
In (K
¨
uc¸
¨
uk and Arslan, 2014) authors propose a
semi-automatic approach to build an ontology model
for wind energy domain. This process involve the
crafting process of different Wikipedia articles related
to the wind energy domain. However as the previous
one this cannot be used to setup the proposed ontol-
ogy model because it reflects a different domain case.
Our solution involves NLP techniques and
Ontology-based information extraction for capturing
syntactic and semantic relations within words in a
text, as specified previously. Data is composed of
maintenance reports regarding WTs maintenance ac-
tivities and provided by Bax Energy. A main short-
coming is that reports provided as input are not
enough, because only a small part of them contains
analyzable data.
These reports make references to maintenance
phrases that can be analyzed to find out what went
wrong or not within the same report. The use of
these techniques will allow identifying domain con-
cepts within the text and if they respect the relation-
ships of the proposed ontology. In fact, it contains all
the relationships that have been built between a single
failure and potential agents of failure that can provoke
it, structural components of the plant that may incur
at that specific failure and the condition that triggered
that failure event for that specific component. So, the
work shows the combination of a semantic module,
that is based on the ontology information retrieval,
with a set of natural language processing techniques
that goes deep into the sentence to clean up it, tok-
enizes it and retrieves possible syntactic dependencies
that are useful to build up domain individuals.
3 THE WEAMS PROJECT
In the context of WTs, if we want to find a correla-
tion between the data shown in the maintenance re-
ports and the possible causes of failure, first of all,
it is necessary to gather a significant data amount for
an analysis purpose. This data can already be col-
lected from the developed software and also from dif-
ferent sources. In fact, the document will show how
the grafting of different types of data, or rather the ex-
tension of our dataset can not only specify and further
enrich our analysis purpose but also improve the final
result in terms of accuracy.
In this section, we will present the solution of
Weams project; the proposal is schematized in fig. 2,
where several modules are present, i.e. the Data col-
lector, data parser, data reader and matcher and fi-
nally the data writer joint with translate module. All
these components are described in the following para-
graphs.
3.1 Data Collector
The first module is named data collector and is re-
sponsible for the acquisition of the relevant parts of
interest from maintenance reports. Uploaded docu-
ments pass through the PDFBox library (PDFbox, ),
used for preliminary reading of each string in the file.
The algorithm starts by collecting all relevant
parts as defined in the user template. They are di-
vided into static and dynamic parts, whose length may
vary within each user document. Since the extraction
area for fixed parts is fixed for every document, such
parts can be extracted immediately, whereas dynamic
parts require the algorithm looks for all the content
that belongs to them. Read strings are stored into a
specific data structure for further processing and as-
sociated with the relevant part of interest. In addi-
tion, the library provides important information like
the reference coordinates within a 2D space of each
character extracted from the document, to be used
later to remap the content. Data acquisition also pro-
vides the removal of unwanted data as specified by
the user (e.g. header/footer).
Once the noise data has been removed, the acqui-
sition operation starts by comparing each read string
with the one entered by the user in order to map that
specific part. If this comparison succeeds, the algo-
rithm will register each subsequent string in a struc-
ture that will be identified by the compared string.
The structures can have different composition, de-
pending on the nature of the current part. If it is re-
ported from the template as table, the related structure
will also have a list of TextPagination objects. Each
object has a list of TextPosition properties, the whole
information that are taken from the PDFBox library,
and the relative page in which this string has been
found. This information is important for the next step
in which the content is ordered before building the re-
lated table structure.
When the part is reported as free text, the struc-
ture stores the starting point as X-Y 2D coordinates of
each string, the related endpoint and the page where
it has been found. Since this occurs for each page, the
content of a page does not depend on others. More-
over, we consider the membership page of each string
and its order and we compare with others’ in the list
to find the maximum value of that page. Such value
is used as the start value of the first string on the next
page. All strings that belong to pages after the first
will see the value of the ordinate changed, which is
Power Plants Failure Reports Analysis for Predictive Maintenance
407
Figure 2: Weams Report Architecture.
not related to the previous page. This value is modi-
fied with the sum of the maximum value found on the
previous page, i.e. the maximum ordinate (Y) of the
last useful string on the previous page, and their cur-
rent position on that page. This procedure for eval-
uating minimum and maximum values for each page
is performed for all the pages in the document. Af-
ter establishing a vertical data order, we then aggre-
gate each row’s content by a horizontal arrangement.
Once these operations are completed, the content thus
ordered and collected for each part of interest is ready
to be further processed by the parser.
3.2 Data Parser
This module takes as inputs the different parts ex-
tracted from Data Collector and processes the parts
in tabular format. In fact, the tabular information is
still not homogeneous therefore it will be necessary to
create a structure in rows and columns for each record
mapped within the part being examined. The process
of building the table foresees the use of the TrapRange
library (Luong, 2015) which has been modified for
our case. Specifically, the changes made allow a cor-
rect interpretation of tables whose headings are di-
vided into several lines, or in which there are columns
that subsequently branch into two or more columns
and a correct interpretation of the rows of a table that
can develop on more lines.
Concerning the first point, the user can specify in-
side the template which columns exhibit bifurcations,
passing the index of the column of the interested table
(0-based). This will allow the algorithm to exclude
the word mapped to this index and to take into con-
sideration the columns underlying it. As a result, the
table will be in a structured and homogeneous form
and the analysis of the number of rows and columns
of the same can be carried out. This analysis is done
by the software using TrapRange which identifies the
limits of each row and each column of the table.
A drawback of this point concerns the identifica-
tion of multi-lines rows, that are composed of sev-
eral lines, that are not correctly recognized in the ta-
ble. Once they are processed by the software they
will have more lines than actually (the content is split
unevenly into different lines). To make sure that the
entire content is collected within a single row, a ref-
erence parameter has been chosen. It must be a value
that exists within each table of interest and for this
reason, it has been chosen as the content of the first
column which represents always an identifying value,
always reported, or a row index, always reported too.
For collecting these reference points the algorithm
create specific intervals for each word/character; these
intervals record the height of the character and repre-
sent the range of integers that start from a lower min-
imum, the ordinate of the character, up to a higher
maximum, the sum of the Y coordinate of the charac-
ter plus the height of the character (in order to map the
interval in which a character is positioned within the
text). Intervals falling under the first column (that has
its own interval range) constitute the content of the
first column for each different row. In this way, ex-
cluding the null values, we have the positional ranges
for each content of the first column for each row and
the algorithm can proceed to the second step.
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
408
In summary, the algorithm computes an interval
for each content that falls in the columns following the
first column, with the same indexes already described,
and the distance between the aforementioned range
and each reference interval. The minimum distance
value represents at which reference range that con-
tent refers. Once this procedure has been carried out
for each reference interval a single line can be con-
structed. It is made up of as many cells as the num-
ber of columns in the table and each cell has its own
content, even if distributed over several lines. Having
done this for each part of interest the content is con-
sidered well structured and can be further processed
to be written into the database.
3.3 Data Reader and Matcher
This module is responsible for the operation of read-
ing the instances already present in the final system.
The model file that has been created by the system on
user entities (see the section below) comes in help for
this purpose. It contains all the properties and rela-
tions and all data types of these properties. Moreover,
it contains the address of each entity that has to be
read from the system and written into the system; it is
necessary to this module and to the writer module. In
addition, in accordance with the parameters that have
been set by the user during the creation of the tem-
plate, some of these data must already be present in
the system. As a result they must be referenced and
not created and conversely, those that are not present
in the database will be created. For the first case, the
software reads the related objects already present in
the system for each entity referenced by the user. For
those entities that must match the software takes and
stores the relative object that match. It will be sent
as input to the writing module. Finally, the match-
ing approach is based on the user-entered properties,
different for each reference entity.
Note that for those objects that provide a match
and consequently must already exist within the final
system, the software will stop the processing of the
current file giving an error in the case the entity is not
matched/found. It will continue with the next file in-
serted by the user in the extraction process, restarting
from the first module (the data collector).
3.4 Data Writer and Translate Modules
The writer module stores data present in the document
on the final system. It receives as input all the refer-
ences to instances already present in the system to be
referenced (via an identification number) and the var-
ious patterns that will be used to create these objects
(template patterns) before writing them. These pat-
terns have been created to allow the user to extract dif-
ferent parts within a single block of text through regu-
lar expressions. The user must specify the name of the
property the related content will refer to. Moreover,
these patterns allow describing the properties of each
column within a table. Then, the rich texts that will
be involved in the analysis are translated through an-
other module. It interfaces with a service exposed by
Yandex dictionary (Yandex, ) to detect text language
and translate it into English (maintenance report in-
clude information in Deutch and other languages). It
has been developed to facilitate the construction of a
common reference dictionary. Finally, structured data
is converted into JSON format and sent to the final
system.
4 CONCLUSIONS
The approach discussed so far shows how textual data
can be crafted and analyzed, starting from mainte-
nance reports, with the application of a set of natural
language processing and ontological techniques. The
obtained results have been satisfactory, regarding the
failure classification, notwithstanding the limited test
dataset. This can be the first step on a true evaluation
step that will be based on a much larger dataset than
this. Its extension will also involve a larger ontologi-
cal model that will be useful to encapsulate other fail-
ure phenomena or other more structural components
within the plant or other components which will fur-
ther enrich the text analysis. A deeper comparison
with other works in the same domain (e.g. (Kusiak
and Li, 2011)) is also planned.
ACKNOWLEDGMENTS
This work has been partially supported by Wind En-
ergy Asset Management System (WEAMS) Project,
endorsed by UE, Italian Ministry of Economic Devel-
opment (MISE) and PON ”Imprese e Competitivita’ -
Iniziativa PMI 2014-20”
REFERENCES
Abichou, B., Fl
´
orez, D., Sayed Mouchaweh, M., Toubakh,
H., Francois, B., and Girard, N. (2014). Fault diag-
nosis methods for wind turbines health monitoring: a
review.
Apple. Siri - apple. https://www.apple.com/siri. Accessed:
Apr 2, 2019.
Power Plants Failure Reports Analysis for Predictive Maintenance
409
Bouti, A. and Kadi, D. a. (1994). A state-of-the-art review
of fmea/fmeca. International Journal of Reliability,
Quality and Safety Engineering, 01(04):515–543.
Carchiolo, V., Longheu, A., and Malgeri, M. (2015). Per-
sonal health record feeding via medical forums. In
2015 IEEE 19th CSCWD Intl. conf., pages 632–636.
Carchiolo, V., Longheu, A., and Malgeri, M. (2015). Using
twitter data and sentiment analysis to study diseases
dynamics. In Proceedings of the 6th ITBAM conf.,
pages 16–24, New York, NY, USA. Springer-Verlag
New York, Inc.
Carchiolo, V., Longheu, A., Malgeri, M., and Mangioni, G.
(2015). Multisource agent-based healthcare data gath-
ering. In 2015 Federated Conference on Computer
Science and Information Systems (FedCSIS), pages
1723–1729.
Council, G. W. E. (2019). Global wind energy council
(gwec) - global wind report. Technical report.
de Azevedo, H. D. M., Ara
´
ujo, A. M., and Bouchonneau, N.
(2016). A review of wind turbine bearing condition
monitoring: State of the art and challenges. Renew-
able and Sustainable Energy Reviews, 56:368 – 379.
Ertek, G., Chi, X., Zhang, A. N., and Asian, S. (2017).
Text mining analysis of wind turbine accidents: An
ontology-based framework. 2017 IEEE Big Data
conf., pages 3233–3241.
Fischer, K., Besnard, F., and Bertling, L. (2012).
Reliability-centered maintenance for wind turbines
based on statistical analysis and practical experience.
IEEE Trans. on Energy Conversion, 27(1):184–195.
Guolin, H., Ding, K., Li, W., and Jiao, X. (2016). A novel
order tracking method for wind turbine planetary gear-
box vibration analysis based on discrete spectrum cor-
rection technique. Renewable Energy, 87:364–375.
Helbing, G. and Ritter, M. (2018). Deep learning for fault
detection in wind turbines. Renewable and Sustain-
able Energy Reviews, 98:189 – 198.
Herbert, G. J., Iniyan, S., and Goic, R. (2010). Performance,
reliability and failure analysis of wind farm in a de-
veloping country. Renewable Energy, 35(12):2739
2751.
Huuhtanen, T. and Jung, A. (2018). Predictive maintenance
of photovoltaic panels via deep learning. In 2018
IEEE Data Science Workshop (DSW), pages 66–70.
IEA. International energy agency - modern bioenergy
forecast. https://www.iea.org/newsroom/news/2018/
october/. Accessed: ”Aug 5, 2019”.
IEA. International energy agency - world energy outlook.
https://www.iea.org/weo/. Accessed: ”Apr 2, 2019”.
K
¨
uc¸
¨
uk, D. and Arslan, Y. (2014). Semi-automatic con-
struction of a domain ontology for wind energy using
wikipedia articles. Renewable Energy, 62:484 – 489.
Kusiak, A. and Li, W. (2011). The prediction and diagnosis
of wind turbine faults. Renewable Energy, 36(1):16 –
23.
Lee, K. and Lee, S. (2013). Patterns of technological in-
novation and evolution in the energy sector: A patent-
based approach. Energy Policy, 59:415 – 432.
Longheu, A., Previti, M., and Mangioni, G. (2016). Tourism
websites network: crawling the italian webspace. In
Proc. of 5th Intl. conf. on Data Analytics, pages 131–
136.
Luong, T. Q. (2015). Traprange: a method to extract table
content in pdf files.
Microsoft. Cortana. your intelligent assistant across your
life. https://www.microsoft.com/en-us/cortana. Ac-
cessed: ”Apr 2, 2019”.
M
´
arquez, F. P. G., Tobias, A. M., P
´
erez, J. M. P., and Pa-
paelias, M. (2012). Condition monitoring of wind tur-
bines: Techniques and methods. Renewable Energy,
46:169 – 178.
Nabati, E. G. and Thoben, K. D. (2017). Big data ana-
lytics in the maintenance of off-shore wind turbines:
A study on data characteristics. In Dynamics in Lo-
gistics, pages 131–140, Cham. Springer International
Publishing.
PDFbox. Pdfbox open source java pdf library. http://www.
pdfbox.org/.
Qiu, Y., Feng, Y., Tavner, P., Richardson, P., Erdos, G., and
Chen, B. (2012). Wind turbine SCADA alarm analysis
for improving reliability. Wind Energy, 15:951–966.
Romero, A., Soua, S., Gan, T.-H., and Wang, B. (2018).
Condition monitoring of a wind turbine drive train
based on its power dependant vibrations. Renewable
Energy, 123:817 – 827.
Selcuk, S. (2017). Predictive maintenance, its implementa-
tion and latest trends. Proceedings of the Institution of
Mechanical Engineers, Part B: Journal of Engineer-
ing Manufacture, 231(9):1670–1679.
Wagner, S. (2016). Natural language processing is no free
lunch. In Perspectives on Data Science for Software
Engineering, pages 175 179. Morgan Kaufmann,
Boston.
Yandex. Yandex online dictionary. https://translate.yandex.
com/. Accessed: Apr 2, 2019.
Zhou, Q., Yan, P., and Xin, Y. (2017). Research on a knowl-
edge modelling methodology for fault diagnosis of
machine tools based on formal semantics. Advanced
Engineering Informatics, 32:92 – 112.
WEBIST 2019 - 15th International Conference on Web Information Systems and Technologies
410