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

net/publication/327558176

A Cloud-based Architecture for Condition Monitoring based on Machine

Learning

Conference Paper · July 2018

DOI: 10.1109/INDIN.2018.8471970

CITATIONS READS
4 1,238

3 authors:

Fernando Arévalo Mochammad Rizky Diprasetya

Fachhochschule Südwestfalen Fachhochschule Soest
12 PUBLICATIONS 20 CITATIONS 4 PUBLICATIONS 9 CITATIONS

SEE PROFILE SEE PROFILE

Andreas Schwung
Fachhochschule Südwestfalen
87 PUBLICATIONS 245 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Recurrent Fuzzy Systems View project

Service Oriented Architectures in automation View project

All content following this page was uploaded by Fernando Arévalo on 12 December 2018.

The user has requested enhancement of the downloaded file.

©2018 IEEE. Personal use of this material is permitted. However, permission
to reprint/republish this material for advertising or promotional purposes or
for creating new collective works for resale or redistribution to servers or lists,
or to reuse any copyrighted component of this work in other works must be
obtained from the IEEE.

Appeared as: F. Arévalo, M. R. Diprasetya and A. Schwung, „A Cloud-based

Architecture for Condition Monitoring based on Machine Learning“, 2018 IEEE
16th International Conference on Industrial Informatics (INDIN), Porto, 2018, pp.
163-168.

DOI: 10.1109/INDIN.2018.8471970
 A Cloud-based Architecture for Condition
Monitoring based on Machine Learning
Fernando Arévalo∗ , Mochammad Rizky Diprasetya† , Andreas Schwung‡
Department of Automation Technology
South Westphalia University of Applied Sciences, Germany
Email: {arevalo.fernando,schwung.andreas}@fh-swf.de, [email protected]

Abstract—In the framework of the Digitalization of the In- This paper presents a cloud-based architecture for con-
dustry, there is an increasing trend to use machine learning ditioning monitoring based on Machine Learning. A Web
techniques to assess condition monitoring, fault detection, and application presents the results to the end-user. This research
process optimization. Traditional approaches use a local IT
framework centralized in a server, in order to provide these is divided in two parts: the front-end development and the
services. Cost of equipment and IT manpower are associated to condition monitoring based on machine learning, see section
the implementation of a condition monitoring based on machine IV and section V respectively. The front-end development
learning. Nowadays, cloud computing allows to replace local IT uses a IaaS cloud computing model for the web application.
frameworks by a remote service, which can be paid according The back-end is based on an IT framework, which comprises
to the customer needs. This paper proposes a cloud-based
architecture for condition monitoring based on machine learning, the aforementioned services. This back-end is implemented
which can assess the end-user through a web application. The in the cloud service. The condition monitoring is developed
condition monitoring is implemented using a fusion of classifica- using machine learning techniques, specifically classification
tion methods. The fusion is implemented using Dempster-Shafer methods. An information fusion approach is used to improve
Evidence Theory (DSET). The results show that the use of DSET the performance of the condition monitoring system.
improves the overall result.
Index Terms—Cloud-based Architecture, Condition Monitor- II. R ELATED W ORK
ing, Machine Learning, IaaS, MQTT, OPC-UA, DSET
The Industry 4.0 refers to the implementation of systems
I. I NTRODUCTION with the concept of the internet of things. Ferrer et al. explains
Nowadays, an Industrial PC (IPC) is used as a monitor- about web-based IoT devices and cloud based manufacturing
ing system, which possesses a data analytics software for platform [5]. To achieve the internet of things concept, an
the industrial process. This IPC could be a dedicated PC, Information and Communication Technologies (ICT) frame-
which is separated from the automation system. When a data work is implemented. ICT helps machines, sensors, embedded
analytics software uses a complex algorithm, which needs devices and software components to be connected with each
more resources, a powerful IPC is needed [1]. The more other, in order to use the information for monitoring pur-
powerful the IPC is, the more expensive it is. Often, a data poses. Cloud Collaborative Manufacturing Networks (C2NET)
analytics software relies on additional services such as access is implemented for the cloud-based platform. The C2NET
to technical documentation, databases, connectivity to mobile architecture uses WebSockets, MQTT and RESTful API for
applications and to the automation platform of the factory the connection between the IoT devices and the cloud-based
[2]. These services are normally contained in an local IT platform.
Infrastructure. The up-front cost for an IT infrastructure can There is an increasing trend in the industry to include
be high e.g. a suitable room, equipment cost, IT manpower, cloud computing in its data analysis. Elazab et al. designed
maintenance. Moreover, a user who does not have experience a condition monitoring system for power plants [6]. The
in setting up an IT infrastructure, requires more time to have application was based on an IaaS model, a LAN connection
a running system [3] and technical training. for the data collection, and in the concept of Cloud Based
There is an increased use of cloud computing as an alter- Condition Monitoring (CBCM).
native for a local IT infrastructure. A cloud service enables a OpenStack is a development toolkit for managing a cloud
user to share and access her system with minimal management computing platform. Zhen-wu et al. developed a cloud infras-
effort, often over the internet. It allows the user to minimize tructure for industrial applications based on OpenStack [7]. An
the up-front IT infrastructure cost and the maintenance cost. automation platform is integrated with the cloud infrastructure.
An Infrastructure as a Service (IaaS) can be used for a user, The automation platform uses Siemens PLC’s.
who wants to implement her system in the cloud [4]. The user There is a term in industrial system applications, which a
does not have to configure the IT infrastructure, since the cloud machine communicates with another machine using different
provider offers it. Cloud computing is a suitable solution for architectures. It is called Machine to Machine (M2M) commu-
small and middle size companies, since there is no need to nication. One proper way to establish an M2M communication
invest in IT equipment, neither to maintain it. is through IoT. Domenech et al. designed a smart industrial
environment by using IoT as Web of Things [8]. The M2M C. Cloud Computing
communication is implemented through a RESTful API over
There are three basic models in cloud computing [4]:
HTTP.
1) Infrastructure as a Service: The cloud provider offers
III. A C LOUD - BASED A RCHITECTURE FOR C ONDITION hardware, software, server, security, storage, and maintenance.
M ONITORING IN THE F RAMEWORK OF THE C ORE S YS A computer software, such as Oracle Virtualbox, VMware,
P ROJECT or Hyper-V, runs a virtual machine as a guest in a server.
One server can run multiple virtual machines. It depends
The completion of this research comprises the understand- on the server capability, in order to maintain the stability
ing of different topics such as software framework, cloud while running the virtual machine. There are many providers
computing, communication protocols and development of end- that offer IaaS such as Google Cloud, Amazon AWS, and
user applications. DigitalOcean.
2) Software as a Service: The cloud provider offers only
A. The CoreSys Framework
the software. The software is developed in the cloud infras-
The CoreSys framework is designed to include different IT tructure. The user can only access the software.
services for supporting the end-user, see Fig. 1. CoreSys is 3) Platform as a Service: The cloud provider offers to
designed to work as a local IT Infrastructure. The main mod- the user: a development toolkit, operating system, web server
ules of CoreSys are: Machine Learning and Data Analytics and database. The user can only access the development
(MLDA), the Apache Server, Historical Data, and Knowledge environment. She cannot access the whole infrastructure of
Modeling. The purpose of MLDA is to provide functions the cloud.
for fault detection assessment and condition monitoring. The
historical data serves for training and testing models, as well D. MQTT Communication Protocol
as to run data analytics. CoreSys collects data from the
automation platform through data communication protocols. Nowadays, a web-based internet of things is used in In-
dustry for monitoring, analyzing, and tracking. One of the
known communication protocols, which is used for web-based
internet of things is Message Queue Telemetry Transport
(MQTT) [9]. MQTT is a simple publish\subscribe commu-
nication protocol that uses a message broker. The broker
listens to all clients that are subscribed to it. A client can
be either a publisher or a subscriber, or both. Publisher of the
client publishes a message to the broker with a certain topic
(broker label). When the broker updates a topic, the subscriber
retrieves the message from the topic.

IV. F RONT-E ND D EVELOPMENT

The development of this Front-End considers a Waterfall
Fig. 1. CoreSys Project Software Development model [10]. The main stages for the
software development are: planning, definition of software re-
CoreSys was designed initally to run in an IPC, as a local quirements, design, implementation, testing and maintenance.
system. Nevertheless, given the increasing trend in the industry This research includes only the software requirements, design,
to use cloud computing, a cloud framework CoreSys-Cloud is implementation and testing stages.
proposed. Due the fact that CoreSys is built on Python, a cloud
service provider that contains Python was chosen. The web A. Software Requirements
application allows the end-user to analyze data using machine
learning techniques, which can be visualized through a PC or This section comprises the software specifications, which
a smartphone through the web service. All the modules of the defines the functionality of the application. The Front-End
CoreSys-Cloud are implemented directly in the cloud server. requirements are summarized as follows:
• MQTT data communication protocol as an interface be-
B. Condition Monitoring based on Machine Learning tween the the automation platform and the Back-End.
The condition monitoring is implemented in the MLDA • MQTT data communication protocol shall be used be-
module of CoreSys-Cloud using machine learning techniques. tween the automation platform and CoreSys-Cloud.
The condition monitoring uses classification methods to pre- • Access to results of the data analytic and machine learn-
dict the machine status based on its input variables. Informa- ing module.
tion fusion is used to combine the classification methods, in • Access to current data of the automation platform.
order to improve the overall performance. • Implement the Front-End using a Web Design.
B. Design
The design is specified using UML diagrams and the
transition of the pages diagram. These diagrams are based on
the information recolected in the software requirements.
The transition of the pages is shown in Fig. 2. Two pages
are developed in this project: homepage and automation page.
Fig. 4. Back-End Sequence Diagram
The homepage consists of a short introduction about the author
and a short description about the project. The homepage has a
navigation bar, which has two navigation menu: Home and main components are: the automation platform, the communi-
Automation. It is located on the header of the page. The cation protocol MQTT, the framework CoreSys-Cloud, a OPC-
automation page consists of a figure, which displays the result UA/MQTT gateway, and the Web Application. The gateway
of the condition monitoring assessment. enables the communication between the OPC Server of the
automation platform and the MQTT broker in CoreSys-Cloud.
MQTT establishes the communication between CoreSys-
Cloud and the web application.
Fig. 2. Transition of the Front-End Screens

The UML diagrams employed are the sequence diagrams,

which are the front-end sequence and back-end sequence
diagrams. The front-end sequence diagram is shown in Fig.
3. When the user accesses the web application, the homepage
will be displayed. The web application requests the home Fig. 5. CoreSys Cloud and Automation Platform
page data from the web framework Flask. Flask retrieves the
data from a MySQL database. Flask provides the data to The software implementation is divided in four steps, see
the homepage. The same scenario is applied when the user Fig. 6.The steps are: choosing the cloud provider, implemen-
accesses the automation page. There is a recurrent function tation of the OPC-UA/MQTT gateway, development of web
in the automation page, which implements this process in a application using Flask, and definition of a function for the
predefined time. web application.

Fig. 6. The Implementation Steps

In the first step, the cloud provider DigitalOcean is chosen.

DigitalOcean is a well-known cloud provider along with
Google Cloud and Amazon AWS. It has a simple cloud
configuration. It automatically creates the virtual server with
predefined specifications offered by DigitalOcean. The user
chooses the specification that suits her nee for her application.
After DigitalOcean creates the virtual server, the IP address
Fig. 3. Front-End Sequence Diagram
of the virtual server is given to the user. The user can access
the virtual server through a Secure Shell (SSH) session.
The back-end sequence diagram is shown in Fig. 4. This In this project, Ubuntu is installed in a virtual server. Linux
diagram describes the data collection between CoreSys-cloud environment is used for the project development. There are
and the automation platform. In the first step, the automation four services implemented in CoreSys-Cloud:
platform sends data to the machine learning and data analytics 1) Machine Learning and Data Analythics: The machine
(MLDA) module through MQTT. The next step, MLDA ana- learning and data analytics (MLDA) module is used for the
lyzes the data for the condition monitoring. Then the result is condition monitoring assessment. MLDA is written in Python.
sent to Flask. Flask stores the result in MySQL. This proccess The library scikit-learn [11] is used for the implementation of
is iterated in a loop, so the data will be continuously updated the machine learning functions of MLDA.
in MySQL. 2) Flask: Flask is a software framework that is used for the
web application development. The reasons for choosing Flask
C. Implementation are its lightweight framework, and it is a Python library. This
The architecture of the integration between the CoreSys allows an easier integration of Flask to the MLDA, since both
Cloud and the automation platform is shown in Fig. 5. The are written in Python.
 3) MySQL Database Server: The database is used to store methods are Random Forest, Multilayer Perceptron Artifitial
the result of the CM assessment. Neural Network, and AdaBoost. The classification methods
4) MQTT Server: MQTT server is used to establish a were chosen due to well performance and wide use in research
communication between the OPC-UA/MQTT gateway and [12] [13] [14]. Dempster-Shafer is chosen as an information
CoreSys-Cloud. The MQTT server communicates with Flask. fusion method, due its capability to combine different infor-
The server provides an authentication method using username mation sources [15]. A summary of each method is provided,
and password. This prevents an unknown connection to the for a deep mathematical understanding refer to [13] [14].
MQTT server. 1) Random Forest: Random Forest method combines ran-
The second step is the implementation of an OPC- domize decision trees, and accumulates their result by averag-
UA/MQTT gateway. The gateway enables the communication ing and voting methods. Each decision tree uses a part of the
between the OPC server of the automation platform and the features to generate a model of classification [16].
CoreSys-Cloud MQTT broker. The gateway communicates 2) Multilayer Perceptron Artificial Neural Network: Nor-
with the automation platform through OPC-UA. mally, a Multilayer Perceptron (MLP) network consists of
The third step is the development of a web application using three layers of nodes: input layer, hidden layer, and output
Flask. Flask serves as an interface between the MQTT broker layer. Each layer uses a non-linear activation function. The
of CoreSys-Cloud and the MySQL database. The websites backpropagation algorithm is used to train the network [17].
Homepage and Automation Page interact with Flask, in order 3) Adaptative Boosting: Adaptative Boosting (AdaBoost)
to request or retrieve data. combines many types of learning algorithms to improve the
The fourth step is to create functions for the web ap- overall performance. Every learning algorithm has its own
plication. The web application obtains data from the OPC- parameters and configurations. It needs to be adjusted to
UA/MQTT gateway through MQTT. A function is called achieve an optimal performance. The output is combined by
whenever the gateway publishes data to the MQTT bro- using a weighted sum of each algorithm [18].
ker. This function stores the published data to the MySQL 4) Dempster-Shafer Rule of Combination: Dempster Shafer
database. A recurrent function based on JQuery is used in the Evidence Theory (DSET) is an information fusion method,
automation page, in order to update continuously the results of which allows to model uncertainty and to combine knowledge
MLDA module. This recurrent function calls a GET method of different sources. The knowledge is defined as a mass
in the web application through REST architecture. The GET function. Each mass function is defined as a set of evidence.
method retrieves data from MySQL database, and sends the The DSET rule of combination combines the mass functions
data to the recurrent function. The recurrent function sends of the different knowledge sources [15].
the data to the website to be displayed. The database can
B. Fusion of Classification Methods using DSET
be only accessed through the GET method. Fig. 7 shows
the visualization of the Automation Page customized for the The prediction of a classification method is transformed into
application on a Bulk Good System. a set of evidence using:
(1 − K)
mprediction = [K ∗ w ∗ w ...
n−1 (1)
(1 − K)
∗ w (1 − w)]
n−1
k = 1 − 10F (2)
where k is the sensitivity to zero [15] factor, F is the
approximation factor, w is the confidence weight of each
method and n is the number of target classes.
C. Methodology
Fig. 7. Automation Page MLDA module collects the machine data, trains the classi-
fication models, and feeds up current data into the models
V. C ONDITION M ONITORING BASED ON M ACHINE to obtain a prediction of the target class. MLDA uses the
L EARNING predicted target class for the machine condition assessment.
This section describes the methodology for the implemen- The methodology for the implementation of the MLDA
tation of the Condition Monitoring using the machine learning follows the steps: Data collection, data preparation, method
and data analytics (MLDA) module. A theoretical background selection and model training, and model testing.
for the machine learning methods is provided. The Data Collection is divided in two types. The first
type collects the data once or in a regular basis, which is
A. Preliminary stored as historical data. MLDA uses the data for training,
MLDA has its backbone on classification methods, as well validation and testing the models. The second type collects
as information fusion techniques. The selected classification data continuously and feeds it up into the trained models, in
order to get the predictions. The Data Preparation consists in TABLE I
splitting the data into training, validation and testing data. A F EATURES AND C LASSES OF THE BGS.
two-step split strategy is followed. The first step divides the
Storage,
data into validation(30%) and model data(70%). The second Features Loading Filling
Weighing
step divides the model data into training(70%) and testing Filling height min. connected x
x
data(30%).The data is scaled to improve the performance. The Filling height max. connected x x
data features are normalized using a normalize function, which Overflow connected x x
scales the features into a number between zero and one [14]. Emergency stop active x x x
Normal speed active x
The Method Selection is assessed through a cross valida- Normal pressure active x x
tion method. This method separates the training data into a Normal suction time active x
predifined number of folds [13]. Each fold is divided into Next station connection active x
Container active x
training and validation data. The purpose of this method is
to use all the data for training and validation of the models. Storage,
Classes Loading Filling
Weighing
The method provides a performance score for each fold. An
Machine condition normal x x x
average of all the performace of the folds is taken. This average
Machine condition warning x x x
for each classification methods serves as a metric to evaluate Machine condition critical x x x
the performance of each method. This average is used as a
confidence weight in the DSET fusion.
TABLE II
The Model Testing involves changing on parameters, and C ROSS VALIDATION OF C LASSIFICATION M ETHODS USING S TATION DATA
checking the performance of the methods. This includes testing
all the classification methods individually, as well as the DSET Method Loading Storage Weighing Filling
fusion of the classification methods. Decision Tree 99.92 99.71 99.71 99.58
Random Forest 99.92 99.71 99.71 99.62
VI. U SE C ASE : A C LOUD - BASED A RCHITECTURE FOR ANN MLP 99.03 97.02 96.97 97.18
C ONDITION M ONITORING ON A B ULK G OOD S YSTEM AdaBoost 75.60 97.61 96.50 96.76
Naive Bayes 99.79 95.20 87.22 94.95
The Bulk Good System (BGS) Laboratory plant is a bulk
good plant in a small scale [19]. It is a machine, which
contains four stations namely loading, storage, weighing and Table III shows the performance of each method and DSET
filling. Each station can operate in standalone modus, or in a fusion for each station. The station performance is divided
configuration of at least two stations. The automation platform per class. The performance indicator is built using a confusion
of the BGS possesses a controller on each station, specifically matrix, which compares the class prediction with the test class.
an Industrial PC for the loading station and a PLC for the rest The station Loading shows a total performance above 83.99%.
of the stations. Profinet serves as a communication protocol The class critical using Multilayer Perceptron presents the
between the stations. lowest result with 52.25%. The rest of the results are pre-
MLDA collects the data from the BGS, train the models, dicted with an accuracy above of 99.31%. The station Storage
and feeds up current BGS data into the models to obtain a presents a total performance above 71.31%. The lowest result
prediction of the target classes (Normal, Warning, Critical) 18.05% belongs to the class critical using MLP. The rest of
of the machine. The machine assessment is performed then the results remains homogeneous with values above 95.88%.
using current features to obtain the machine status based on The station Weighing has a total performance above 84.78%.
the target class. The lowest result for the class critical is 71.72% using MLP.
The lowest results for the warning class belongs to MLP and
A. Experimental Data Analysis and Results
AdaBoost with values of 82.74% and 86.76% respectively.
The Data Collection from the BGS extracted data in a lapse The rest of the results presents values above 99.76%. The
of 2 hours using a sample rate of 1s. This data is used as station Filling displays a total performance above 87.80%.
historical data, in order to perform the training, validation and The lowest result belongs to the class warning with a value
testing of the models. Table I displays the features and classes of 63.91% using AdaBoost. The rest of the results remains
of the BGS data, for more details regarding the dataset refer homogeneous with values above of 94.32%. DSET fusion
to [20]. presents the same results of the method Random Forest for
The Data Preparation consists into splitting the data into the stations Loading, Storage and Filling with values above
training, validation and testing data. The features are nor- 99.83%. The station Weighing presents a slightly difference
malized. MLP presents better results using standard data. between the two approaches.
Nevertheless, a normalized data was chosen for MLP, due
to the information fusion can be more evidently shown. The B. Discussion
Method Selection is performed through a cross validation The results were improved using DSET rule of combination,
method using 10 folds. The average of the cross validation this relies on the combination of set of evidences with low
is used as a confidence weight for each method, see Table II. uncertainty. The uncertainty was based on the cross validation
 TABLE III [4] M. Moravcik, P. Segec, J. Papan, and J. Hrabovsky, “Overview of cloud
P ERFORMANCE OF C LASSIFICATION M ETHODS PER S TATION computing and portability problems,” in ICETA 2017. [Piscataway,
New Jersey]: IEEE, 2017, pp. 1–6.
Method Normal Warning Critical Total [5] B. R. Ferrer, W. M. Mohammed, E. Chen, and J. L. M. Lastra,
“Connecting web-based iot devices to a cloud-based manufacturing
Station Loading platform,” in IECON 2017 - 43rd Annual Conference of the IEEE
Industrial Electronics Society. IEEE, 10/29/2017 - 11/1/2017, pp. 8628–
Random Forest 100.00 99.86 100.00 99.95
8633.
ANN MLP 100.00 99.72 52.25 83.99
[6] E. Elazab, T. Awad, H. Elgamal, and B. Elsouhily, “A cloud based
AdaBoost 100.00 99.31 100.00 99.77
condition monitoring system for industrial machinery with application
DSET Fusion 100.00 99.86 100.00 99.95
to power plants,” in 2017 Nineteenth International Middle East Power
Station Storage Systems Conference (MEPCON). IEEE, 2017, pp. 1400–1405.
[7] L. Zhen-wu, Z. De-min, S. Yun-tao, and S. De-hui, “Research on the
Random Forest 100.00 100.00 100.00 100.00 architecture and application of industrial cloud experimental platform
ANN MLP 100.00 95.88 18.05 71.31 based on openstack,” in Proceedings of the the 29th Chinese Control
AdaBoost 100.00 100.00 100.00 100.00 and Decision Conference (2017CCDC). Singapore: IEEE Industrial
DSET Fusion 100.00 100.00 100.00 100.00 Electronics (IE) Chapter, 2017, pp. 7401–7406.
Station Weighing [8] M. C. Domenech, L. P. Rauta, M. D. Lopes, P. H. Da Silva, R. C. Da
Silva, B. W. Mezger, and M. S. Wangham, “Providing a smart industrial
Random Forest 99.88 99.76 100.00 99.88 environment with the web of things and cloud computing,” in 2016 IEEE
ANN MLP 99.88 82.74 71.72 84.78 International Conference on Services Computing, J. Zhang and J. A.
AdaBoost 99.88 86.76 100.00 95.55 Miller, Eds. Los Alamitos, California: IEEE Computer Society, 2016,
DSET Fusion 99.88 99.76 100.00 99.83 pp. 641–648.
Station Filling [9] K. Grgic, I. Speh, and I. Hedi, “A web-based iot solution for moni-
toring data using mqtt protocol,” in Proceedings of 2016 International
Random Forest 99.48 100.00 100.00 99.83 Conference on Smart Systems and Technologies (SST), D. Žagar, Ed.
ANN MLP 99.61 96.03 94.32 96.65 Osijek, Croatia: Faculty of Electrical Engineering, Computer Science,
AdaBoost 99.48 63.91 100.00 87.80 and Information Technology Osijek, 2016, pp. 249–253.
DSET Fusion 99.48 100.00 100.00 99.83 [10] E. J. Braude and M. E. Bernstein, Software engineering: Modern
approaches, 2nd ed. Hoboken, NJ: Wiley, 2011.
[11] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vander-
results, as a basis to determine the confidence of each method. plas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duch-
The confidence of each method has values above 96.50%, esnay, “Scikit-learn: Machine learning in python,” Journal of Machine
Learning Research, vol. 12, pp. 2825–2830, 2011.
with the difference of AdaBoost with a value of 75.60% for [12] Abraham J. Wyner, Matthew R Olson, Justin Bleich, and David Mease,
the station Loading. The results of the fusion are equal or “Explaining the success of adaboost and random forests as interpolating
better than the other classification methods, except for slightly classifiers,” Journal of Machine Learning Research, vol. 18, pp. 48:1–
48:33, 2017.
differences in the performance of the weighing station. The [13] K. P. Murphy, Machine learning: A probabilistic perspective, ser.
results are attributed to the nature of the dataset, since there Adaptive computation and machine learning series. Cambridge, Mas-
are only 5 features per station, and it does not possess complex sachusetts and London, England: The MIT Press, 2012.
[14] I. H. Witten, C. J. Pal, E. Frank, and M. A. Hall, Data mining:
relationships between the variables. Practical machine learning tools and techniques, fourth edition ed.
Cambridge, MA: Morgan Kaufmann, 2017. [Online]. Available:
VII. C ONCLUSION http://proquest.tech.safaribooksonline.de/9780128043578
This paper presented a cloud-based architecture for condi- [15] S. R. Mousavi, Dempster-shafer theory and modified rules to determine
uncertainty in mineral prospection. Clausthal-Zellerfeld: Univer-
tion monitoring based on machine learning. CoreSys-Cloud sitätsbibliothek Clausthal, 2011.
proves to be a cost effective alternative for small and middle [16] C. Hao, S. Sivanesan, M. Majmudar, and K. S. Rajput, “Combinational
size companies, in order to have access to machine learning feature based random forest classification for enhanced bundle branch
block beat detection,” in 2018 IEEE EMBS International Conference on
techniques and data analytics in a software framework. The Biomedical & Health Informatics (BHI). IEEE, 3/4/2018 - 3/7/2018,
front-end development section provides with a methodology pp. 319–322.
for the implementation of a web application. The condition [17] A. Thakur and D. Mishra, “Hyper spectral image classification using
multilayer perceptron neural network & functional link ann,” in Pro-
monitoring section provides with a methodology for the imple- ceedings of the 7th International Conference Confluence 2017 on Cloud
mentation of a fusion of classification methods. DSET fusion Computing, Data Science and Engineering, A. Bansal and A. Singhal,
proved to be a suitable alternative to combine the predictions Eds. Piscataway, NJ: IEEE, 2017, pp. 639–642.
[18] K. Randhawa, C. K. Loo, M. Seera, C. P. Lim, and A. K. Nandi, “Credit
of different classification methods. card fraud detection using adaboost and majority voting,” IEEE Access,
vol. 6, pp. 14 277–14 284, 2018.
R EFERENCES [19] Philip van Bergerem, “Concept development and installation of a modu-
[1] A. Matsunaga and J. A. Fortes, “On the use of machine learning to lar bulk good laboratory plant and implementation of a self-reconfiguring
predict the time and resources consumed by applications,” in 10th control system,” Master Thesis, South Westphalia University of Applied
IEEE/ACM International Conference on Cluster, Cloud and Grid Com- Sciences, Soest, 2016.
puting (CCGrid), 2010, M. Parashar, Ed. Piscataway, NJ: IEEE, 2010, [20] F. Arevalo, T. Mohammed, and A. Schwung, “Fault detection using
pp. 495–504. probabilistic prediction and data fusion on a bulk good system,” in 2017
[2] L. Wang and R. X. Gao, Condition monitoring and control for intelligent 52nd International Universities Power Engineering Conference (UPEC).
manufacturing, ser. Springer series in advanced manufacturing. London: Piscataway, NJ: IEEE, 2017, pp. 1–6.
Springer, 2006.
[3] T. Hegazy and M. Hefeeda, “Industrial automation as a cloud service,”
IEEE Transactions on Parallel and Distributed Systems, vol. 26, no. 10,
pp. 2750–2763, 2015.

View publication stats