Computers in Industry 134 (2022) 103555

Contents lists available at ScienceDirect

Computers in Industry
journal homepage: www.elsevier.com/locate/compind

Cognitive analytics platform with AI solutions for anomaly detection

⁎
Vaia Rousopoulou, Thanasis Vafeiadis , Alexandros Nizamis, Ioannis Iakovidis,
]]
]]]]]]
]]

Leonidas Samaras, Alkis Kirtsoglou, Konstantinos Georgiadis, Dimosthenis Ioannidis,

Dimitrios Tzovaras
CERTH/ITI (Center for Research and Technology Hellas, Information Technologies Institute), 57001 Thessaloniki, Greece

a r t i cl e i nfo a bstr ac t

Article history: This work presents a cognitive analytics platform for anomaly detection which is capable to handle, analyze
Received 4 June 2021 and exploit resourcefully machine data from a shop-floor of factory, so as to support the emerging and
Received in revised form 13 September 2021 growing needs of manufacturing industry. The introduced system contributes to industrial digitization and
Accepted 5 October 2021
creation of smart factories by providing a generic platform which is a complete solution supporting stan­
Available online 9 November 2021
dards-based factory connectivity, data management, various AI models training and comparisons, live
predictions and real-time visualizations. The proposed system is built on a micro-service architecture, in
Keywords:
Cognitive platform order to be extendable and adaptive over time, and contains three core modules, the Data Acquisition, the
Predictive maintenance Knowledge Management and the Predictive maintenance, which contribute to machine failure prediction and
Failure prediction activate predictive maintenance procedures, to efficient production schemes and decision making, to
Decision making monitor anomalies and handle unforeseen conditions, to predict future behaviours on time series etc. The
Advanced visualization techniques proposed platform utilizes continuous re-training mechanisms enabling a self learning approach for the
Industry 4.0 delivery of AI solutions, usable also for various production data, guaranteeing the quality of results without
continuous monitoring and human-resources allocation for AI models’ retraining. This cognitive platform is
supported by machine learning techniques and deep learning architectures in order to achieve the desired
performance in the management of factory processes and needs. All the information generated by the
proposed platform is provided to the end user through a user interface that utilizes advanced visualization
techniques.
© 2021 The Authors. Published by Elsevier B.V.
CC_BY_NC_ND_4.0

1. Introduction Gutzmer very accurately captured the current situation in Industry

4.0 regarding the moves that need to be made in order to evolve
In recent years, manufacturing industry becomes larger, more further: “Industry 4.0 is not the digitalization of the mechanical
complex and is accelerated by several growing technologies as industry, because that is already here. It is about getting real-time
customers demands on product quality and reduction of product data into the supply and manufacturing chain. If we use this IoT data
delivery time are steadily increasing. On the other hand, raw ma­ and combine it in a different way, we can be more flexible. We can
terial and components prices are increasing in production markets, adapt faster.”
competition is rising constantly and the most significant of all, shop- There is a vast amount of literature on predictive maintenance,
floor or factory data from industrial equipment, automation systems smart manufacturing and cognitive applications. Decision making
and sensors installed in production lines and more other places, are and real time monitoring are an integral part of smart manu­
growing exponentially. In order to succeed in the era of Industry 4.0, facturing and empower predictive maintenance operations in
it is vital and for manufacturers to work with and dig in historical Industry 4.0. Machine learning methods are integrated in predictive
and real-time data, along with unstructured data that needs serious maintenance implementations and contribute to innovative systems.
pre-processing, to be able to make decisions that balance product The most common machine learning techniques used in industry are
and work quality, efficiency of industrial equipment and sustain­ investigated in (Cinar et al., 2020). This particular work highlights
ability of shop-floor or factory. A few years ago, professor Peter the industrial areas which use machine learning applications in
order to improve efficiency of the machines, it demonstrates the
challenges of such implementations and the most used machine
⁎
Corresponding autho: learning methods, it identifies the data types used with special
E-mail address: [email protected] (T. Vafeiadis).

https://doi.org/10.1016/j.compind.2021.103555
0166-3615/© 2021 The Authors. Published by Elsevier B.V.
CC_BY_NC_ND_4.0
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

mention to real time data and finally indicates limitations but also supervised analysis which cannot guarantee the high accuracy of the
proposes some future steps on the predictive maintenance field. predictions indefinitely. There is no update or evaluation mechanism
Indisputably, the methodologies applied for a predictive main­ in this solution, and since the training is implemented offline there is
tenance venture depend on the industry needs and also the mon­ a lack of adaptability in changes over time.
itoring equipment (IoT devices) and available data (historical and Several European Commission projects address the problems and
real time) (Vafeiadis et al., 2018). This topic is widely studied in challenges of the 4th industrial revolution by introducing innovative
(Carvalho et al., 2019) based on exploring main papers of predictive digital analytics platforms utilizing AI solutions. The Z-Bre4k
maintenance through machine learning. This study points out the (Strategies and Predictive Maintenance models wrapped around
feasible and promising characteristics of predictive maintenance and physical systems for Zero-unexpected-Breakdowns and increased
reveals the linkage between any specific machine learning approach operating life of Factories)1 project proposes a novel predictive
with the available equipment and sensor technologies. Finally it maintenance platform to eliminate unexpected-breakdowns and
proposes the use of the development of sensor techniques, the extend the life of production systems. A modular Smart Maintenance
comparison of machine learning methods and even use of deep Platform encapsulates known from scientific literature machine
learning techniques, as well as ensemble learning for future re­ learning approaches is used to provide a day-to-day assisting tool for
searchers. the detection and the prediction of the failures of the machinery
The amount of data coming from factories has led to smart equipment to the maintenance engineers (Naskos et al., 2021). Smart
manufacturing including process monitoring, analytics and high- Maintenance Platform offers the micro-services: Data Provider, Data
level decision making. Data-driven solutions and applications bring Consumer, Fault Detection, Failure Prediction, Fusion and Reporter.
research and analytical methods to address real-world problems and The work in (Anaya et al., 2018) presents the vf-OS (Virtual Factory
increase efficiency and decrease costs. Machine learning resolves Open Operating System)2 project, an approach to IoT Analytics and
active applications such as process and sensor monitoring,but also how vf-OS IO (Input/Output) components and vf-OS analytics can be
active applications such as control optimization and decision making used to capture data from sensors and Program Logic Controllers
(Shang and You, 2019; Vafeiadis et al., 2019). The automation of data (PLCs) to generate and run machine learning models to do analytics
collection and analysis and the intelligence exploitation are the in the cloud and in the edge. In the same direction, the recently
major challenge addressed by data-driven smart manufacturing (Tao launched project named knowlEdge3 aims to deliver a platform
et al., 2018). Sergio et al. (2021) that offers AI powered manufacturing services,
Smart manufacturing is further augmented by the incorporation processes, and products in an edge-to-cloud-knowledge continuum
of deep learning techniques. Deep networks can easily handle the for humans in-the-loop. The EFPF (European Connected Factory
characteristics of Big Data coming from manufacturing. The ad­ Platform for Agile Manufacturing)4 project is a federated smart
vantages of deep learning in terms of computational intelligence and factory ecosystem and a digital platform that interlinks different
potential are the key to upgrade smart processes,machine opera­ stakeholders of the digital manufacturing domain. The EFPF platform
tions, early fault detection, decision support and product quality enables users to utilise innovative functionalities, experiment with
insurance. The work of (Wang et al., 2018) thoroughly investigates disruptive approaches and develop custom solutions to maximise
the advances of deep learning in predictive maintenance and deci­ connectivity, interoperability and efficiency across the supply chains.
sion making of smart manufacturing applications and proves the The EFPF platform introduces an Anomaly Detection Service that
power of transforming the unprecedented amount of data into ac­ allows users to create, train and initiate the trained models in real-
tionable and insightful solutions. time data streams. The COMPOSITION (Ecosystem for collaborative
Various complete and integrated solution towards predictive manufacturing processes)5 project has developed an integrated in­
maintenance propose predictive maintenance platforms with mul­ formation management system (IIMS)Vafeiadis et al. (2019) which
tiple features and functionalities. A work of this category is optimises the internal production processes by exploiting existing
(Efthymiou et al., 2012), which proposes an integrated predictive data, knowledge and tools to increase productivity and dynamically
maintenance platform with data acquisition and knowledge man­ adapt to changing market requirements. The project developed an
agement resulting in a sustainability maintenance dashboard. This ecosystem supporting the interchange of data and services between
paper presents a review on the predictive maintenance approaches, factories and their suppliers, thus facilitating the entry of new
methods and tools in manufacturing systems and proposes an in­ market actors into the supply chain. Other research projects have
tegrated predictive maintenance platform. The knowledge manage­ promoted to large initiatives such the International Data Space As­
ment and fault diagnosis part of this research is based on semantics sociation6 (IDSA) that promotes a reference architecture for stan­
and an inference machine. The diagnosis of failures is implemented dardized, secure and sovereign data exchange for building data
with the use of stored rules and reasoning. This is a decent fault spaces able to realise various use cases including industrial sector.
diagnosis proposition, but still it is a rigid, non-expendable solution Analytic tools as apps and services can be applied over the data
which does not incorporate full exploitation of data. space digital infrastructure and by consuming data using the avail­
Another work in the same context is introduced in (Li et al., able data connectors. In the same direction, the Gaia-X7 project
2017). This work proposes a predictive maintenance framework proposes a data infrastructure for building an open, transparent and
based on Industry 4.0 concepts and machine centres. The framework secure digital ecosystem, where data and services can be made
includes sensor selection and data acquisition, data pre-processing, available, and shared in an environment of trust. Similar to IDSA,
data mining, decision support, and maintenance implementation. Gaia-X provides an association, hubs and a community.
The system provides a wide variety of methods inside the available The motivation behind this study is to present a cognitive ana­
modules but it has a high dependency with the use case that is being lytics platform that combines all those elements necessary for data
applied to and also there are points in which the different modules
do not act automatically as a complete pipeline.
Finally, (Canizo et al., 2017) is a data-driven solution for wind 1
https://www.z-bre4k.eu/
2
turbines failures prediction. The solution entails predictive model https://www.vf-os.eu/
3
analytics using Random Forests, real time monitoring for wind tur­ https://www.knowledge-project.eu/
4
https://www.efpf.org/
bines state prediction and real-time visualization. The system 5
https://www.composition-project.eu/
achieves to predict and visualize the state of the specific machines, 6
https://internationaldataspaces.org/
but its main limitation is that the system provides a specific 7
https://www.data-infrastructure.eu/GAIAX/Navigation/EN/Home/home.html

2
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 1. Cognitive analytics platform capabilities.

analysis and efficient decision making. Thus the proposed platform, installing it in a cloud server needs purchasing of some MS Azure or
has the ability, firstly, to connect with the shop-floor or factory, re­ Amazon cloud services. In addition to aforementioned platform, SAP
trieve data from industrial equipment through IoT devices and to provides another ready to market solution its SAP Analytics Cloud.10
move on with pre-processing, for the modules and methods to Data exploration, visualizations, BI solutions, augmented and pre­
follow. Secondly, to apply data analytics techniques, such as su­ dictive analytics is among its core functionalities. Of course, a
pervised and unsupervised machine learning techniques, depending functionality to create custom analytic apps is available to this
on the pre-processed data, so as end-user to get estimations, pre­ platform as well. Besides these platforms that provide strong ana­
dictions and fault detection from several places in the shop-floor or lytic services and options to create custom dashboards and compare
factory and make critical decisions. Finally, with the predictive ML models with a more general scope including multi domains,
maintenance methodologies incorporated, the proposed platform Siemens offers its platform name Mindsphere11 which is primarily
offers to maintenance workers the prospect of being a step ahead focused on industrial solutions. Mindsphere is an Industrial IoT as a
towards controlling and preventing machine damages. Overall, the service solution using advanced analytics and AI. It provides services
proposed cognitive analytics platform is adapted to new and com­ such as monitoring, real-time analytics and connection of various
plex conditions of Industry 4.0, is extendable due to micro-services assets from the shop-floor. The analytic solutions are available as
architecture with which it is built, has a wide range of machine apps in the Mindsphere Store. Custom apps can be created by using a
learning (supervised, unsupervised, ensemble) and deep learning low-code development platform connected to Mindsphere.
techniques empowering the analysis, has the ability to train online The remainder of the paper is organized as follows. In Section 2, a
the machine and deep learning techniques and to interact in real- brief presentation of the three core modules, Data Acquisition,
time with the end user and finally has the cognition mechanism in Knowledge Management and Predictive maintenance, of the cognitive
order to maintain constant accuracy and quality results. Fig. 1 below analytics platform is presented along with their sub-modules. In
presents in brief the capabilities of the proposed platform. Section 3, a precise description of experimental results and a dis­
Besides research approaches and technologies, methodologies cussion about them along with other aspects is provided and in
and algorithms that led our motivation to built the introduced Section 4 our conclusions are drawn.
platform it worth’s to mention some indicative example of the
world-leading analytics platform that consecrates a lot of common 2. Materials and methods
characteristics with the proposed work in this article. Microsoft’s
Power BI8 is a BD analytics platform that enables the creation of The current work proposes a complete system that can handle,
custom analytics dashboards with features such as monitoring, se­ analyze and exploit resourcefully machine data from factory. The
lection of the monitoring assets and features, creation, training and aim of the proposed system is to offer a thorough predictive main­
comparison of ML models and connection with a large variety of tenance tool which can facilitate the job of maintenance workers
data sources. The focus of the applications and dashboards is BI and and achieve an efficient production scheme, handle unforeseen
customer solutions. KNIME9 Analytics Platform supports creation conditions, predict failures and aid the decision making process.
and training of ML/DL models and comparison option among dif­ Besides the maintenance workers, the introduced platform aims to
ferent models as well. It is a desktop type application and for provide a tool for factories’ data scientist to setup, train and compare

8 10
https://powerbi.microsoft.com/en-us/ https://saphanajourney.com/sap-analytics-cloud/
9 11
https://www.knime.com/knime-analytics-platform https://siemens.mindsphere.io/en

3
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 2. General architecture.

Table 1
A description of the correspondence of each User Interface with the allowed user access and the responsible module of the system’s architecture.

User interface Module Data Scientist End User

Machine Configuration Data Acquisition

Data Visualization Data Acquisition X X
Data Cleansing Data Acquisition X X
Online training Knowledge Management X
Neural network training Knowledge Management X
Live anomaly monitoring Predictive Maintenance X X
Time-series forecasting Predictive Maintenance X X

effortlessly, by using a graphical interface, various machine learning construct the integrated architecture of the proposed module and
models in order to select the best available one to run for real time are fully demonstrated through the data analytics platform. Fig. 2
anomaly detection. This feature is of great importance as a data shows the architecture of the proposed system and the connections
scientist can create models that will be used for machines’ mon­ between the distinct modules.
itoring by maintenance workers and with connection to self-trained The proposed methodology is realized through the cognitive
capabilities of the platform these models’ performance and accuracy analytics platform and each module is demonstrated with the spe­
will be in a high level without the need of further involvement by cially designed user interfaces. The system has been designed with
data scientist. The system contains three distinct modules in a different views for each user category. Apart from the Super Admin
micro-service architecture and provides each of them to the user which is responsible for the initialization and maintenance of the
through advanced visualization techniques. The Data Acquisition platform, the two user views are the data scientist view and the
module involves the connection of the factory data with the system regular end user view. In order to correlate the user interfaces with
and the initial processing of them. The Knowledge Management the architecture module and the specific user views, a descriptive
module incorporates the data analytics part of the system including table is shown below (see Table 1). The platform’s development was
classification and clustering approaches and additional neural net­ based on well known programming languages and open source
work methods applied on the machine data and generating pre­ frameworks. Python language was used for the development of core
dictive models which interact with the other modules. Last but not back-end functionalities related to analytics services and JavaScript
least, the Predictive Maintenance module which encompasses all (Node.js12) for the development of the rest back-end functionalities.
processes regarding real-time data. These are the Anomaly
Monitoring, the Performance Monitoring, the cognitive Automated
Retraining mechanism and the Time-series Forecasting. All these parts 12
https://nodejs.org/en/

4
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

MongoDB used for data storage. The front-end part is based on 2.2. Knowledge Management
AngularJS framework and use of libraries such as Chart.js 13 and
D3.js.14 Docker containers were used for the deployment of various The proposed cognitive data analytics framework provides sev­
platform’s parts and Portainer15 for the container’s management. eral AI techniques in the direction of early fault detection of in­
dustrial machines. The Knowledge Management module offers the
2.1. Data Acquisition basic functionality of the proposed system, the Online Training (OT)
consisting of three sub-modules, the OT with Classification Methods,
The proposed system handles historical and real-time factory the OT with Clustering Methods and the OT with Neural Networks. It
data through the Data Acquisition module. The connection of the accommodates a list of classification, clustering and deep neural
data analytics framework and the factory site is accomplished with network methodologies that can be initialized, run and saved by the
the use of the IDS Trusted Connector (Otto and Jarke, 2019; Otto users in order to result to accurate and robust prediction models.
et al., 2019). Secure Big Data exchange is achieved with the config­ Fig. 4 shows the various algorithms implemented in each category of
uration of IDS Trusted Connectors ecosystem. As depicted in Fig. 3 model training which are described in detail later on this chapter.
the system encompasses two IDS Trusted Connectors: . The user can select the data to be trained, select the algorithm
parameters and split the dataset into training and testing sets, run
• The first connector is deployed on the factory site. The IDS con­ the desired machine learning algorithm and finally visualize the
nector receives the machines’ data through an MQTT broker evaluation results. The modular and scalable architecture based on
which is connected with the factory’s cloud infrastructure. The micro-services enables the integration of new algorithms and the
factory cloud repository is the Data Provider of IDS Architecture. extension of the OT functionality of the system.
• The second IDS Connector is placed in the Data Consumer site, The proposed platform addresses the anomaly detection of in­
specifically the Cognitive Analytics framework. An MQTT Broker dustrial machine data through classification and clustering methods.
is also used in order to enable data exchange with the data Based on the needs of the data set and specifically the existence or
provider. absence of quality indicator features, meaning labeled and unlabeled
data, the system provides supervised and unsupervised learners for
The IDS Connector is an open-source system for sovereign data training. The algorithms reported at this chapter are included in the
exchange based on usage policies. The connector is a core compo­ primary version of the system and were selected due to their sa­
nent for building data spaces according to the DIN SPEC 27070 tisfactory results in respect to the prediction models’ performance.
standard. The connector uses the most recent and established However, new methods can be incorporated due to fact that the
technologies, and identity certificates and TLS encryption to promote overall architecture is extensible.
secure factory connectivity and transmission of the sensitive and Furthermore, the Cognitive Analytics platform is supported by a
private industrial data. The IDS connector was selected as it can be machine learning sub-module for OT. A designated neural network
integrated into existing software systems with minimum effort to architecture is used for time series reconstruction and outliers are
easily enhance them in terms of data sovereignty. It provides a REST detected based on the training results. This method adds extra value
interface that allows data to be easily added and accessed by other to the system functionality as it can successfully handle unlabeled
connectors, a fact that increases the inseparability of the proposed and unstructured data, achieve accurate results and manage the
cognitive platform. increasing volume of data as the system continuously accepts new
The data analytics platform has the Machine Administration User real-time data streams.
Interface tab that contains machine data in the form of time series. It The methodology of the specific sub-module is depicted in Fig. 5.
could be different measurements of a machine along with a specific The historical data are fed to the supervised learners for classifica­
timestamp. Each machine has to pass by a configuration process tion or to the unsupervised learners for clustering. On top of these
where all information is gathered for the specific data set, including methods there are ensemble learners available that can boost the
the type of data, the size of the data set, the names of the features, performance of simple models. The result of the training is the
the quality indicator, when it is available. prediction models that will be delivered to the Live Anomaly Detec­
The raw data set has to be transformed into useful information. A tion sub-module. The generated prediction models will receive real-
basic cleansing of the data is enforced by removing constant, empty time data as input and produce the corresponding anomaly predic­
or duplicated valued features. Also, low variance features are re­ tions. The OT and prediction processes are connected through a
moved as they cannot improve the prediction model’s performance. cognitive check mechanism which reserve the smooth and accurate
This step is automatically performed when the data is loaded on the performance of the prediction based on the model performance and
platform. The data can be visualized in the Data Visualization User the quality of new incoming data.
Interface of the platform. Correspondingly, historical data are fed to the neural network
Following is the Data Cleansing User Interface which provides the and achieve time series data reconstruction and then anomaly de­
ability to pre-process the data, so as to fit properly in the machine tection. The generated prediction models are also deployed on Live
learning algorithms. After data cleansing, potential inconsistency or Anomaly Detection sub-module and receive real-time machine data.
incompleteness in the remaining features has to be excluded. The The pipeline of the machine learning sub-module is plainly de­
Data Preparation step uses the following methods and is available to scribed in Fig. 5 from the beginning of the method with historical
the end-user in order to process and visualize the initial data: . data, through training, prediction and retraining stages, to the live
anomaly detection procedure. Each stage is visualized in the plat­
• Interpolation for estimating missing values between known data form and the user-platform interaction is strictly necessary at each
points is used. point.
• Zero removal process for eliminating zero values.
• Normalization for scaling each feature individually. 2.2.1. Online training with classification methods
Anomaly detection is addressed as a binary classification pro­
blem for the labeled data sets. The two classes are the normal and
13
https://www.chartjs.org/
the abnormal data points. Well-known supervised training methods
14
https://d3js.org/ are available for OT. Fig. 4 depicts the set of algorithms implemented
15
https://www.portainer.io/ at this work. Specifically, for the supervised learning, the deployed

5
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 3. IDS connection.

Fig. 4. Available algorithms of Online Training.

algorithms are the tree-like model of decisions, Decision Tree Clas­ over-fitting of the model, the user has the option of applying cross
sifier (Lee and Siau, 2001), the meta estimator that fits a number of validation. The user can edit and train multiple models and review
decision tree classifiers, Random Forest (Breiman, 1999), the robust the results.
predictor for two-group classification problems, Support Vector Thus, for the selected machine, the prediction models are as­
Machines with Polynomial and Radial Basis Function and lastly, the sessed by calculating precision, recall, accuracy and f-measure me­
Back Propagation Network method with one hidden layer, Multi- trics on the testing set. The models are trained in real-time and the
layered Perceptron (Rumelhart et al., 1986). results are the evaluation metrics calculated on the testing set. The
On the top of individual learners, an ensemble method is avail­ system proposes the optimal prediction model and offers a com­
able, in order to generate a more accurate solution comparing with parison visualization between each model.
the results of single models. Adaptive boosting (or AdaBoost) tech­
nique combines a set of weak learners and manages to improve
classification performance (Vafeiadis et al., 2015; Shin, 2015; Arditi 2.2.2. Online training with clustering methods
and Pulket, 2005; Shin et al., 2012; Gavrishchaka, 2006). There are On the other side, for the OT of unlabeled datasets, unsupervised
many forms of boosting algorithms (Nath and Behara, 2003; learning techniques were implemented aiming to detect anomalies
Schapire and Freund, 2012; Freund and Schapire, 1995). AdaBoost through data clustering. Density-Based Spatial Clustering of
SAMME - Stagewise Additive Modeling with Multi-class Exponential Applications with Noise (DBSCAN) is the data clustering algorithm
loss function is used in this work, which is an extension of Ada­ which discovers clusters of arbitrary shape in spatial data-spaces
Boost.M1 algorithm and performs well in two-class classification with noise. Next is the Local Outlier Factor (LOF), which provides a
scenarios. The system provides supervised and ensemble supervised factor of how close is a data point to its neighbors in respect to its
learning algorithms for training along with a set of parameters for neighbor being also close to it. The One-Class Support Vector
each one of them. The data set is divided to training and testing set, Machine (One-class SVM) algorithm classifies the points that lie
the percentage of which is defined by the user. In order to avoid outside some boundaries of the data space as outliers. Finally,

6
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 5. Machine learning sub-module.

Kmeans iteratively tries to partition the dataset into clusters with using the model itself (Rousseeuw, 2005). Another evaluation metric
each data point belonging to only one cluster. in case that the ground truth labels are not known, is the Davies-
Majority Voting (MV) is the ensemble method which aims to Bouldin index (Davies and Bouldin, 1979). The “similarity" between
boost unsupervised learners. This method generates a common clusters is measured by this metric by comparing the distance be­
consensus from multiple learners by selecting the result voted from tween clusters with the size of the clusters themselves. Next is the
the majority of these learners (Jung and Lease, 2012). Specifically, Calinski-Harabasz index also known as the Variance Ratio Criterion
the procedure of the method is to calculate the average result (Caliński and Harabasz, 1974). The index is the ratio of the sum of the
coming from multiple learners and round according to a decision dispersion between-clusters and inter-cluster for all of them. Last, is
threshold. MV is also used as a substitute for the ground truth in case Dunn index (Dunn, 2008), another metric that aims to identify the
of unlabeled datasets, especially on the Automatic Retraining Mon­ compact sets of clusters and the well separated ones.
itoring sub-module of the proposed methodology (See 2.3.2). The Thus, for the selected machine, the prediction models are as­
system provides unsupervised and ensemble unsupervised learning sessed by calculating performance evaluation metrics. The models
algorithms for training along with a set of parameters for each one of are trained, again, in real-time and the results are the evaluation
them, as in the supervised case. metrics calculated on the testing set. The system proposes the op­
The unsupervised model assessment is performed by four clus­ timal prediction model and offers a comparison visualization be­
tering performance evaluation metrics: Silhouette Coefficient, tween each model.
Calinski-Harabasz Index, Davies-Bouldin Index and Dunn Index.
Those are metrics for evaluating clustering algorithms following an 2.2.3. Online training with neural network
internal evaluation scheme, where the metric result is based on the The prediction model is an LSTM network (see Fig. 7) and is
clustered data. The Silhouette Coefficient is an example of evaluation based on an LSTM autoencoder architecture (see Fig. 6). The aim of

Fig. 6. LSTM autoencoder network architecture.

7
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 7. LSTM network model architecture.

an autoencoder is to learn a compressed representation (encoding) factory machines. The Live Anomaly Monitoring sub-module is en­
of the input data that retains the most important information of the hanced by Automatic Retraining Monitoring sub-module aiming to
input. The autoencoder then uses this encoding to produce an output sustain the accurate performance of predictors through a dedicated
that is as close as possible to the original input. After the model has cognitive check. Machine Feature Early Prediction is a supplementary
created the reconstruction of the original time series, the re­ sub-module for machine feature prediction and fault detection at a
construction error for each step of each time series is calculated. For very early stage. These three sub-modules offer the machine main­
each time step, the reconstruction errors for all time series are used tenance workers or administrators the prospect of being a step
as input to a 1-layer neural network, whose output is the model’s ahead towards controlling and preventing machine damages.
fault prediction. The input data to the LSTM network consist of d
different time signatures t, each accompanied by n values s, for ex­ 2.3.1. Live Anomaly Monitoring
ample sensor readings (plus label l signifying whether there was a The primary functionality the proposed system offers to the end
fault for the train dataset). When training, we use 10-fold cross- user is the Live Anomaly Monitoring. Real-time data are constantly
validation to split the data into a training, a validation and a test arriving through the IDS Connector as described in the previous
dataset. The whole LSTM model architecture is presented below: chapters (See 2.1). Live data are visualized and renewed every one
The encoder consist of a bidirectional LSTM, followed by dropout minute. Concurrently, live data are fed to the optimal predictive
layer for regularization. The bidirectional LSTM has an embedding models generated at the OT phase. The input data of Live Anomaly
dimension size equal to n/2, so that the concatenated output of both Monitoring sub-module follow the exact same data scheme with the
LSTMs has dimension of n. The decoder consist of a repeater layer, historical data of the predefined machines.
bidirectional LSTM and a reversing layer. The repeater layer repeats The machine data are constantly monitored and for each instance
the encoding K times so that it can be used as input to the bidirec­ of measurements, the system recognizes normal behavior or detects
tional LSTM. The bidirectional LSTM has an embedding dimension anomalies and depicts them as colored points on the data visuali­
size equal to n, so that the concatenated output of both LSTMs has zation widget. On the same diagram the system exposes the action
dimension of 2 *n, equal to that of the input. Finally, the reversing of the Automatic Retraining Monitoring with red lines, which is
layer flips the output of the bidirectional LSTM along the time axis. described in the next section. The red line indicates that the system
The reconstructed time series from the autoencoder, along with the has automatically triggered the retraining mechanism. The Live
original time series, are passed through a subtraction layer, the Anomaly Monitoring also contains a clock widget which counts the
output of which is passed through an absolute layer, producing the anomalies detected in the last hour and an alert in case of retraining
absolute error of the reconstruction for each time step of each time triggering. A table with information about the retraining is also
series. For each time step, the errors are fed to an 1-layer fully- available at the same module.
connected network with a single output and sigmoid activation. This
output is the model’s fault prediction for that time step.Fig. 7. 2.3.2. Automatic Retraining Monitoring
The OT of the LSTM model described above, is arranged by the The Automatic Retraining Monitoring is a sub-module running on
user. This model configuration includes the set of various para­ the background of the Live Anomaly Monitoring. The aim of this
meters, such as optimizer, loss function, number of epochs for module is to secure the high accuracy of the prediction models
training, look-back parameter, early stopping and validation folds. running in real-time through an automatic and cognitive way. This is
After training the user can visualize the results of the LSTM achieved by automatically updating the prediction models that are
model (or models). The evaluation summary includes the loss and deployed for the live data. A double-oriented cognitive mechanism
validation loss of the reconstruction during the epochs and the Mean was designed in order to detect the occasions when retraining is
Absolute Error (MAE) from the reconstruction result. A comparison needed. The system continuously examines the dataset’s statistical
between trained models with different sets of hyper-parameters are characteristics and the models’ performance in order to capture
presented to the user. changes. The architecture of the retraining mechanism is presented
in Fig. 8.
2.3. Predictive Maintenance The new incoming data are constantly reviewed in terms of
statistical measurements like mean value, standard deviation,
The proposed system supports predictive maintenance opera­ maximum, minimum values. These metrics are compared to the
tions in this specific module, in a threefold level. Thus, Predictive corresponding measurements of the trained dataset. In case of high
Maintenance module consists of three sub-modules, Live Anomaly variation between these values, the model has to be updated. The
Monitoring, Automatic Retraining Monitoring and Machine Feature system triggers the retraining mechanism and the model is fed and
Early Prediction. Live Anomaly Monitoring, offers through an advanced trained with the latest data. The cognitive mechanism observes re­
User Interface, to the user to observe real time data and anomaly peated measurements with variations and triggers the retraining so
detection on them in order to overtake potential breakdowns of as to eliminate accidental discrepancies of the machine live data and

8
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 8. Model-based retraining mechanism.

prevent the generation of poor anomaly detection results. This used for the procedure of the future values predictions. There is a
specific method is called data-driven retraining for ease of reference. variety of forecasting methods to choose from, including simple
Correspondingly, the next method is called model-driven retraining autoregressive models, moving average and the two most widely
as it monitors the prediction model performance. The model-driven used approaches to time series forecasting, exponential smoothing
retraining method. and ARIMA models.
An Initial Prediction Window (IPW) is determined at the training Data pre-processing is an essential stage where data are handled
phase, which is a specific number of real-time predictions tested appropriately in order to provide them as inputs to the forecasting
against the real ones when those are available. In case of the ma­ algorithms later. First of all, a dataframe containing the timestamps
chines with labeled data, the real values are given and compared to and the corresponding data values is created from the input data
the predicted ones. In case of the machines with unlabeled data, the which are either coming from a pre-defined stored machine or from
result of major voting method substitutes the labels of data instances a csv file. Then missing values are eliminated with the help of
and is compared with the predicted results in order to extract the backward and forward filling methods to ensure the continuity and
performance metrics. In both cases, a confusion matrix is created the validity of the dataset. From then on, the timestamps are con­
and the metrics, precision, recall, accuracy and f-measure are cal­ verted into date objects where the format of the dates is handled in a
culated. Based on the values of f-measure, the IPW is changed (in­ way that the forecasting methods will be able to determine the data
creases or decreases) or remains the same. More specifically, a frequency (yearly, monthly, hourly, etc.).
minimum and maximum value is defined for the IPW values, along Depending on the user’s choice, a forecasting method calculates
with a threshold for the f-measure value. Starting from the max­ the predictions of the specified feature for a certain time period in
imum IPW value, f-measure is calculated for this window. If f- the future. The first available method is autoregressive models
measure exceeds the defined threshold, the training model remains (Nerlove and Diebold, 1990). In these types of models, the featured
as it is, whereas IPW increases by 10 if f-measure is higher than 90%, variable is forecasted using a linear combination of its past values.
decreases by 10 if f-measure is lower than 80% and remains as it if f- The term autoregression specifies that it is a regression of the
measure falls between 80% and 90%. This process is repeated until variable against itself. The next available forecasting method is the
the IPW equals the minimum IPW. In case that f-measure falls be­ moving average (Nerlove and Diebold, 1990). Instead of using past
hind the defined threshold, the retraining mode is triggered and the values of the variable in a regression, moving average model uses
IPW value resets to the maximum value. past forecast errors in a regression-like model or alternatively can
be thought as a weighted moving average of past forecast errors.
2.3.3. Machine Feature Early Prediction ARIMA models are also engaged for the feature early prediction
The Cognitive Analytics Platform includes also the sub-module since all of the above methods belong to the ARIMA family (Hussan
of Machine Feature Early Prediction. This specific functionality et al., 2018). Specifically, ARIMA is a combination of both Auto­
achieves forecasting of specific machine features and outlier de­ regressive (AR) and Moving Average (MA) models. Moreover, there
tection on the predicted time series. The purpose of the forecasting is a differencing pre-processing step called integration, which
module is to make predictions of future values of time series data renders a non-stationary time series stationary by computing the
and give an insight on their forthcoming behavior, achieving an differences between consecutive observations. The term statio­
even more earlier fault diagnostic step for the machine. It is a tool narity refers to a time series whose statistical properties (mean,
where the user provides the data, the forecast time period (fore­ variance, etc), are all constant over time. The last available method
cast horizon) of interest and the forecast method which will be is exponential smoothing (Shastri et al., 2018). The general idea is

9
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 9. (a) ARIMA, (b) autoregression and (c) exponential smoothing time series forecasting.

that the produced forecasts are weighted averages of previous Grid search provides a way to test different configurations of
observations with higher weights assigned on the more recent ones each forecasting method (Li and Abdallah, 2020; Tran and Le, 2020;
and decreasing exponentially for the older observations. There are Yu and Zhu, 2020). Knowing the possible parameter values, it creates
three variations in this category emanating from the existence of various configurations of parameter sets for each method (ex­
the trend and seasonality components in the time series (Billah ponential smoothing or ARIMA) and then passes them onto the
et al., 2006): cross-validation algorithm, that decides the method and the set of
parameters which yield the best results.
• Simple exponential smoothing (Suitable for data with no clear Cross-validation (Bergmeir et al., 2018) is the part of the method
trend or seasonal pattern) that decides which configuration produces the best forecasts. It is
• Double exponential smoothing (Suitable for data with trend) based on the philosophy of cross-validation used in machine
• Triple exponential smoothing (Suitable for data with both trend learning problems, but in the case of time series it is implemented
and seasonal patterns) differently due to the time dependence of the data. Train and test
splits are not random but they form an expanding window training
The output of this module includes the predicted values of the approach. Each split contains the training data followed by the test
selected method. Figs. 9a, 9b and 9c are some indicative plots de­ data in their original chronological order. Then the training set ex­
picting forecasts for ARIMA, autoregressive and exponential pands to include the previous test set and the new test set is formed
smoothing methods, respectively. The implementation of the accordingly. The size of these sets depends on the forecast horizon
aforementioned methods is based on python’s statsmodel module, the user provided. After each split, the predicted values are com­
which provides all necessary functions for the estimation of many pared to the actual values with the help of the Root Mean Squared
statistical models. Each method is implemented as a python function Error (RMSE) metric. Then an overall average value is extracted from
and all share a similar structure. This means they have the same them to yield the final error score for the current configuration. In
inputs (dataset, forecast horizon) and they all return the calculated the end all configurations passed from grid search algorithm to
predictions in a similar manner. cross-validation algorithm are assigned an error score. The best
All of the methods used in this sub-module address the fore­ configuration is the one with the lowest RMSE score.
casting problem, although the success of each one depends on the On top of the time-series forecasting for machine features a simple
behavior of the input data. Moreover, implementations of the same anomaly detection process is carried out with the intention of deliver
category may vary significantly in results, due to the different set of the machine condition monitoring at an extra early stage. Z-score
parameters that must be passed in order to handle the various statistical measure is deployed for anomaly detection. The rationale of
components of the time series. This is the reason why cross-vali­ this method is to consider the time-series data as a Gaussian dis­
dation and grid search algorithm were created for the cases of ex­ tribution and anomaly points are the data points which deviate a lot
ponential smoothing and ARIMA. The user can choose whether to from the population mean. How far is this deviation is determined by a
enable or not these algorithms when choosing one of the above defined threshold. So, Z-score indicates how many standard deviations
methods. When cross-validation and grid search are disabled a away a data point is from the mean and the threshold pinpoints the
simpler configuration is utilized instead. outliers which are the data points that exceed it.

10
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 10. Machine Administration User Interface.

3. Results and discussion Last but not least is the Data Cleansing User Interface (see Fig. 12).
The user can select specific feature of the data set, time range and
3.1. Case study pre-processing method(s) to apply to the selected data frame. The
user interface contains the raw and the clean data and the user can
At this point, a case study is presented in order to demonstrate select to view one of this time series or both on the same diagram.
the Cognitive Platform’s use and functionality in a practical way. The
use case scenario is the Cognitive Platform application on Injection 3.3. Knowledge Management
moulding machines. This practical application of the platform is
presented in detail in the work of (Rousopoulou et al., 2020), so in The Knowledge Management module consists of supervised, un­
current work the presentation of platform’s functionalities is the supervised and neural network learning. Each of these methods are
main focus of the result’s section and not the analysis of algorithmic available in the dedicated user interfaces and the user can follow the
outcomes. The machines’ basic operation is shaping rubber or plastic process and complete a model training. The user can accomplish su­
parts by injecting heated material into a mould, specifically, for the pervised or ensemble training on labelled data sets by selecting the
fabrication of plastic components for electric shavers. Six machines corresponding algorithms as shown in Figs. 13a and b. At this point, the
are included in this specific factory site which generate machine log user can set the algorithm parameters and start the model training.
data in the form of time series. The data includes measurements When model training is over, the platform provides the results of
such as temperature, pressure, energy consumption and time. The training. As shown in Fig. 14a the evaluation metrics of training are
system indicates “bad” shots during injection as anomaly points, visualized at the corresponding user interface. A comparison be­
which lead to rejected products. The advantages given by the Cog­ tween the results and the previously trained models is also available
nitive Platform to this case study are the machine condition mon­ (Fig. 14b).
itoring and early predictions in order to optimize machine operation Respectively, the unsupervised training is realized in the same
and minimize losses and costs. Following in this chapters are the way. Fig. 15 shows the Unsupervised model training preparation User
demonstration of the Cognitive Platform applied in Injection Interface, which provides unsupervised (see Fig. 15a) and ensemble
Moulding machines. unsupervised (see Fig. 15b) learning algorithms for training along
with a set of parameters for each one of them.
3.2. Data Acquisition Next, the evaluation of unsupervised models is demonstrated at
the defined User Interface as depicted in Fig. 16a. Again, the system
Starting from the Data Acquisition module, the Cognitive proposes the optimal prediction model and gives a comparison be­
Platform provides the Machine Administration User Interface which tween each model (see Fig. 16b).
contains the configuration of the factory machines and is the main The last capability of the Online Training module is the set up and
page of the platform. As shown in Fig. 10, the user can see the running of an LSTM neural network. The user has the opportunity to
available machines and select any of them to train data or even use a insert preferences and configurations to the LSTM model (see Fig. 17).
“Demo" machine with new data from a csv file. After training the user can visualize the results of the LSTM
Machine data are visualized in the Data Visualization User model (or models) (see Fig. 18a). A comparison between trained
Interface as it can be seen in Fig. 11. The user has the option to select models with different sets of hyper-parameters is also presented to
specific feature of the data set and specific time range to explore. the dedicated user interface as depicted in Fig. 18b).

11
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 11. Data visualization User Interface.

3.4. Predictive maintenance about the latest model retraining including the occurrence time, the
f-measure percentage before and after retraining, the type of re­
The Cognitive Analytics Platform performs predictive main­ training (data-driven or model-driven), the algorithm and the
tenance operations such as machine state operation and early fault training type. An alert visualization is also available, which turns
detection through a constantly updated methodology of advanced into red in case of retraining activation and a clock visualization
analytics and cognitive retraining. These functionalities are de­ which shows the mean of outliers detected in the last hour.
monstrated through the Anomaly Live Monitoring User Interface. As Finally, the Cognitive Analytics Platform provides the user in­
depicted in Fig. 19, the interface provides a data visualizer which terface of Machine Feature Early Prediction. In order to achieve an
consists of a diagram of the time series along with detected outliers even earlier indicator of the machine state, the system displays a
(green dots) and points where retraining is triggered (red lines). The time series prediction in the dedicated interface (Fig. 20). The user
Anomaly Live Monitoring User Interface also provides information can explore this functionality and select specific machine, machine

Fig. 12. Data cleansing User Interface.

12
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 13. Supervised model training preparation User Interface. (a) Supervised learning techniques selection and (b) Ensemble supervised learning techniques selection.

Fig. 14. Supervised training results User Interface. (a) Machine selection tab and (b) Online Training results for specified machine.

feature, method and horizon for the future prediction. The historical 3.5. Comparison with market-ready solutions
values and the predicted ones are visualized by different color.
The data depicted in each user interface figure come from the A short comparison between the introduced platform and the
case study of injection moulding machines. This case study was used market-ready solutions were mentioned in the introduction section
in order to demonstrate the potential of the proposed system and is documented in this section. The comparison is based on the public
proves the multiplicity and wide range of available functionalities available information for each platform and primarily in the
towards smart manufacturing in a real-world scenario. common functionalities with the proposed platform. The scope of

Fig. 15. Unsupervised model training preparation User Interface. (a) Unsupervised learning techniques selection and (b) Ensemble unsupervised learning techniques selection.

13
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 16. Unsupervised training results User Interface. (a) Machine selection tab and (b) Online Training results for specified machine.

Fig. 17. LSTM model configuration User Interface.

Fig. 18. LSTM training models result User Interface for a specified machine. (a) Model results tab and (b) comparison of trained models.

14
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Fig. 19. Anomaly live monitoring User Interface.

Fig. 20. Machine feature early prediction User Interface.

15
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

Table 2
Platform’s strengths and weaknesses compared to market-ready ones.

The proposed solution focused on Industry 4.0. On the other hand Power BI, SAP Power BI, SAP Analytics Cloud and Siemens Mindsphere is well-known, widely
Analytics Cloud and KNIME analytics provide general solutions but not used and tested platforms coming from world-leading companies in the analytics
specialized to Industrial sector. The market-ready solutions have a generic focus domain.
especially for customer and business analytics.
The proposed solution supports secure and sovereign connectivity with industrial In the proposed platform less functionalities are available and less connectivity
environments (IDS standard) based on industrial and IoT standards. On the options than the ones were supported by market-ready solutions. Power BI
other hand the market-ready solutions provide their proprietary security provides support for many data types and connections with many other
options same for various types of customers and not so to specific to industrial Microsoft Services and tools. SAP Analytics Cloud contains a large amount of
domain. Mindsphere has introduce an idea of IDS and FIWARE compatibility but packages that can be connected to the analytics main product. SIEMENS
it is not delivered yet. Mindsphere can be extended with apps from its Store in order to enhance the
main functionality. On the other hand these kind of extensions seems not
possible for KNIME platform.
Cognitive Functionalities for real-time data analytics is available by the introduced Graphical drag and drop interfaces for creating real time data flows are available
platform. Power BI provides cognitive services only available for data on KNIME and Power BI platforms.
preparation stage and for other application usages not for industrial
applications focused on anomaly detection. KNIME and SAP analytics seem to
miss this kind of functionality. Same for Mindsphere.
ML models comparison functionalities and interfaces are available to the introduced Power BI, SAS Analytics Cloud and Mindsphere provides detailed documentation,
platform. The same kind of functionalities are available on both Power BI and support and developers communities etc.
KNIME. However, it seems that SAP Analytics Cloud and Mindsphere have not
adopt this approach.

this comparison is to summarize some features and directions that an effort to follow the continuously growing industry needs
are common in all platforms, some that are missing in the market- (Padmanaban and Mukesh, 2018; Gröger et al., 2014;
ready ones and to point out some of the weakness that a newly Thammaboosadee and Wongpitak, 2018; Ansari and Glawar, 2019).
introduced platform has. So best to our knowledge we collect the In addition, a human-knowledge injection module will be added in
information available at Table 2. order to enable human to add their observations and domain
knowledge for further improvement of platform’s analytic and pre­
4. Conclusions dictive capabilities. Besides this, a further extension form moving
from a complete cloud platform to a kind of a ’hybrid’ platform that
A cognitive analytics platform is presented in this work in an will enable the edge execution of analytic models based also on
effort of making a big step to digital transformation of traditional federated learning, is included in the next steps related to platform’s
manufacturing industry by utilizing exponentially growing tech­ further improvement and update.
nologies from the field of computer science. The platform is based on
a micro-service architecture so as to be adaptive and extendable over CRediT authorship contribution statement
time, supports different types of users, such as data scientists for
setting up AI models and simple end-users for monitoring factory Vaia Rousopoulou: Conceptualization, Methodology, Software,
conditions based on previous trained models. Furthermore, the Validation, Formal analysis, Investigation, Data curation, Writing –
platform contains three core modules, the Data Acquisition, the original draft, Writing – review & editing, Visualization.
Knowledge Management and the Predictive Maintenance. Based on Thanasis Vafeiadis: Conceptualization, Methodology, Software,
these modules, a wide range of jobs is covered in the shop-floor or Validation, Formal analysis, Investigation, Data curation, Writing –
factory regarding secure and inter-operable factory connectivity, original draft, Writing – review & editing, Visualization.
achieving efficient production schemes, handling unforeseen con­ Alexandros Nizamis: Conceptualization, Methodology, Formal ana­
ditions, predicting and detecting machine failures and making de­ lysis, Investigation, Writing – original draft, Writing – review &
cisions. Machine learning techniques, deep learning architectures, editing, Visualization. Ioannis Iakovidis: Software, Validation, Data
clustering methodologies and time series models known from sci­ curation, Writing – review & editing, Leonidas Samaras: Software,
entific literature were employed so as to extract the information Validation, Data curation. Alkis Kirtsoglou: Software, Data curation.
from the industrial data imported in proposed platform from IoT Kostantinos Georgiadis: Software, Data curation.
devices. All the information mentioned above is provided to the end Dimosthenis Ioannidis: Conceptualization, Supervision, Funding
user through a user interface that utilizes advanced visualization acquisition. Dimitrios Tzovaras: Supervision, Project administration,
techniques. This work constitutes an integrated solution and highly Funding acquisition All authors contributed in changes and correc­
adaptive to shop-floor and factory environment in an effort to adopt tions before the final submission.
information and communication technologies and create an in­
tensive connection with the surrounding physical world. Declaration of Competing Interest
Future work will comprise the addition of semantic technologies
for the description of ML models and the application of reasoning The authors declare that they have no known competing fi­
functionalities over them. The models would be described by well nancial interests or personal relationships that could have appeared
known data models such as Predictive Model Markup Language to influence the work reported in this paper.
(PMML) and Portable Format for Analytics (PFA) (The Data Mining
Group, 2021). An ontology will be built on top of it in order to enable Acknowledgment
rules appliance and reasoning related to various models metrics and
parameters. These functionalities will replace the functional pro­ This project has received funding from the European Union’s
gramming logic during the retraining phase and will increase the Horizon 2020 research and innovation programme under grant
interoperability, knowledge extraction capability and extend-ability agreement No 780732 - BOOST 4.0 and from BIG-SMART-LOG project
of the proposed platform. Furthermore, prescriptive analytics, as an under CHIST-ERA BDSI call and General Secretariat for Research and
effort to take advantage of the predictive future and to make deci­ Innovation (GSRI), Greece with code T9EPA3-00013. Furthermore, we
sions on any time horizon (immediate or long term), will be added as thank Bas Tijsma from Philips Drachten for providing us with

16
V. Rousopoulou, T. Vafeiadis, A. Nizamis et al. Computers in Industry 134 (2022) 103555

anonymized data and access to infrastructure that enabled us to apply S.V. Nath and R.S. Behara Customer churn analysis in the wireless industry: a data
our platform for the needs of the current study.This paper reflects mining approach, Annual Meeting of the Decision Sciences Institute, 505–510,
2003.
only the authors’ views and the Commission is not responsible for any Nerlove, M., Diebold, F.X., 1990. Autoregressive and moving-average time-series
use that may be made of the information it contains. processes. Time Series and Statistics. Palgrave MacMillan, UK, London. https://doi.
org/10.1007/978-1-349-20865-4-3
Otto, B., Jarke, M., 2019. Designing a multi-sided data platform: findings from the
References International Data Spaces case. Electron. Mark. 29 (4). https://doi.org/10.1007/
s12525-019-00362-x
Anaya , V. , Fraile , F. , Aguayo , A. , García , O. , Ortiz , Á. ,Towards IoT Analytics. A vf-OS Otto , B. , Lohmann , S. , Steinbuss , S. , Teuscher , A. , IDS Reference Architecture Model
Approach, International Conference on Intelligent Systems (IS), 10.1109/IS.2018. Version 3.0, International Data Spaces Association, Dortmund, https://www.
8710476, 2018. internationaldataspaces.org/wp-content/uploads/2019/03/IDS-Reference-
Ansari, F., Glawar, R., Nemeth, T., 2019. PriMa: a prescriptive maintenance model for Architecture-Model-3.0.pdf, 2019.
cyber-physical production systems. Int. J. Comput. Integr. Manuf. Padmanaban, R., Mukesh, R., 2018. Hadoopsec: sensitivity-aware secure data place­
Arditi, D., Pulket, T., 2005. Predicting the outcome of construction litigation using ment strategy for big data/hadoop platform using prescriptive analytics. J.
boosted decision trees. J. Comput. Civil Eng. 19 (4), 387–393. Comput. 6 (1), 116–121. https://doi.org/10.5176/2251-3043_6.1.109
Bergmeir , C. , Hyndman , R.J. , Koo , B. , A note on the validity of cross-validation for Rousopoulou, V., Nizamis, A., Vafeiadis, T., Ioannidis, D., Tzovaras, D., 2020. Predictive
evaluating autoregressive time series prediction, Computational Statistics and maintenance for injection molding machines enabled by cognitive analytics for
Data Analysis, 27th Color and Imaging Conference Final Program and Proceedings, industry 4.0. Front. Artif. Intell. 3, 86. https://doi.org/10.3389/frai.2020.578152
120, 70–83, 10.1016/j.csda.2017.11.003, 2018. Rousseeuw, P.J., 2005. Silhouettes: a graphical aid to the interpretation and validation
Billah, B., Maxwell, K.L., Ralph, S.D., Koehler, A.B., 2006. Exponential smoothing model of cluster analysis. J. Comput. Appl. Math. 20, 53–65. https://doi.org/10.1016/0377-
selection forecasting. Int. J. Forecast. 22 (2), 239–247. https://doi.org/10.1016/j. 0427(87)90125-7
ijforecast.2005.08.002 Rumelhart, D.E. , Hinton , G.E. , Williams , R.J. , Learning representations by back-
Breiman, L., 1999. Random Forests - random features. Nature 567 (7), 5748–5757. propagating errors, Technical Report, 323, 533–536, 1986.
〈ftp://ftp.stat.berkeley.edu/pub/users/breiman〉. Schapire, R.E., Freund, Y., 2012. Boosting: Foundations and Algorithms (Adaptive
Caliński, T., Harabasz, J., 1974. A dendrite method for cluster analysis. Commun. Stat. 3 Computation and Machine Learning Series). MIT Press.
(1), 1–27. https://doi.org/10.1080/03610927408827101 Sergio Alvarez-Napagao et al. , knowlEdge Project - Concept, Methodology and
Canizo , M. , Onieva , E. , Conde , A. , Charramendieta , S. , Trujillo S. , Real-time Innovations for Artificial Intelligence in Industry 4.0, International Conference on
predictive maintenance for wind turbines using big data frameworks, IEEE Industrial Informatics, 2021.
International Conference on Prognostics and Health Management, 70–77, 10.1109/ Shang, C., You, F., 2019. Data analytics and machine learning for smart process man­
ICPHM.2017.7998308, 2017. ufacturing: recent advances and perspectives in the big data era. Engineering 5
Carvalho, T.P., Soares, F.A.A.M.N., Vita, R., Francisco, R.P., Basto, J.P., Alcalá, S.G.S., 2019. (6), 1010–1016. https://doi.org/10.1016/j.eng.2019.01.019
A systematic literature review of machine learning methods applied to Shastri, S., Sharma, A., Mansotra, V., Sharma, A., Bhadwal, A., Kumari, M., 2018. A study
predictive maintenance. Comput. Ind. Eng. 137. https://doi.org/10.1016/j.cie.2019. on exponential smoothing method for forecasting. Int. J. Comput. Sci. Eng. 6,
106024 482–485. https://doi.org/10.26438/ijcse/v6i4.482485
Çínar, Z.M., Abdussalam, N.A., Zeeshan, Q., Korhan, O., Asmael, M., Safaei, B., 2020. Shin, Y., 2015. Application of boosting regression trees to preliminary cost estimation
Machine learning in predictive maintenance towards sustainable smart manu­ in building construction projects. Comput. Intell. Neurosci. https://doi.org/10.
facturing in industry 4.0. Sustainability 12 (19), 37–42. https://doi.org/10.3390/ 1155/2015/149702
su12198211 Shin, Y., Kim, T., Cho, H., Kang, K.I., 2012. A formwork method selection model based
Davies, D.L., Bouldin, D.W., 1979. A cluster separation measure. IEEE Trans. Patter Anal. on boosted decision trees in tall building construction. Autom. Constr. 23, 47–54.
Mach. Intell. 2, 224–227. Tao, F., Qi, Q., Liu, A., Kusiak, A., 2018. Data-driven smart manufacturing. J. Manuf. Syst.
Dunn, J.C., 2008. Well-separated clusters and optimal fuzzy partitions. Cybern. Syst. 4, 48, 157–169. https://doi.org/10.1016/j.jmsy.2018.01.006
95–104. https://doi.org/10.1080/01969727408546059 Thammaboosadee , S. , Wongpitak , P. , An Integration of Requirement Forecasting and
Efthymiou, K., Papakostas, N., Mourtzis, D., Chryssolouris, G., 2012. On a predictive Customer Segmentation Models towards Prescriptive Analytics For Electrical
maintenance platform for production systems. Procedia CIRP 3, 221–226. https:// Devices Production, International Conference on Information Technology (InCIT),
doi.org/10.1016/j.procir.2012.07.039 10.23919/INCIT.2018.8584864, 2018.
Freund, Y., Schapire, R.E., 1995. 3. CIE colorimetry. A Desicion-theoretic Generalization The Data Mining Group (DMG) official web page http://dmg.org/ 2021.
of On-line Learning and an Application to Boosting. Springer. Tran, T., Le, V.D., 2020. Grid search of exponential smoothing method: a case study of
Gavrishchaka, V.V., 2006. Boosting-based frameworks in financial modeling: appli­ Ho Chi Minh City load demand. Indones. J. Electr. Eng. Comput. Sci. 19. https://doi.
cation to symbolic volatility forecasting. Adv. Econometr. 25–35. org/10.11591/ijeecs.v19.i3.pp1121-1130
Gröger , C. , Schwarz , H. , Mitschang , B. , Prescriptive analytics for recommendation- Vafeiadis, T., Diamantaras, K.I., Sarigiannidis, G., Ch, K., 2015. Chatzisavvas A com­
based business process optimization, 17th International Conference on parison of machine learning techniques for customer churn prediction. Simul.
Business Information Systems (BIS), 25–37, Springer, 10.1007/978–3-319–06695- Model. Prac. Theory 55, 1–9. https://doi.org/10.1016/j.simpat.2015.03.003
0_3, 2014. Vafeiadis , T. , Nizamis , A. , Alexopoulos , N. , Ntinas , C. , Mastos , T. , Papadopoulos , A.
H.J. Jung and M. Lease Evaluating classifiers without expert labels, https://arXiv:1212. , Ioannidis , D. ,Tzovaras D. , Data analytics and data modeling for predictive
0960, 2012. maintenance and automated waste management: An integrated (IoT-Industry 4.0)
Hussan , A.-C. , Yamur, A.-D. , Lundberg , J., Time Series Forecasting using ARIMA sensor-based solution to improve factory procedures, International Conference on
Model: A case study of mining face drilling rig, The Twelfth International Internet of Things and Smart Manufacturing, 2018.
Conference on Advanced Engineering Computing and Applications in Sciences, Vafeiadis , T. , Nizamis , A. , Apostolou , K. , Charisi , V. , Metaxa , I.N. , Mastos , T. ,
2018. Papadopoulos , A. , Ioannidis , D. , Tzovaras , D. , Intelligent Information
Lee, S.J., Siau, K., 2001. A review of daata mining techniques. Ind. Manag. Data Syst. Management System for Decision Support: Application in a Lift Manufacturer’s
101 (1), 41–46. Shop Floor, International Symposium on INnovations in Intelligent SysTems and
Li, Y., Abdallah, S., 2020. On hyperparameter optimization of machine learning algo­ Applications, 1–6, 2019.
rithms: theory and practice. Neurocomputing 415, 295–316. https://doi.org/10. Vafeiadis , T. , Nizamis , A. , Pavlopoulos , V. , Guigliano , L. , Rousopoulou , V. , Ioannidis
1016/j.neucom.2020.07.061 , D. , Tzovaras , D. , Data Analytics Platform for the Optimization of Waste
Li, Z., Zhe, Wang, Y., Wang, K.-S., 2017. Intelligent predictive maintenance for fault Management Procedures, 15th International Conference on Distributed
diagnosis and prognosis in machine centers: Industry 4.0 scenario. Adv. Manuf. 5, Computing in Sensor Systems, 10.1109/DCOSS.2019.00074, 2019.
1–11. https://doi.org/10.1007/s40436-017-0203-8 Wang, J., Ma, Y., Zhang, L., Gao, R.X., WuDeep, D., 2018. learning for smart manu­
Naskos , A. , Nikolaidis , N. , Naskos , V. , Gounaris , A. , Caljouw , D. , Vamvalis , C. , A facturing: methods and applications. J. Manuf. Syst. 48, 144–156. https://doi.org/
micro-service-based machinery monitoring solution towardsrealizing the 10.1016/j.jmsy.2018.01.003
Industry 4.0 vision in a real environment, 4th International Conference on Yu , T. , Zhu , H. , Hyper-Parameters Optimization: A review of Algorithms and
Emerging Data and Industry 4.0, 184, 565–572, 10.1016/j.procs.2021.03.071, 2021. Applications, arXiv:2003.05689, 2020.

17