Space vector pulse width modulation svpwm

Artificial Intelligence Techniques
for Cyber-Physical, Digital Twin
Systems and Engineering Applications

Lecture Notes in Networks and Systems

Fernando Gomide, Department of Computer Engineering and Automation—DCA, School of Electrical and Computer Engineering—FEEC, University of Campinas—UNICAMP, São Paulo, Brazil

Okyay Kaynak, Department of Electrical and Electronic Engineering Bogazici University, Istanbul, Turkey

Jun Wang, Department of Computer Science, City University of Hong Kong, Kowloon, Hong Kong

The series “Lecture Notes in Networks and Systems” publishes the latest developments in Networks and Systems—quickly, informally and with high quality. Original research reported in proceedings and post-proceedings represents the core of LNNS.

More information about this series at

Editors
Tawfik Masrour
Department of Mathematics and Computer Science
National Graduate School for Arts and Crafts
Meknes, Morocco

Ibtissam El Hassani
Department of Industrial
and Manufacturing Engineering National Graduate School
for Arts and Crafts
Meknes, Morocco

The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

– Smart Operation Management
– Artificial Intelligence: Algorithms and Techniques
– Artificial Intelligence for Information and System Security in Industry– Artificial Intelligence for Energy
– Artificial Intelligence for Agriculture
– Artificial Intelligence for Health care
– Other Applications of Artificial Intelligence

In A2IA’2020 conference proceedings, about 141 papers were received from around the world. A total of 58 papers are selected for presentation and publication. In order to maintain a high level of quality, a blind peer review process was performed by a large international panel of qualified experts in the conference topic areas. Each submission received at least two reviews, and several received up to

– Artificial Intelligence and Industrial Applications:
Smart Operation Management (Volume 1)
In: Advances in Intelligent Systems and Computing
– Artificial Intelligence and Industrial Applications:
Artificial Intelligence Techniques for Cyber-Physical, Digital Twin Systems and Engineering Applications (Volume 2)
In: Lecture Notes in Networks and Systems

We hope that our readers will discover valuable new ideas and insights.

Tawfik Masrour Department of Mathematics and Computer

Science, Artificial Intelligence for Engineering

t.masrour@ensam.umi.ac.ma

t.masrour@umi.ac.ma

Keynotes Speakers

Amal El Fallah Seghrouchni Sorbonne University, Paris, France

Andrew Kusiak University of Iowa, USA

vii

Aalaoui Zinab, Morocco
Abawajy Jemal H., Australia
Aboulaich Rajae, Morocco
Aghezzaf El-Houssaine, Belgium Ahmadi Abdeslam, Morocco
Ait Moussa Abdelaziz, Morocco Akhrif Iatimad, Morocco
Aksasse Brahim, Morocco
Al-Mubaid Hisham, USA

Brouri Adil, Morocco
Buccafurri Francesco, Italy
Carrabs Francesco, Italy
Castillo Oscar, Mexico
Cerulli Raffaele, Italy
Chaouni Benabdellah Abla, Morocco Chaouni Benabdellah Naoual, Morocco Charkaoui Abdelkabir, Morocco
Chbihi Louhdi Mohammed Reda,

Ezziyyani Mostafa, Morocco
Faquir Sanaa, Morocco
Fassi Fihri Abdelkader
Fiore Ugo, Italy
Fouad Mohammed Amine, Morocco Gabli Mohamed, Morocco
Gaga Ahmed, Morocco
Gao Xiao-Zhi, Finland
Garza-Reyes Jose Arturo, UK
Ghobadian Abby, UK
Giuseppe Stecca, Italy
Govindan Kannan, Denmark
Grabot Bernard, France
Hajji Tarik, Morocco
Hamzane Ibrahim, Morocco
Harchli Fidaa, Morocco
Hasnaoui Moulay Lahcen, Morocco Herrera-Viedma Enrique, Spain
Itahriouan Zakaria, Morocco
Jaara El Miloud, Morocco
Jaouad Kharbach, Morocco
Jawab Fouad, Morocco
Kacprzyk Janusz, Poland
Kaya Sid Ali Kamel, Morocco
Khadija Bouzaachane, Morocco
Khireddine Mohamed Salah, Morocco Khrbach Jawad, Morocco
Kodad Mohssin, Morocco
Krause Paul, UK
Kumar Vikas, UK
Laaroussi Ahmed, Morocco
Lagrioui Ahmed, Morocco
Lasri Larbi, Morocco
Lazraq Aziz, Morocco
Lebbar Maria, Morocco
Leung Henry, Canada
Manssouri Imad, Morocco
Marcelloni Francesco, Italy
Massoud Hassania, Morocco
Medarhri Ibtissam, Morocco

Mkhida Abdelhak, Morocco
Mohiuddin Muhammad, Canada
Moumen Aniss, Morocco
Moussi Mohamed, Morocco
Najib Khalid, Morocco
Nee Andrew Y. C., Morocco
Nfaoui Elhabib, Morocco
Nguyen Ngoc Thanh, Poland
Nouari Mohammed, France
Noureddine Boutammachte, Morocco Novák Vilém, Czech
Ouazzani Jamil, Morocco
Ouerdi Noura, Morocco
Oztemel Ercan, Turkey
Palmieri Francesco, Italy
Pesch Erwin, Germany
Pincemin Sandrine, France
Rachidi Youssef, Morocco
Rahmani Amir Masoud, Iran
Raiconi Andrea, Italy
Rocha-Lona Luis, Mexico
Saadi Adil, Morocco
Sabor Jalal, Morocco
Sachenko Anatoliy, Ukraine
Sael Nawal, Morocco
Saidou Noureddine, Morocco
Sekkat Souhail, Morocco
Senhaji Salwa, Morocco
Serrhini Simohammed, Morocco
Sheta Alaa, USA
Siarry Patrick, France
Soulhi Aziz, Morocco
Staiano Antonino, Italy
Tahiri Ahmed, Morocco
Tarnowska Katarzyna, USA
Tyshchenko Oleksii K., Czech
Tzung-Pei Hong, Taiwan
Zemmouri Elmoukhtar, Morocco
Zéraï Mourad, Tunisia

Badr Abou El Majd, Morocco
Ibtissam El Hassani, Morocco
Ercan Oztemel, Turkey
Cherrafi Anass, ENSAM
El Hassani Ibtissam, ENSAM
Hajji Tarij, ENSAM
Masrour Tawfik, ENSAM
Najib Khalid, ENSMR
Saadi Adil, ENSAM
Sekkat Souhail, ENSAM
Zemmouri El Moukhtar, ENSAM

Poster Chairs

Ibtissam El Hassani, Morocco
Zemmouri El Moukhtar, Morocco

Organization xi

Ph.D. Organizing Committee

Amhraoui ElMehdi, Morocco
Benabdellah Chaouni Abla, Morocco
Eddamiri Siham, Morocco
El Mazgualdi Choumicha, Morocco
El Mekaoui Fatima Zahra, Morocco
Fattah Zakaria, Morocco
Hadi Hajar, Morocco
Jounaidi Ilyass, Morocco
Khdoudi Abdelmoula, Morocco
Moufaddal Meryam, Morocco
Raaidi Safaa, Morocco
Rhazzaf Mohamed, Morocco
Zekhnini Kamar, Morocco

Spam Filtering System Based on Nearest Neighbor Algorithms . . . . . . . 36 Ghizlane Hnini, Jamal Riffi, Mohamed Adnane Mahraz, Ali Yahyaouy,
and Hamid Tairi

A Robust Control of Permanent Magnet Synchronous Generator
Based Wind Energy Conversion System via Online-Tuned Artificial
Neural Network Compensators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 Mohsin Beniysa, Aziz El Janati El Idrissi, Adel Bouajaj,
and Mohammed Réda Britel

TRNSYS Simulation of a Solar Cooling System Under

Oujda Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

Saida Driouache, Najib Naja, and Abdellah Jamali

Toward Intelligent Solution to Identify Learner Attitude

Image Texture Component and Machine Learning Techniques . . . . . . . 119

Saloua Senhaji, Sanaa Faquir, Fidae Harchli, Hajji Tarik,

Mining Online Opinions and Reviews Using Bi-LSTM

for Reputation Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

and Reduce Dropout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

Mouhcine Sabri, Jaber El Bouhdidi, and Mohamed Yassin Chkouri

SQL Generation from Natural Language Using Supervised Learning

and Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

and an Intelligent Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

Yassin Bendriss, Youssef Hamdaoui, and Fatima Guerouate

	Dr. Tawfik Masrour				is a professor of Applied
	Mathematics and Artificial Intelligence in National
	High	School		for	the
	Engineering		Sciences		(AIES)		and	Laboratory		of
	Mathematical		Modeling,			Simulation		and	Smart
			(L2M3S).		He		graduated		from

Email: i.elhassani@umi.ac.ma

	Dr. Anass Cherrafi			Is an Assistant Professor at the
		of	Industrial		Engineering,

Method Based on the Distance Transform

Adam Hammoumi1(B), Maxime Moreaud1, Elsa Jolimaitre1, Thibaud Chevalier2, Alexey Novikov3, and Michaela Klotz3

Modeling physico-chemical phenomena in porous materials is a topic of great interest. It allows the study of real materials but also the design of new ones. Methods for flow simulation inside a porous media have been widely investigated in the last decades. The accuracy of these methods depends on a wide vari-ety of factors that can be categorized into two families: physical and chemical

⃝ Springer Nature Switzerland AG 2021
T. Masrour et al. (Eds.): A2IA 2020, LNNS 144, pp. 1–13, 2021.

segmentation that proceeds, after applying a distance transform to the void space, to remove all local maxima that lead to an oversegmented watershed [7]. In this work, a distance transform based algorithm is proposed. We extract from the distance map local maxima. A filtering operation is then applied to avoid overlapping issues. Then, we cover the void space by the geodesic distance trans-form that yields then again a discretized space of pores. Our approach takes as an input a binary image that represents a porous material microstructure. The binary aspect of the image refers to the co-existence of both solid and porosity space. Our algorithm is called the Pore Network Partitioning method (PNP) and it aims at treating a wide spectrum of microstructures.

2 Method

DTd(x) = min y̸=Pd(x, y) x, y ∈ E (4)

where the related n-dimensional metric space is denoted (E, d), X is the solid space and P = E \ X is the porosity space. Given a square orthogonal grid in E, the used algorithm throughout this paper is the raster scanning algorithm which is well established in the literature [16]. It consists of a forward and a backward pass. The two passes algorithm ensures obtaining the correct distance to the nearest foreground elements. This technique is widely used for the problem of connected component labeling [17]. Algorithms based on distance transform can produce Voronoi Diagrams and Delaunay triangulation [18]. In Fig. (1-a), we show side by side the original image and its corresponding distance transform. We obtain a digital representation of the porosity space, where the intensity of the white color of background elements is inversely proportional to their proximity to foreground elements. The distance transform can also yield a skeletonization transform – an operation in image processing that simplifies an object while retaining its topology – [20]. The accuracy of the distance transform depends on whether we use exact or approximate transformations.

Where point x′In Fig. (1-b), the dots correspond to the local extracted maxima. kbelongs to the neighborhood V(xk).

2.3 Maxima Filtering

(c)

(d)

A subsequent operation of creating disks around the remaining points is then applied. In Fig. (2), the center of the black disk is included into a bigger disk. Applying the filter will remove the associated maximum point and its disk accordingly. Considering intersection of disks, two filtering tecnhiques are intro-duced: standard filtering, shown in Fig. (2-a) and filtering with intersections removal, shown in Fig. (2-b). We define the partition function M and F for each filtering technique:

M(x) =⎧⎨⎪ F(x) = rk

(9)

M and F refer to standard filtering and filtering with intersections removal respectively. For M, we keep the intersections between pores. Whereas, for the other technique, we create intermediate pores by removing the intersections. The interest of these filtering methods will be explored further in what follows. Figures (1-c) and (d) illustrate the obtained results before and after the standard filtering operation. For the 3D case, all the operations described above remain the same, replacing disks with spheres.

Filtering with intersections removal

Efficient Pore Network Extraction Method Based on the Distance Transform 7

Fig. 3. The geodesic distance between P and Q is given by the red path, whereas the

Euclidean distance is the length of the blue path.

Fig. 4. Illustrations of the two types of local maxima filtering applied to a 2D image of

Our algorithm allows pore network extraction on images of 2003voxels in 1.6 s and bigger images of 5003in 29 s. The computations were performed using a personal computer (CPU: intel core i7 2.6 GHz, RAM: 16 GB). For the sake of

Efficient Pore Network Extraction Method Based on the Distance Transform 9

Fig. 6. Illustrations of the extracted pore networks by the PNP method. (a) 3D Pore network of a multi-scale Boolean model of platelets. (b) Section of the pore network (a) with partition labeling.

Consider the family of sets (each set is made of pores), and every set has a unique label (pore diameter ≃ DTd). We consider the family of sets {Lr} such that r = 1, 2, ....n and n = max(DTd). A Granulometry function is obtained with:
Gr= Cardinal(Lr) (12)

The pore size distribution ‘PSD’ is the cumulative sum of the granulometry function values and is defined by:

A new algorithm for pore network extraction has been described. The algorithm relies on simple and well-known methods in image processing. The method is

12 A. Hammoumi et al.

Rep. 5, 10635 (2015)
6. Pudney, C.: Distance-based skeletonization of 3D images. In: Proceedings of the 1996 IEEE TENCON Digital Signal Processing Applications, TENCON 1996, vol.

201, pp. 209–214 (1996)
7. Gostick, J.T.: Versatile and efficient pore network extraction method using marker- based watershed segmentation. Phys. Rev. E 96, 023307 (2017)
8. Zeng, Y., Fan, C., Do, D.D., Nicholson, D.: Evaporation from an ink-bottle pore: mechanisms of adsorption and desorption. Ind. Eng. Chem. Res. 53(40), 15467– 15474 (2014)
9. Xiong, Q., Baychev, T.G., Jivkov, A.P.: Review of pore network modelling of porous media: experimental characterisations, network constructions and appli- cations to reactive transport. J. Contam. Hydrol. 192, 101–117 (2016)
10. Jeulin, D.: Morphology and effective properties of multi-scale random sets: a review. Comptes Rendus M´ecanique 340(4–5), 219–229 (2012)
11. Moreaud, M., Chaniot, J., Fournel, T., Becker, J.M., Sorbier, L.: Multi-scale stochastic morphological models for 3D complex microstructures. In: IEEE Con- ference on 17th Workshop on Information Optics (WIO), Quebec city, Canada (2018)
12. Maurer Jr., C.R., Qi, R., Raghavan, V.: A linear time algorithm for computing exact Euclidean distance transforms of binary images in arbitrary dimensions.

Springer, Berlin (2004)
15. Borgefors, G.: Distance transformations in arbitrary dimensions. Comput. Vis.

Graph. Image Process. 27(3), 321–345 (1984)
16. Shih, F.Y., Wu, Y.-T.: Fast Euclidean distance transformation in two scans using a 3x3 neighborhood. Comput. Vis. Image Underst. 93(2), 109–205 (2004) 17. He, L., Chao, Y., Suzuki, K.: A run-based two-scan labeling algorithm. In: Kamel, M., Campilho, A. (eds.) Image Analysis and Recognition, ICIAR 2007. LNCS, vol.

1 LIIAN Laboratory, Computer Science Department, Faculty of Sciences Dhar Mehraz, Sidi Mohamed Ben Abdellah University, Fes, Morocco chahinaze.am@gmail.com, ayahyaouy@yahoo.fr
2 Faculty of Sciences Engineering, Fes Private University, Fes, Morocco sanaa.faquir@usmba.ac.ma

Powering remote sites with renewable sources require the use of storage devices due to the fluctuating nature of the power production. In this paper, a novel Multi-Agent-System strategy is developed to perform multi-function strategy for smooth

© Springer Nature Switzerland AG 2021
T. Masrour et al. (Eds.): A2IA 2020, LNNS 144, pp. 14–25, 2021.

Energy management for HRES was done before using fuzzy logic [7, 8], neural network detailed in [9] and genetic algorithm in [10]. The results obtained after applying the fuzzy modeling showed good estimation of the PV and WT powers and the proposed method in [10] has been applied to the analysis of a hybrid system which supplies power for a telecommunication relay station, and good optimization perfor-mance has been found.

Many papers are reported in literature which deals with system control for energy flow management of HRES. In [11] stabilization problems in the integrated system using the intelligent fuzzy logic controller on the basis of flatness property for DC grid voltage regulation is discussed. In [12] an optimum configuration and dispatch strategies in solar-wind based hybrid system is presented. In [13] PI/PID controllers to regulate the output power from the sources and demand under varying condition of load and generation and this reduced the frequency deviation is used. In [14] authors, proposed a strategy that was compared with a conventional strategy. They minimized number of change over between FC and battery with the help of measurement and time delay elements. In [15], first fuel cell system supplied the deficit power to load and later UC bank satisfied the remaining energy for short duration that cannot be fulfilled by the FC system. In [16] transient analysis of a self-excited induction generator with elec-tronic load controller for stand-alone applications is presented. Analysis includes the effect of switching of loads on dump power, load power and generated power. In [17] authors proposed three power management strategies (PMS) and compared them on the basis of sensitivity analysis, considering state of charge of batteries and output power from FC. They also observed the effect of these PMS on lifetime of FC and electrol-yser. In [18] based on the battery storage energy, authors proposed six operating points

The HRES used in this research is a stand-alone hybrid PV/Wind/Battery system that should supply the electricity to a private house or an apartment in an isolated site. Solar and wind energies were combined in this system as they have the advantage of complementing each other. To implement an HRES on an isolated site, a study was done to be able to determine the required input and output sources required for the system to satisfy a load demand besides the changeable weather conditions: the system requires a combination in series of 16 PV panels composed of 36 cells to provide 1 kWc of maximum energy per day, along with a generic wind turbine with a rated power 1 kW peak under STC conditions (Temperature 25 °C and lighting 300 W/m2). The maximum power delivered by the system is 2 kW. The system also includes a number of batteries to either store the excess of energy generated by the sources or provide the energy demanded by the load when there is low renewable energy produced [24] (Fig. 1).

The hybrid system used in this paper is composed of:

• A wind turbine connected to the DC bus using static converters.

18 C. Ameur et al.

Fig. 4. Daily load demand variation

The SOC represents the second parameter and is calculated at each period of time t (1 h) by the following formula:

SOC ¼ Pbat=BC ð2Þ

NewBat ¼ Batterie þ Ppv þ Pw � Pload ð5Þ

20 C. Ameur et al.

Fig. 6. Block diagram of the Fuzzy controller applied to the HRES to manage the flow of energy

The figures below show the membership functions of the input and output vari-ables. The membership function for the battery Soc is normalized and represented by the following chart (Fig. 7):

Fig. 8. Membership function for the output variable Battery Status

The proposed multi-agent structure is composed of a cognitive agent called“Supervisor Agent”, several other cognitive agents, each called “WT Agent” “PV Agent” “Load Agent” and “Battery Agent”.

(a) Supervisory Agent: This is a unifying agent allowing:
(1) Receive and calculate the sum of the powers generated by turbine and PV panels.

(5) Take the necessary decisions in case a turbine and PV panels agent is not able to provide the power.

(b) WT Agent/PV Agent: the role of each WT/PV is: (1) Calculate the power pro- duced by the WT/PV.

The two methods were tested with batteries filled at their minimum level and other batteries filled to their maximum level as follows:
Empty batteries: where the battery starts with its lowest value (Pbat = Pbmin = 1200 W). The variation of the batteries level for one day is represented by the graph below: Full batteries: where the battery starts with its maximum value (Pbat = Pbmax = 3200 W). The variation of the batteries level for one day is represented by the graph below: The chart shows that the batterys level started with its maximum level (3200 W). Then it was discharged until the minimum level (1200 W) was reached. The batteries then remained constant at their minimum until they were charged again. Once the batteries level reached their maximum, it remained constant. As it is shown from

A Study of Energy Reduction Strategies in Renewable Hybrid Grid 23

Fig. 12. Battery power curve obtained from the multi-agent in case of a system starting with a full battery

References

1. Koua, B.K., Koffi, P.M.E., Gbaha, P., Tour, S.: Present status and overview of potential of renewable energy in Cote d’Ivoire. Renew. Sustain. Energy Rev. 41, 907–914 (2015) 2. Zadeh, L.: The concept of a linguistic variable and its application to approximate reasoning I.

photovoltaic/diesel hybrid generating system considering fluctuation of solar radiation.

Sol. Energy Mater. Solar Cells 67(14), 535–542 (2001)
19. Ameur, C., Faquir, S., Yahyaouy, A.: Intelligent optimization and management system for renewable energy systems using multi-agent. In: 2018 Modern Intelligent Systems Concepts (MISC), vol. 8(4), December 2018
20. Ameur, C., Faquir, S., Yahyaouy, A.: Intelligent optimization and management system for renewable energy systems using multi-agent. IAES Int. J. Artif. Intell. 8(4), 352–359 (2019).

DFS often uses data replication and distribution on several machines to ensure their durability, optimize paralleled queries, and enable data recovery in case of machine failure. The data placement on machines of a cluster generally respects a strategy that promotes data security (having replicas on different machines, racks, sites …). However, it often ignores the optimization of response times, which depends on the capacities of storage machines, the network and the intensity of use of this data.

In this study, we propose to predict response time of a DFS Cluster, by taking into account a set of criteria such as locations, file sizes …, using the techniques of “Deep Learning”.

Because of the large size of those systems, response times are highly affected by the architecture of the cluster and the performance of used hardware but also by the data placement strategy used by the DFS.

The choice of nodes involved in the replication of stored data has a big impact on the response time. Placing a high-demanded file in a network of nodes that does not have the best resources will seriously impact read/write times [3, 4]. Improving

In this study, a model for predicting the processing time in a cluster is proposed, using Deep Learning techniques. This model can be improved to increase the accuracy of response time predictions by periodically feeding the learning dataset by new metrics.

The mean objective of this work is to give an answer to the following question: what is the time required to execute a job on a node Y that reads an entire file of a size T, residing in a node X.

2 State of the Art

Several research works have proposed to calculate the response time in a cluster, either by relying on mathematical formulas using technical characteristics of the hardware components such as the speed of the processors, the bandwidth of the routers and gateways, the latency [5–7] … or by trying to predict it based on the processing history collected from system logs.

3 Simulation Description

OptorSim is a Data Grid simulator written in Java and designed to enable experiments and evaluations on various replication strategies and job execution scenarios, in dis-tributed environments like distributed file systems (DFS).

The Resources Broker is responsible for submitting jobs to the Computing Entities, which are responsible for performing data file jobs on the Cluster sites.

The Storage Entities are responsible for storing the data used by jobs.

DFS Response Time Prediction 29

– Site: composed by calculation elements (CE) and storage elements (SE), it allows to define the basic components of the Cluster and its global architecture by defining the hardware capabilities (computing power expressed in kSI20001.

To simulate the operations performed on the cluster, files was stored on different cluster nodes, and jobs calling those files was executed on other nodes.

In each run, three jobs was executed involving the same nodes and files and the mean execution time corresponding to the three jobs was collected as “The mean job time” (MJT).

1 kSI2000 is the unit for 1000 of the CINT2000 measurement unit, an Intel P4 Xeon at 2.8 GHz is approximately 1 kSI2000, see ).

30 A. Elomari et al.

The storage entity IDs and calculation entity IDs (SE ID and CE ID) is representing the physical locations in the network of entities. In fact, instead of integrating the network speed, the number of gateways and the other physical characteristics of the network, we have included the identifiers of the nodes, which is in this case an illus-trative factor of the response times.

4 Deep Learning Model

To implement our “Deep Learning” model, a sequential model was created using two highly dense neural networks layers as shown in Fig. 3.

A third layer returns a continuous and unique value which is the mean job pro-cessing time (Mean Job Time).

– ID SE: ID of the storage entity where the file is stored in the Cluster
– ID CE: ID of the computing entity that uses the file in the Cluster and where jobs are executed.

– File Size: the size of the file (in Mega Byte)
– Power: Power of the calculation unit CE in kSI2000
– MJT: Mean execution time (in milliseconds) of jobs executed on CE, using a file from SE (three jobs consuming the same file and running on the same CE sequentially was lunched, MJT is the sum of the times of execution of those Jobs, divided by 3).

1) It can accept multiple inputs (i.e., multivariate features) and use all these properties to generate the prediction.

2) RMSProp does not engage in a parametric form, unlike formula-based approaches. Instead, the RMSProp templates are very flexible.

5 Evaluation of Deep Learning Model Results

The Mean Squared Error and the Mean Absolute Error (Fig. 4 and 5) was used to evaluate the model loss.

Where yi is the value and ŷi is the corresponding prediction, calculated by the model.

DFS Response Time Prediction 33

varying the network bandwidth occupation in the simulation, this can perfectly explain this value.

Fig. 5. Evaluation of the mean absolute error

Fig. 6. Prediction simulation

6 Conclusion

Thus, we have proved that the response time of a DFS cluster can be predicted using a deep learning model.

In our future works, this model will be used to define a better file placement strategy for a DFS like HDFS. By predicting response time, it will be possible to reject or accept the proposed data placement options, in order to respect response time criteria predefined by the users of the DFS.

Comput. Appl. 17(4), 403–416 (2003)
9. What is machine learning.
10. da Silva, I.N., Hernane Spatti, D., Andrade Flauzino, R., Liboni, L.H.B., dos Reis Alves, S.F.: Artificial Neural Networks. Springer, Cham (2017).
11. Yazan, E., Talu, M.F.: Comparison of the stochastic gradient descent based optimization techniques. In: International Artificial Intelligence and Data Processing Symposium (IDAP) (2017)
12. Sai, Y., Jinxia, R., Zhongxia, L.: Learning of neural networks based on weighted mean squares error function. In: Second International Symposium on Computational Intelligence and Design (2009)
13. Mendo, L.: Estimation of a probability with guaranteed normalized mean absolute error.

IEEE Commun. Lett. 13(11), 817–819 (2009)

Keywords: K-NN � WKNN � K-d tree � Machine learning � Feature extraction

1 Introduction

Spam Filtering System Based on Nearest Neighbor Algorithms 37

the email’s vector of all emails. Secondly, classify the emails using Nearest Neighbors (NN) classifiers such as: K-Nearest Neighbors (KNN) [7], Weighted KNN (WKNN) [8] and K-d tree [9].

These techniques are not learning-based, as they have advantages they also suffer from many problems. In the first, Blacklisting and whitelisting required constructing the lists of IP addresses and domains and keeping them up to date, which surely causes a waste of time. In the second time, all these techniques do not take into account the context and the semantic of the email.

Filtering may also be done by more complicated and accurate filtering techniques to deal with the spam phenomenon, some of them are based on Bayesian decision or some others machine learning techniques which are considered as learning-based technique. In the rest of this part, we present an overview of some used learning-based techniques and some works in text mining and classification tasks.

Decision Trees [14], are composed of a hierarchical structure in the form of a tree. It works well to solve some difficult and significant problems. Quinlan [15] develops decision trees for classification tasks. Each tree is processed as follow, starting from the root and down to its leaves.

In the next section, we detail our system architecture based on three known tech-niques namely Bag-Of-Words [5], Term Frequency-Inverse Document Frequency [4] and Word N-Gram [6], then show and discuss the results obtained.

a. Pre-processing
Pre-processing is an essential element that aims to clean all data (emails) from any word that does not make sense in order to increase the performance. Pre-processing is a set of implementing a series of processes. As shown in Fig. 2, Tokenization is one of the primordial process used in Natural Language Processing (NLP), aims at trans-forming the entire text into a list of words (terms). Moreover, Elimination of capital letters transforms all capital letters to lowercase; indeed the word spam and SPAM will be looked different even though they have the same meaning. In addition, removing stop words is a significant technique used in classification task and text mining, so the main intention behind this technique is to remove all the words that are repeated frequently in all documents, because they do not have any effect in the process of text

Fig. 2. Pre-processing step

processed emails

Fig. 3. Feature extraction step

b. Feature Extraction
Feature extraction [16] is the most essential technique to achieve a high performance in classification task. The choice of feature extraction methods has a major impact in attaining a high performance. In this paper, we focus on using the three known methods specifically Bag-of-words (BOW) [5], N-gram [6] and Term Frequency-Inverse Doc-ument Frequency (TF-IDF) [4]. In the following, we give details of how each method works:

information, which has an impact in improving the performance of the system. In our experiments we use word n-gram with n ¼ 2.

3. Term Frequency-Inverse Document Frequency (TF-IDF)
TF-IDF [4] is based on term weighting. Otherwise, TF-IDF calculates the relevance of a given term in a particular email compared to the inverse proportion frequency of that term over the entire email corpus. The weight wi;j of each term i in email j is calculated according the following rule [4]:

1. KNN Classifier
K-Nearest Neighbor (KNN) [7] classifier is non-parametric and lazy learning instance-based method. The KNN classifier saves instances of training data which employed in

test phase [17]. Notice T ¼ corresponding class label. To classify a new incoming email x0, KNN algorithm looks f ð xi; ci Þ gQ i¼1the collection of training emails where ci the

Where c is a class label, cNN i

nearest neighbors. d c ¼ cNN and zero otherwise.�

• Email Dataset

This comparison is done across two different dataset LingSpam and Enron. In all our experiments, we use 70% for training and the remaining for test.

42 G. Hnini et al.

cross validation. The k-fold cross validation involves splitting the training set into k

Fig. 5. Varying k using TF-IDF

Fig. 7. Varying k using BOW

Spam Filtering System Based on Nearest Neighbor Algorithms 43

The figures (Fig. 10, Fig. 11, Fig. 12, Fig. 13, Fig. 14 and Fig. 15) present the

Fig. 11. ROC curve using TF-IDF (k = 43)

Fig. 13. ROC curve using BOW (k = 1)

44 G. Hnini et al.

same datasets. The last representation method TF-IDF has showed a high performance of 98% and 90%.

The Table 1, Table 2 and Table 3 present the performance measures results on LingSpam dataset. We can see that KNN, WKNN and K-d tree classifiers have the same results in term of Precision, Recall, F1-score and Accuracy.

Using BOW representation, we obtained an accuracy of 95,04% which is less than N-Gram representation. Furthermore, TF-IDF achieved a high accuracy of 98,38%, another essential remark the Precision, Recall and F1-score are identical.

Spam Filtering System Based on Nearest Neighbor Algorithms 45

Table 3. Performance measures using N-Gram (k = 3)

• Enron dataset

In this paper, we compared the three nearest neighbor methods namely: KNN, WKNN and K-d tree for spam classification task in both Enron and LingSpam datasets. We have seen that KNN classifier gives good results when we apply TF-IDF contrary to BOW. In addition, the decrease in performance is caused by the size of data, in other words if we have a small dataset we can get a good performance, but in the other case make training faster and testing phase slower and costlier in time and memory.

46 G. Hnini et al.

J. Inf. Comput. Sci 9(6), 1429–1436 (2012)
18. Hechenbichler, K., Schliep, K.: Weighted k-Nearest-Neighbor Techniques and Ordinal Classification, vol. 399 (2004). Projekt partner Weighted k -Nearest-Neighbor Techniques 19. Klimt, B., Yang, Y.: The Enron Corpus: A New Dataset for Email Classification Research.

In: Boulicaut, J.-F., Esposito, F., Giannotti, F., Pedreschi, D. (eds.) ECML 2004. LNCS (LNAI), vol. 3201, pp. 217–226. Springer, Heidelberg (2004).
20. Androutsopoulos, I., Koutsias, J., Chandrinos, K.V., Paliouras, G., Spyropoulos, C.D.: An Evaluation of Naive Bayesian Anti-Spam Filtering (2000)

Compensators

Mohsin Beniysa(B)	, Aziz El Janati El Idrissi, Adel Bouajaj
and Mohammed R´eda Britel

Keywords: Parametric uncertainty · Mechanical torque disturbances ·Artificial neural network compensators · DTC · PMSG · Wind turbine

1 Introduction

In recent years, artificial neural network (ANN) has became an efficient method for controlling linear and nonlinear systems, due to its high accuracy and robustness against internal and external disturbances [19–21]. Extended versions of ANN control techniques have been found to be the most popular and conventional tools in control systems engineering, identification and func-tional approximations, such as those presented in [23–26]. The ANN control has a strong capability of handling uncertain information, and can be easily used in the control of systems that are too complex to have an exact mathematical model. Actually, ANN controller, ANN estimator and ANN predictor have been developed both in continuous and discrete-time. The techniques implemented in discrete-time have the advantage of being able to be directly implemented in digital hardware [27–29].

In this paper, the objective is to use a dynamic inversion approach to con-struct two online-tuned artificial neural network compensators (ANNCs), which can adapt their weights dynamically to any uncertainty and external disturbance in PMSG drives. In order to validate the reliability of the proposed approach, several desired input references for the wind speed are chosen for respective simulation tests. The results are discussed and compared to those obtained by classical direct torque control.

The two equations below show the kinetic power Pw captured by the turbine, and the resulted mechanical one Pm that is transmitted to the PMSG, respec-tively [17].

The mechanical torque Tm can be expressed by the quotient of the power trans-mitted to PMSG by its mechanical rotor speed, as follows [17].

x8(k + 1) = x8(k) + x6(k)Ts

Where the parameters defined by cap depict parameter actual values, the tiled ones depict individual errors and the parameters without accent depict constant (nominal) values. Substituting nominal parameters in Eq. (4) by those caped in Eq. (5), we obtain the model of uncertain PMSG. In the case of clas-sical controllers, the uncertainties can destroy the control of PMSG based wind turbine. As a solution, a new control scheme based on two ANNCs is proposed.

A Robust Control of PMSG Based WECS 51

the errors due to this uncertainty two artificial neural network backlash com-pensators are added to the current loops. More details about design procedure of the proposed control strategy are presented in the Fig. 1.

The dynamic model results of the proposed system are simulated using the Dev-C++ software and plotted with Matlab. In this simulation, the rectifier is simulated by the ideal switching frequency 20 KHz. The generator is represented by its dynamic model in the Park frame. Thus, after several simulation tests, the average execution time of our program is around 0.012 s for 10000 simulation cycles, i.e., the execution time needed for one cycle is around 1.2 µs. Bearing in mind that in practice, the sensors, the analog/digital converters, the switching elements of the rectifier and the digital signal processing (DSP) algorithm are time consuming, it is practically difficult to achieve such system with small sampling period. Thus, convenient sampling periods, such as 100 µs or larger are normally selected for processing. So, to make the proposed adaptive method feasible for the PMSG with the parameters shown below, a simulation with a sampling time of Ts = 100 µs is executed, meaning that for 10000 samples the time of simulation is T= 1 s.

In this study, the values of PMSG parameters are given in Table 1 [17]. The two gains Kp = 0, 3669 and KI = 88, 5612, of PI controller are calculated accord-ing to the expressions of [15]. In order to better show the effectiveness of the proposed method, the parameters of the two backlash actuators are considered as follows: md = 0.5, mq = 0.55, dd+ = 0.2, dd− = 0.2, dq+ = 0.2 and dq− = 0.2.

indicated above, where the wind speed is a step (Fig. 2).

Vw (m.s-1)	13	A Robust Control of PMSG Based WECS										53
	12
	11
	10
	9
	8
	7
	6
	5
	4
			0.1	0.2	0.3	0.4	0.5	0.6	0.7	0.8	0.9		1

(a)

150

Time (s)

Fig. 3. Speed response under mechanical torque and parametric uncertainties.

DTC

Time (s)

Fig. 4. Tracking errors of the speeds.

100

Time (s)

Fig. 5. Mechanical power evolution.

0.5

0.4

Adaptive method

Since the speed is designed with a step speed reference under parametric uncer-tainties and mechanical torque disturbances, a robustness evaluation of the pro-posed technique is performed by applying a random wind speed reference (Fig. 7), under high frequency fluctuations, with the same conditions indicated above. The dynamic reference and actual speed responses related to the proposed adaptive method and DTC are given in Fig. 8. It is worth noting that the disturbance rejection capacity of the proposed adaptive method under parameter uncertain-ties and mechanical torque disturbances leads to a good speed tracking perfor-mance. In addition, the asymptotic speed tracking objective is obtained with a good accuracy.

200

150

Time (s)

Fig. 8. Speed response under parametric uncertainties and mechanical torque distur-

Time (s)

0.4

0.35

Time (s)

Fig. 10. Power coefficient Cp.

1. Sumathi, S., Ashok Kumar, L., Surekha, P.: Solar PV and Wind Energy Conver- sion Systems: An Introduction to Theory, Modeling with MATLAB/SIMULINK, and the Role of Soft Computing Techniques. Springer International Publishing, Switzerland (2015)
2. Quang, N.P., Dittrich, J.-A.: Vector Control of Three-Phase AC Machines.

Springer, Heidelberg (2008)
3. Bodson, M., Chiasson, N., Novatnak, R.T., Rekowski, R.: High-perfomance nonlin- ear feedback control of a permanent magnet stepper motor. IEEE Trans. Control Syst. Technol. 1(1), 5–13 (1993)
4. Zhang, Z., Zhao, Y., Qiao, W., Qu, L.: Space-vector-modulated sensorless direct- torque control for direct-drive PMSG wind turbines. IEEE Trans. Ind. Appl. 50(4), 2330–2341 (2014)
5. Kim, H.-W., Kim, S.-S., Ko, H.-S.: Modeling and control of PMSG-based variable- speed wind turbine. Electr. Power Syst. Res. 80, 46–52 (2010)
6. Li, S., Haskew, T.A., Ling, X.: Conventional and novel control designs for direct driven PMSG wind turbines. Electr. Power Syst. Res. 80, 328–338 (2010) 7. Lee, S.-H., Joo, Y., Back, J., Seo, J.-H., Choy, I.: Sliding mode controller for torque and pitch control of PMSG wind power systems. J. Power Electron. 11(3), 342–349 (2011)
8. El Magri, A., Giri, F., El Fadili, A., Dugard, L.: Adaptive nonlinear control of wind energy conversion system with PMS generator. In: 11th IFAC International Workshop on Adaptation and Learning in Control and Signal Processing, Caen, France, July 3–5, pp. 318–325 (2013)
9. Youness, E.M., et al.: Implementation and validation of backstepping control for PMSG wind turbine using dSPACE controller board. Energy Rep. 5, 807–821 (2019)
10. Yang, B., et al.: Passivity-based sliding-mode control design for optimal power extraction of a PMSG based variable speed wind turbine. Renew. Energy (in press)

60 M. Beniysa et al.

28. Lewis, F.L.: Neural Network Control of Nonlinear Discrete-Time Systems. CRC Press, Taylor & Francis Group, Boca Raton (2006)
29. Sanchez, E.N., Ornelas-Tellez, F.: Discrete-Time Inverse Optimal Control for Non- linear Systems. CRC Press, Taylor & Francis Group, Boca Raton (2013)

In order to contribute to the classification and the security of Medical Images (storage, sharing, transfer …), we present, in this paper:

• a comparative study between the famous Convolutional Neural Network architectures and ACSA-Net that we propose concerning the classification of Brain Tumors detected in the Magnetic Reso-nance Imaging.

62 Y. Douzi et al.

1 Introduction

In the past, researchers relied on Magnetic Resonance Imaging (MRI) [8] to detect brain tumors and brain progression. According to National Institute of Biomedical Imagine and Bioengineering (USA) a medical imaging technique that provides two or three dimensional views of the interior of the body in a non-invasive manner with relatively high contrast resolution [9].

Classification and Watermarking of Brain Tumor 63

The structure of this paper is organized as follows: in Sect. 2 there’s an overview on the deep learning concepts and architectures, Sect. 3 described the steps of the proposed methodology, Sect. 4 presents the experimental results and the conclusion and future work is given in Sect. 5.

2 Related Work

AlexNet (2012) – In 2012, Alex Krizhevsky (and others) [17] released AlexNet which was a deeper and much wider version of the LeNet and won by a large margin the difficult ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2012. It was a significant breakthrough with respect to the previous approaches and the current widespread application of CNNs can be attributed to this work [14].

ZF Net (2013) – The ILSVRC 2013 winner was a Convolutional Network from Matthew Zeiler and Rob Fergus. It became known as the ZFNet [18] (short for Zeiler & Fergus Net). It was an improvement on AlexNet by tweaking the architecture hyperparameters [14].

In 2018, Heba Mohsen and all [10] presented a methodology which combines the discrete wavelet transform with the Deep Neural Network to classify the brain MRIs into Normal and 3 types of malignant brain tumors: glioblastoma, sarcoma and metastatic bronchogenic carcinoma. The methodology architecture resemble the CNN architecture but requires less hardware specifications and takes a convenient time of processing for large size images (256 � 256), in addition the classifier shows high accuracy compared to traditional classifiers.

And in June 7, 2019 a 3-dimensional (3-D) CNN called HeadXNet was developed by Allison Park and a group of researchers. Its role is the segmentation of intracranial aneurysms from computed tomographic angiography CTA scans [22]. This model of the deep learning successfully detected clinically significant intracranial aneurysms on CTA. As a conclusion, the integration of an artificial intelligence–assisted diagnostic

3.1 CNN Architectures -Comparative Study-

Our methodology starts by a comparative study of these convolutional architectures: LeNet, Alex Net, ZF Net, Google Net, VGG Net, Res Net and Dense Net. Then we propose our CNN architecture to brain tumors detection and classification. This is the presentation of our approach plan:

66 Y. Douzi et al.

As a project, we designed a Java Maven named ACSA-Net (Fig. 4) to manage our CNNs; ACSA-Net can be used to directly generate all CNNs architectures used in this study and to generate a specific CNNs architecture.

Classification and Watermarking of Brain Tumor 67

68 Y. Douzi et al.

To implement our CNN, DL4J was used. As well, the same dataset that presented in the comparative study and the same machine performances to build and run it. And finally we compared the result of training and recognition with the CNNs used in our comparative study.

The Fig. 8 show the propagation of one input x into the ACSA-Net layers and the result was the last layer 81 features.

Table 1. The evolution of the quantity of data in the 4 layers of the ACSA-Net

Table 2. The filters used in each convolution layer

3.2 Compression and Digital Watermarking Signature

The result of an ANN compression operation and a digital tattoo algorithm were used in this new method to inject the signature mark into the original image. In this case, the image tablet was used by an ANN as a signature mark. First, a comparative study was conducted to select the most effective ANN architecture for image compression and restoration without much loss of information. Several ANN have been instantiated with different architectures, same learning base, same quadratic error and even learning algorithm (reverse propagation of gradient). Then, we grouped the results in the table. Finally, we concluded this study by selecting ANN (64, 16, 64) as the most effective for image compression and restoration. Figure 9 presents an operation to compress and restore an image by ANN (64, 16, 64). In the middle of Fig. 9, we have the signature mark that is injected into the original image by a tattoo algorithm:

Note that the same ANN is used in two modes: compression and restoration as explained in Fig. 10.

Classification and Watermarking of Brain Tumor 71

Fig. 11. Compression signature process

And to check the signature, we did the reverse. The signature mark was extracted via the reverse of the tattoo algorithm. Then, for each 4-pixel block, the corresponding 64-pixel block was restored through the restoration portion of the ANN. Figure 12 shows the principle of verification.

Fig. 13. Principle of approach

We presented the results of our experience in the Table 3: we use the following colors

74 Y. Douzi et al.

4.2 Compression and Digital Tattoo Signature

• The rate of convergence (restoration) after the test is 30% (70% information is lost).• Table 4 shows that the convergence rate is directly related to the structure of the

Classification and Watermarking of Brain Tumor 75

Table 4. Results of the ANN comparative study to select the best architecture which retains the maximum information at the time of restoration

5 Conclusion

We started this work with a comparative study between the most popular deep learning architecture, then, we proposed a more flexible architecture that gives a perfect clas-sification for the timers studied.

References

1. Douzi, Y., Kannouf, N., Hajji, T., Boukhana, T., Benabdellah, M., Azizi, A.: Recognition textures of the tumors of the medical pictures by neural networks. J. Eng. Appl. Sci. 13, 4020–4024 (2018)
2. Hajji, T., El Jasouli, S.Y., Mbarki, J., Jaara, E.M.: Microfinance risk analysis using the business intelligence. In: 2016 4th IEEE International Colloquium on Information Science and Technology (CiSt), pp. 675–680. IEEE (2016)
3. Hajji, T., Jamil, O.M.: Rating microfinance products consumers using artificial neural networks. In: Rocha, Á., Serrhini, M. (eds.) EMENA-ISTL 2018. SIST, vol. 111, pp. 460– 470. Springer, Cham (2019).
4. Tarik, H., Ouerdi, N.: EMV cards vulnerabilities detection using ANN. In: 2016 International Conference on Information Technology for Organizations Development (IT4OD), pp. 1–5. IEEE (2016)
5. Ouerdi, N., Hajji, T., Palisse, A., Lanet, J.-L., Azizi, A.: Classification of ransomware based on artificial neural networks. In: Rocha, Á., Serrhini, M. (eds.) EMENA-ISTL 2018. SIST, vol. 111, pp. 384–392. Springer, Cham (2019).
6. Hajji, T., Ouerdi, N., Azizi, A., Azizi, M.: EMV cards vulnerabilities detection using deterministic finite automaton. Procedia Comput. Sci. 127, 531–538 (2018)
7. American Association of Neurological Surgeons, © 2020, i
8. Khambhata, K.G., Panchal, S.R.: Multiclass classification of brain tumor in MR images. Int.

Machine Learning Application for Blood

Pressure Telemonitoring over Wireless

Abstract. Human blood pressure is a health indicator related to var-ious chronic diseases. With the development of telemedicine applica-tions, innovative devices are being marketed to monitor remotely human health and to increase more autonomy and wellbeing. The application of machine learning algorithms for continuous blood pressure (BP) monitor-ing is a viable method for analyzing, predicting and classifying BP. In this paper, we applied the decision tree classifier (DT) for BP classification based on human data collected from the Wireless Sensor Network (WSN) combined to the IoT network (Internet of Things). Cross-validation is used to choose the best parameters for the DT model construction to avoid over-fitting. Obtained results show high accuracy and good per-formances by calculating the mean absolute error. The DT model was embedded in the first layer of the multi-levels architecture of WSN, where BP is measured. The second layer performs data clouding to be visual-ized in the third layer which is an IoT platform. When an anomaly is detected by the DT model, the WSN & I alert health center. Experi-mental results are provided to ensure the effectiveness of the proposed approach in terms of collection, classification, transmission, and super-vision in real-time.

Keywords: Machine learning · Wireless Sensors Network · Internet of Things · Health telemonitoring · Blood pressure

The machine learning (ML) can be classified into (a) - supervised learn-ing, where inputs and outputs data are known as the case of the decision tree algorithm; (b) - unsupervised learning where there are no labeled outputs and learning is done according to the inherent structure of the data itself [2]. DT is a supervised learning algorithm used widely in literature in order to classify, predict or decide patients’ disease status according to a set of attributes such as the major health information, including heart rate HR, blood saturation with oxygen SPO2, blood pressure BP, body temperature BT, etc [3].

In the last decade, blood pressure measurement labor as a conventional tech-nique for health professionals to obtain information on patients’ condition sta-tus. First, blood pressure is the force generated when the heart pushes the blood against the walls of the body’s arteries. Its represented by two information, Sys-tolic blood pressure that occurs while heartbeats or contracts, and Diastolic blood pressure which related to the rest between heart’s beats [4]. The abnor-malities in the blood pressure are created when systolic BP and diastolic BP exceeds normal classes as Table 1 shows (Sect. 4) [5].

This paper describes an approach contemplated of the method proposed by [5], where machine learning algorithms and the Internet of Things (IoT) tech-nology are deployed using the Wireless Sensor Network to follow-up patients in the need of blood pressure control remote of the health center. A decision tree was used to analyze the human BP collected continuously by a Body Sensor Node (BSN) and carry out BP prevision. According to the output status of the DT algorithm, the BSN decides on transmission mode. The second layer of WSN performs data clouding collected from BSN of blood pressure. Then, data can be redeemed in real-time and continuously into an IoT platform terminal accessible by health professionals and patients.

This paper is organized into four sections. Section 2 presents the system overview. Section 3 describes the implementation scenarios. Section 4 is reserved for the showing of experimental results. Finally, the last section concludes the paper.

– Normal BP: compress the data for the next transmission step.

– Incorrect BP: no transmission is performed, and repeat the measurement.

Machine Learning Application for BP Telemonitoring over WSN 81

Fig. 1. Proposed WSN architecture based BP telemonitoring

In this paper, the in BP status is taken as the object of the DT model. The systolic and diastolic BP denotes the combination of features vector Xi = [ SBP, DBP], and the prediction vector can be represented as Yi = [y1, y2, y3, y5], which signifies the False BP, High Hypertension, Hypertension, Normal BP, and Hypotension. The criterion of minimization of the square error appoints the selection of the features and the split node in the DT model construction.

82 A. El Attaoui et al.

3. for Xi in X:
For yi in y: 4.

5. Search Min MSE of Cross Validation 6. endfor
7. endfor
8. if (Current Depth < Best Depth )
Divide DT to D1 and D2 9.

The Blood pressure sensing node can be deployed in the patient’s hand to acquire BP measurement in the arteries using a sphygmomanometer. In the prototype

Raspberry Pi3

NRF24L01

The gateway collects data transmitted by the BPSN and another developed node. It should also be connected to the Internet to communicate data with the Cloud. The Raspberry Pi 3 (Fig. 3) is used as a Gateway node that contains an embedded Wi-Fi medium on this device and allows it to be continuously connected to the IoT platform using the MQTT protocol to improve real-time performance [16]. Besides, it can establish a secure channel with an IoT platform allowing health professionals to obtain information on human health. Besides, from the data frame sent by the BSN, the Gateway node can detect whether the received BP information contains an anomaly to alert or simply data to aggregate on the Cloud.

Decision tree classification

Close wireless

Communication

With the aim to illustrate the results, an actual experimentation of the telemon-itoring of blood pressure with WSN is presented in this section. First, the blood pressure classification by the decision tree algorithm was simulated before imple-mentation in the BPSN. Actual BP measurements of database [18,19] has been used as input data. A number of 5000 of records consider systolic and diastolic BP features are used for DT model construction. The five classes are labeled as Table 1 show [5]:

The algorithm was trained with 70% of input records and used the remaining 30% of records to the validation. Accuracy, Mean Absolute Error and Misclas-sified samples are calculated to evaluate the model and presented in Fig. 5. The obtained model of decision tree based CART algorithm is presented in Fig. 6.

Fig. 6. Constructed decision tree model

node. Figure 7 shows an example of a raw measurement data from of BP sensor (Kodea) recorded from 80 volunteers with 23 middle age. In addition, this figure shows also the decision about the BP records based on the systolic and diastolic features.

Fig. 7. Extract of BP measurements in

Figure 11 shows a graphical and numerical representation of the BP measure-ment received in the IBM IoT platform. By using this IoT platform the health professionals can check continuously and in real-time the patient’s BP status any time without the need for the manual measurement inside the health center.

Figure 12 shows the process of the data acquisition cared out in our research laboratory.

88 A. El Attaoui et al.

– Reduce the movement of patients to the health center for consultation or health control by wearable and remote monitoring.

– Improving patients’ data management and patient overload in the hospitals by providing an automatic and real-time diagnosis that can be updated con-tinuously in the IoT platform.

Esp. Cardiol. (Engl. Ed.) 72(12), 1065–1075 (2019)
2. Kilic, A.: Artificial intelligence and machine learning in cardiovascular healthcare.

Ann. Thoracic Surg., S0003497519316121 (2019)
3. Predicting Acute Hypotensive Episodes - The PhysioNet Computing in Cardiology Challenge 2009 v1.0.0”. [En ligne]. Disponible sur: . Accessed 16 Dec 2019
4. Singha, S.K., Sarkar, K., Ahmad, M.: An easy approach to develop a digital blood pressure meter. In: 2019 International Conference on Electrical, Computer and Communication Engineering (ECCE), Cox’sBazar, Bangladesh, pp. 1–5 (2019) 5. Satoto, B.D., Yasid, A., Syakur, M.A., Yusuf, M.: Wireless health monitoring with fuzzy decision tree for the community patients of chronic hypertension. J. Phys.

. Accessed 27 Feb 2019
16. Pi, R.: Raspberry Pi Compute Module (CM1) Raspberry Pi Compute Module 3 (CM3) Raspberry Pi Compute Module 3 Lite (CM3L) (2016)
17. Marolleau, B., Bacle, C., Lalevee, C.: Copyright IBM Corp. 2018 Materials may not be reproduced in whole or in part without the prior written permission of IBM, p. 41 (2018)
18. Data Visualisations > NCD-RisC. [En ligne]. Disponible sur: . Accessed 16 Dec 2019
19. Liang, Y., Liu, G., Chen, Z., Elgendi, M.: PPG-BP Database, 06 February 2018

TRNSYS Simulation of a Solar Cooling System Under Oujda Climate

1 Introduction

Most of the world’s energy consumption comes from fossil fuels (oil, natural gas, coal, etc.). Massive use can lead to the depletion of these reserves and a real threat to the environment owing to the increase in greenhouse gas (GHG) emissions. Therefore, the main concern now is to find rational conventional energy sources with very strict criteria and the use of renewable energy.

and absorption cooler were used in simulations with TRNSYS. The performance of the system was optimized by varying the size of the thermal storage tank, the area and the angle of inclination of the solar collector. The results showed that the optimal collector slope is around 30°, and that the optimal volume of the storage tank is of 0.6 m3 (600 L).

In this study, numerical simulations of a solar-assisted single-effect absorption cooling system using TRNSYS [5] were carried out for predicting the performance of the solar cooling system.

As illustrated by the configuration sketched in Fig. 1, the storage is a common element between the solar collector ETC [9] and the absorption chiller loop. It is the most adopted configuration in practice. The absorption chiller receives hot water from the thermal storage tank at 110 °C to evaporate the refrigerant, as long as the tem-perature of hot water at the outlet of storage tank remains below the specified value, the auxiliary boiler will be switched on to control the value temperature of fluid at the inlet of generator in the absorption chiller (i.e., Tg,i). Until it raise the required 110 °C, for this purpose a thermostat was carried out to monitor the temperature of hot water at the storage outlet. It is pertinent to mention that as the inlet hot water temperature can be varied in the limited range of 108–116 °C, therefore, temperature of 110 °C is used here as the minimum driving temperature (Tg,i) for simulating the operation of the absorption chiller [10]. The default rated capacity specified and COP of the chiller are 1494 kW and 0.53.

Figure 2. Illustrate the pictorial view of the solar cooling systems in TRNSYS.

Fig. 2. Pictorial view of the TRNSYS model.

TRNSYS Simulation of a Solar Cooling System Under Oujda Climate 95

The cooling load used in simulation, generated by Type 686 as synthetic demand, is based on defined peak load, time-of-day variations and weekday/weekend differences. The model represents a quick method of providing realistic loads without the time-intensive modeling required of a real building.

The heat energy of absorption and condensation rejected by the absorption chiller (_Q.cw) which is around 600 to 1000 Kw is defined as sum of the energy consumed by

The next important parameter is the chilled water outlet temperature. The hot water that receives the absorption chiller generator from the storage tank it comes out as chilled water. The chiller remains in operation as long as there is hot water and a demand for chilled water. It is noticeable from Fig. 6 that a constant chilled water temperature of 6.67 °C is obtained throughout the year.

In this study, Numerical simulation based performance analysis of a solar absorption cooling system is investigated using TRNSYS with the collector model input parameters taken from the measured performance data of an ETC.

The results obtained of the simulated solar based absorption indicate that an area of 400 m2of ETC, 10 m3of storage tank and an absorption chiller of 298 kW driven at a temperature of 108–116 °C are sufficient to maintain a temperature of 6.67 °C of air conditioning. The system achieves a maximum COP of 0.52 which is relatively high.

4. Florides, G.A., Kalogirou, S.A., Tassou, S.A., Wrobel, L.C.: Modelling and simulation of an absorption solar cooling system for Cyprus. Sol. Energy 72(1), 43–51 (2002)

Networks

Saida Driouache1(B), Najib Naja1, and Abdellah Jamali2

always best connected, and handover decision approaches are relevant

to attain efficient mobility solutions for mobile users. However, accom-

development of real-time scenarios. For the considered set up, the pro-

posed approach is faster than rank average by 97%, prevents ping pong

A Mobile Terminal (MT) executes a VH when it switches between two access networks of different technologies. VH happens to prevent a possible severe loss of performance or connection caused by technical reasons. Vertical Handover Decision Making (VHDM) is the most crucial step as it needs to deal with the best access network selection. Thus, this paper mainly concentrates on VHDM [15], where high computational effort conflicts with a low response time restric-tion. Also, the VHDM approach must be fast and able to give efficient solutions near to real-time [7]. Fuzzy Logic (FL) has the capability to reason precisely under imprecise information. So, this paper proposes a FL based VHDM app-roach over HetNets. The aim is to prevent ping pong handovers, reduce VHDM delay, and maintain the quality of real-time services at the highest level.

⃝ Springer Nature Switzerland AG 2021
T. Masrour et al. (Eds.): A2IA 2020, LNNS 144, pp. 99–109, 2021.

In [3], RSS thresholds are considered an input for the FL controller. So, VH is still susceptible to signal variation to some extent, making switching among different networks back and forth happen repeatedly. It is not the right choice to have only RSS measurement in HetNets, where every network has different characteristics. There is a need to consider multiple VHDM criteria.

When the VH does not happen at the right time, the connection is released, and sometimes, there is a ping-pong handover. That needs to be solved, since it may worsen the Quality of Service (QoS) or even disrupt the underlying com-munication. Therefore, authors in [4,11,13,16] used FL to reduce the ping pong effect. The proposed approach in [16] reduces the handover number and network load, and ensure QoS. The approach [4] well adapts to the high-speed environ-ments, guarantee flexible and reasonable vehicles access to a variety of networks. Authors [13] considered only LTE networks. The proposed approach suppresses the ping-pong effect. Even though, the proposed approach uses 243 rules which increases the VHDM delay.

Fuzzy Logic Based Vertical Handover 101

to determine the output. The inference process deduces a set of facts using if-then rules [6]. The defuzzification process converts the fuzzified output from a linguistic term to a crisp output value.

Making

This Section presents the proposed approach; VHDM-FL. This proposal seeks to exploit the advantages of FL; the aim is to have a fast and reliable VHDM for real-time services. Figure 2 illustrates VHDM-FL, which decides whether to continue with the currently connected network or to switch to another one. VHDM-FL receives the VHDM criteria for available access networks. Many input variables may cause VHDM-FL less sensitive, and handover might not occur when necessary. Therefore, there are only two input variables and one output variable Q.

102 S. Driouache et al.

Next, VHDM-FL applies the Mean of Maximum (MoM) defuzzification method to determine the output QCN for every candidate network. In other words, the fuzzy controller takes the typical value of the consequent term of the most valid rule as the crisp output value. Finally, VHDM-FL recommends the CN with the maximum value Q.

5 Performance Evaluation

Table 1. Simulation parameters

‘Medium’, and ‘High’. Nine labels describe the output variable; ‘Very Low’ (VL),‘Low’ (L), ‘Somewhat Low’ (SL), ‘Medium Low’ (ML), ‘Medium’ (M), ‘Medium High’ (MH), ‘Somewhat High’ (SH), ‘High’ (H), and ‘Very High’ (VH). To make a successful VHD, Table 3 shows that VHDM-FL uses nine fuzzy control rules.

Table 2. MFs parameters for antecedents (3 labels) and consequents (9 labels)

This subsection discusses the results obtained from the simulation. Throughput, end to end delay, jitter, VHDM delay, and a ratio of VHDM delay to overall VH delay are metrics that help assess the reliability and performance of VHDM-FL. RA approach presented in [5] outperforms some powerful MCDM approaches (TOPSIS and VIKOR). Hence, it would be interesting to compare the VHDM-FL (FL) to RA (MCDM).

Fig. 5. Throughput

106 S. Driouache et al.

– Jitter is a critical metric to measure because high jitter values can lead to poor voice quality. Jitter must be ≤ 20 ms [10] to have good voice quality. Accord-ing to Fig. 7, RA and VHDM-FL provide good jitter results. For VHDM-FL, jitter does not exceed 1 µs.

those presented here. They are affected by factors that differ from one simu-lation to another; the way an operating system runs threads and the number of simultaneously run processes. The minimum values of multiple runs are considered to obtain reliable results. Compared to RA, VHDM-FL offers a low VHDM delay. VHDM-FL guarantees the low VHDM delay requirement for real-time applications. It is relevant to consider this result; mobile devices should give a high-performance level at a lower price.

robust; however, VHDM-FL allows better overall performance. The next step in this research work will be to add some learning capabilities to VHDM-FL via reinforcement learning and evaluate the impact of learning on VHDM-FL for real-time applications. The goal is to see how results may change. Moreover, the approach developed should be expandable to include additional QoS parameters of interest to VHDs.

References

15. Yew, H.T., Supriyanto, E., Satria, M.H., Hau, Y.W.: A vertical handover manage-ment for mobile telemedicine system using heterogeneous wireless networks. Int.

J. Adv. Comp. Sci. Appl. (IJACSA) 7, 1–9 (2016)
16. Zhang, L., Ge, L., Su, X., Zeng, J.: Fuzzy logic based vertical handover algorithm for trunking system. In: 2017 26th Wireless and Optical Communication Confer- ence (WOCC), pp. 1–5. IEEE (2017)

Keywords: Source code understanding � Natural language understanding �Computer science teaching and learning � Machine learning

1 Introduction

The global objective of our research work is to understand programmer attitude in learning context based on his actions on the code. We aim to develop a kind of source code understanding based on machine learning in general and particularly natural language understanding mechanism.

To obtain rich information about programmer interactions with code environment we favored using a source code watcher based on Git objects. Even Git commits are not enough to understand all programmer actions. The information we can gather from this are very rich and we believe that it can be exploited in different levels. To start with this promising research we will try first to track programming scenario to match it with a preconfigured scenario. The objective is to evaluate ideas orders of the learner to identify how he think while trying to solve problems.

NLP can be divided into two kind of processing:

• Natural language understanding (NLU): It is the NLP undertaking of extracting insights from natural language inputs. The objective is to give computer or machine the ability to understand Natural language.

• Machine translation
• Information retrieval
• Information extraction
• Natural language Assistant
• Spell check

2.1 Natural Language Understanding and Code Understanding

• Machine learning model: Uses a statistical approach to find entities and relation- ships in documents. This type of model can adapt as the amount of data grows.• Rule-based model: Uses a declarative approach to finding entities in documents.

This type of model is more predictable, and is easier to comprehend and maintain.

(a) Dynamic analysis, which is based on software execution traces, and examines programs runtime.

(b) Static analysis, based on static source code information, such as slicing, control or data flow graphs.

As we plan to use this intelligent system in different manners. We are trying to design a solution with a general purpose. The objective is to collect mass data related to source code and to build machine learning models that we can train for multiple source code learning concepts. The following figure represent general architecture of proposed intelligent system (Fig. 1).

114 Z. Itahriouan et al.

The process of this component aims to collect and refine data related to coding actions. In development environments source code is the space were programmer use his skills to build programs according to already expressed requirements. In teaching and learning acts, student uses IDEs to solve exercise problems (problem based learning) or to build projects (Project based learning).

Toward Intelligent Solution to Identify Learner Attitude 115

While coding history can contain more information as explained previously, data collected in this intelligent solution will be extracted from source code and also its history.

3.2 Analysis Technics

116 Z. Itahriouan et al.

sentiment or emotions from text with a score [16]. IBM offers a ready to use pre-trained model to extract sentiment, emotion, keywords, entities, categories, concepts and semantic roles from different text formats [17, 18]. We Intent mainly by this study to use NLU based on deep learning models to build similar applications that can extract meanings from student code. The objective is to understand learner coding style and attitude from his source code.

• Annotation: to use to identify entity types and relationship types to form ground truth. In source code context, entity types can be variables, functions, structures, and comments … relation can be type related to returning function. Rule based Model can recognize patterns based on syntax rules.

• Training: use ground truth to train Machine Learning Model. Based on imported training files.

We have introduced first a general architecture to analyze learner code and actions according to analysis technics used for code understanding. This architecture is based on IBM Watson in general and IBM Watson knowledge studio in particular. Therefore, we have explained how to use machine learning to build domain specific Model for source code teaching and learning domain.

As building machine learning Model adapted to domain specific need iterative process to become performant, selecting training data is a very important step to build a performant Model. Before that we need to specify exactly learner code style and

1. Itahriouan, Z., Aknin, N., Abtoy, A., El Kadiri, K.E.: Building a web-based IDE from web 2.0 perspective. Int. J. Comput. Appl. 96(22), 46–50 (2014)
2. Itahriouan, Z., Aknin, N., Abtoy, A., El Kadiri, K.E.: An experimental study of software engineering learning using IDE 2.0. In: 2016 4th IEEE International Colloquium on Information Science and Technology (CiSt), pp. 559–563 (2016)
3. El Bahri, N., Itahriouan, Z., Abtoy, A., Brahim Behaouari, S.: Proceeding of the 11th International Conference on Education and New Learning Technologies (EDULEARN 2019), Palma, Spain, 1–3 July 2019.
4. Hurd, S.: Towards a better understanding of the dynamic role of the distance language learner: learner perceptions of personality, motivation, roles, and approaches. Distance Educ.

27, 303–329 (2006).
5. Dörnyei, Z.: New themes and approaches in second language motivation research. Annu.

118 Z. Itahriouan et al.

15. Arabyarmohamady, S., Moradi, H., Asadpour, M.: A coding style-based plagiarism detection. In: Proceedings of 2012 International Conference on Interactive Mobile and Computer Aided Learning (IMCL), pp 180–186 (2012)
16. Chatterjee, A., Gupta, U., Chinnakotla, M.K., et al.: Understanding emotions in text using deep learning and big data. Comput. Hum. Behav. 93, 309–317 (2019).

Laboratoire Systèmes et Environnement Durables (SED),
Faculté des sciences de l’ingénieur (FSI), Université Privée de Fès (UPF), Fez, Morocco
salouasenhaji@gmail.com, sanaa.faquir@usmba.ac.ma,

fidae.harchli@gmail.com, hajji-tarik@hotmail.com,

1 Introduction

The Internet of Things (IoT) is a system of interconnected computing devices, mechanical and digital machines, objects, animals or people that are provided with unique identifiers (UIDs) and the aptitude to transfer data during a network without requiring human-to-human or human-to-computer interaction [1–3].

These features comprise heterogeneity with regard to of platforms complicated per network, broad range of central processing architectures and multiple modes of system imaging [6].

Analysis of malware as images has shown remarkable results in the recent years.

Several methods of malware analysis have been proposed in the literature [18].

The existing malware image analysis and classification methods can be divided into two: the deep learning-based and handcrafted feature-based approaches. To begin, we review the deep learning-based approach. Liu et al. [19] present a method to classify malware using gray-scale images and ensemble learning.

In handcrafted approach, the image features are extracted and the classical machine learning features applied to learn them and classify the malware into its respective class. Different researchers have extracted variety of image features and applied vari-ants of classical machine learning algorithms to classify malware.

Recently, there have been several attempts to use image analysis on Internet of Things malware [21, 22, and 23].

Let u be the true image and f the distorted image. Then the model of ROF invented by Rudin, Osher and Fatemi to solve is:

122 S. Senhaji et al.

This model decompose an image f into a component u and a component v = f − u; which is supposed to be the noise. In [9] Meyer out some limitations of the ROF model.

Thus, decomposition is given by minimizing a functional f:

Fig. 1. Diagram of supervised learning using decomposition model

Naïve Bayes algorithm: Naïve Bayes classifier is a probabilistic classifier that applies Bayes theorem with strong independence assumptions. The linear classifier is easy to construct and is robust to noise [24].

124 S. Senhaji et al.

Feature extraction is based on a method introduced in Haralick et al. [27]. On each texture component image Gray Level Co-occurence Matrix (GLCM) is computed and then the texture features based on the GLCM are calculated. The computation of GLCM is three fold: grayscale compression, generation of co-occurence matrix and calculation of feature. The work of Haralick offered 14 statistical measures of texture features that can be calculated on GLCM parameters.

The results of precision and recall are presented in the Table 1. we can deduct that all the methods performed above average with results closer to perfect case of one as follows (Random forest = 0.99; K-nearest neighbour = 0.9 and Naïve Bayes = 0.95). Random forest outperformed K-nearest neighbour with overall accuracy of 95.38% compared to K-NN that had 85% on grayscale images. The results shows that usage of images in malware analysis is a promising approach that culls the architecture and platform bottlenecks offering promising results.

In this paper we have presented a new method to analysis of internet of things malware based on decomposition of the images. It consists to split image into two components where the first contains the texture and the second is referred to geometrical charac-teristic. The global texture features are extracted from the texture component image using the gray level co-occurence matrix. The extracted texture features are then used with classical machine learning classifiers that can be applied to IoT malware.

Our approach is evaluated and the result indicates that the proposed decomposition lead a strong potential towards the improvement of the performance.

Faraj Zainab1(&), Aboussaleh Mohamed1, and Zaki Smail2

1 Mechanical and Integrated Engineering (M2I), Laboratory of Sciences and Professions of the Engineer, Moulay Ismail University, ENSAM, Marjane 2, BP: 29850050, Meknes, Morocco
faraj.zainab@gmail.com, m.aboussaleh@umi.ac.ma
2 Advanced Mechanics and Industrial Applications (M2AI),
Laboratory of Mechanics, Mechatronics and Control (L2MC),
Moulay Ismail University, ENSAM, Marjane 2, BP: 29850050,
Meknes, Morocco
s.zaki@umi.ac.ma

Although this technology has many advantages, it also has disadvantages, such as thermal resistance and the effect of stairs that often appear when it comes to inclined surfaces. To reduce the problem of thermal resistance, Chen and Al [4] have been working on the optimization of laser volume energy to improve the possibility of

© Springer Nature Switzerland AG 2021
T. Masrour et al. (Eds.): A2IA 2020, LNNS 144, pp. 126–134, 2021.

The influence of all these factors on the SLM process makes it a complex multi-factorial process which to be optimized will need a powerful tool. The Taguchi method is a statistical tool for carrying out experimental designs, applied in industry, this method aims to minimize disturbances to the process and to make the latter robust and more resistant to variations caused by noise factors, taking into account several factors while basing itself on the principle of orthogonality which evaluates the effects of the factors on the mean and on the variation of the response and this with the objective of reducing the duration and the cost of the experiment [11–14]. The goal of this work was first to make a feasibility study of the parts produced by SLM and this by printing a test plate in AlSi7Mg06 in order to know the limitations of the process, then to study the effect of treatment parameters on the mechanical properties of the parts produced by SLM, and conclude which one has the most impact. The Taguchi method was used for the analysis and optimization of the SLM parameters.

2 Experiment

Fig. 1. The schematic diagram of the SLM equipment

2.2 Experimental Material

Phase 1: Feasibility Study of the SLM Process
The Phase consists of conducting a feasibility study of the parts produced by SLM in order to evaluate the performance of the process and minimize undesirable malfunc-tions, by printing an aluminum test tray.

Cubes located along the X and Y axes, the objective of this test is to determine the effect of the location of the part on its printability and mechanical properties.

Cylinders placed in the center of the table to see the capacity of the machine to print centered parts.

Fig. 2. Orthogonal experiment scanning strategy (a) S-type x, (b) S-type y, (c) contours offset, and (d) X–Y inter-layer stagger scanning.

The printing result of the test plate in AlSi7Mg0.6 is shown in Fig. 3, the work lasted 30 h 2 min 15 s.

3.2 Phase 2

For this study, the combinations proposed by the Taguchi table have been grouped in Table 3, according to the experimental repeatability of Taguchi’s approach which is three times, 16 � 3 = 48 cubes of 10 � 10 � 10 mm3will be produced to evaluate the influence of SLM parameters on the density of the produced parts.

The Effect of Layer Thickness on the Density of the Parts Manufactured. The first image in Fig. 4 shows the values taken by the density of the parts produced at different values of the layer thickness, which varies from 0.02 mm to 0.05 mm. It is remarkable that the density of the samples is on a decreasing curve at different speeds and which becomes more and more important when the layer thickness passes from the

132 F. Zainab et al.

Mean density (%)	95							95
	93							93
	91							91
	89							89
	87							87
Mean density (%)	85						Mean density (%)	85
		LT1	LT2	LT3		LT4			LED1	LED2	LED3
		a-Layer thickness (mm)							b-Linear energy density (J/mm)
	95							95
	93							93
	91							91
	89							89
	87							87
	85							85
		HS1	HS2		HS3	HS4			SS1	SS2	SS3	SS4
		c-Hatching space (mm)

The present work evaluated the performance of the SLM process by printing a test plate in AlSi7Mg06 containing several structures in order to know the limitations of the machine. Then, an experimental design was developed based on the Taguchi method in order to optimize the SLM parameters: hatching space, scanning strategy, powder thickness and linear energy density, in order to obtain parts with better maximum density. The Taguchi table chosen offers sixteen different combinations, the maximum density obtained is 95, 2234%, the optimal process parameters are: powder thickness of 0.02 mm, scanning strategy of X-Y inter-layer stagger scanning, linear energy density of 0, 5 J/mm and hatching space of 0.08 mm.

References

Mech. Eng. 323, 27–63 (2017)
11. Canel, T., Kaya, A.U., Celik, B.: Parameter optimization of nanosecond laser for micro drilling on PVC by Taguchi method. Opt. Laser Technol. 44, 2347–2353 (2012)

134 F. Zainab et al.

elhabib.nfaoui}@usmba.ac.ma

Abstract. Nowadays, people express their opinions and thoughts about an item (products, movies, services, and brands, etc.) directly online. They use platforms like IMDB, Amazon, eBay, and more. This large amount of people’s opinions is very valuable for an automatic computing of reputation score for companies. Consequently, developing a system able to generate reputation from textual opinions and their attached rating will be a helpful tool for companies to have an idea about the quality and the issues of their product, and for potential customers to assist them during their ecommerce decision-making. However, a very limited number of studies have investigated mining reviews expressed in natural lan-guage for reputation generation. Therefore, we propose an approach and a whole system for generating reputation using bidirectional long short-term memory Recurrent Neural Network (Bi-LSTM RNN) and Natural Language Processing (NLP) techniques. The experimental results conducted on the IMDB dataset show that our method provides the nearest reputation value to the ground truth (IMDb weighted average vote). This implies that the proposed approach can be applied in practice to generate reputation.

136 A. Boumhidi et al.

attached to the textual comment to its actual corresponding text meaning, thus we fuse the generated text with the comment to make the opinion denser and more prosperous in term of sentiments toward the entity. Next, we build a Recurrent Neuron Network (Bi-LSTM) in order to perform the classification and predict the sentiment polarity of each opinion from the previous step. Finally, we used both the result from the clas-sification phase and the rating attached to the comment to generate a reputation value applying arithmetic means. The reminder of this paper is organized as follows. Sec-tion 2 provides background and related work about reputation generation and senti-ment analysis. In Sect. 3, we explain our method. We show experimental results and discussions in Sect. 4. Finally, we present a conclusion and summary of the main points of this paper, including perspectives.

2.2 Sentiment Analysis

Sentiment analysis is the role of identifying opinions, emotions, and evaluations in a given text. The important concept in sentiment analysis is to identify the polarity of the overall text. It helps to indicate whether the text is positive, negative, or neutral. In

Fig. 1. The workflow of the proposed reputation generation system.

Pre-processing is necessary to eliminate text noises. It consists of several steps, and it should be performed before the text classification process, especially with textual data collected that has many non-standard text spelling words. The pre-processing steps applied in this paper are as follows:

• Case Folding: All words are converted into lower case• Special character, URL, Punctuation mark are removed• Replacing slang word
• All Stopwords are removed

Mining Online Opinions and Reviews Using Bi-LSTM for Reputation Generation 139

After fusing the comments with words generated from the rating using natural language processing techniques, the opinions are ready for the classification phase.

Based on the result from the classification phase and the ratings attached to every comment we are now able to calculate the reputation expressed by a single value that summarizes people’s opinions about that specific target entity. The result from the classification phase Stotalis a value between 0 and 10 that represent the positive opinions, where the symbol Rtotalrepresents the average of all ratings. We propose the following formula (Formula 1) which is the average of the result from the classification phase Stotaland the average of all ratings Rtotalto finally compute a reputation value:

R ¼Stotal þ Rtotal ð1Þ

4.1 Datasets Description

IMDB Reviews Dataset. To train our model we used the IMDB Movie Review Dataset [12]. This is a dataset for binary sentiment classification containing a set of 25,000 highly polar movie reviews for training, and 25,000 for testing.

An experimental study conducted indicates that the comments carrying words generated from the rating impact positively the accuracy of our deep learning model. Table 3 shows the important improvement of accuracy because of the ratings con-version and integration.

Table 3. Comparison between the accuracy of sentiment polarity of comments with and without words generated from ratings

Fig. 3. Classification result for the dataset 1 “The Irishman (2019)”

The next step is to calculate the reputation score based on the result from the classification phase, where we only take percentage of the positive comments divided by 10 to have a value Stotalbetween 0 and 10.

Comments

Nega�ve

References

1. Jøsang, A., Ismail, R., Boyd, C.: A survey of trust and reputation systems for online service provision. Decis. Support Syst. 43(2), 618–644 (2007).

144 A. Boumhidi et al.

11. Thongtan, T., Phienthrakul, T.: Sentiment classification using document embeddings trained with cosine similarity. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, Florence, Italy, pp. 407–414 (2019).
12.

benghabrit@enim.ac.ma
2 LMAID Laboratory, ENSMR, Mohammed V University, Rabat, Morocco

Abstract. The proliferation of linked medical data on the Web has increased rapidly in terms of both the number of repositories and the datasets’ sizes. These data have a formal format that is RDF (Resource Description Framework). However, the size and complexity of these medical linked data, leading to an error in modeling, missing concepts as well as missing relationships and inconsistencies in critical clinical applications and biomedical research. Thus, to overcome these issues, this paper aims to describe an approach that uses neural language models for RDF data clustering to identify a thematic view of the clustered entities to extract labels or tags even if there are missing entities or links in the integration of medical RDF datasets. Experiments on a bench-marking dataset give us promising results.

© Springer Nature Switzerland AG 2021
T. Masrour et al. (Eds.): A2IA 2020, LNNS 144, pp. 145–157, 2021.

146 S. Eddamiri et al.

The rest of this paper is structured as follows: In Sect. 2, we present related works structured regarding Linked Data technologies in medical data and explores how semantic information is used in different data mining applications for the healthcare domain. Section 3 describes our approaches of extracting feature vectors from the RDF graph in order to prepare the clustering mechanism integrating with a technique for generating themes for the obtained clusters in a medical RDF dataset. The obtained experimental results and their analysis are depicted in Sect. 4. The last section contains conclusions and paves the way for some potential-work suggestions.

2 Related Work

148 S. Eddamiri et al.

agglomerative clustering. Escudero et al. [38] classify Alzheimer’s disease in patho-logic groups or non-pathologic or detect breast cancer recurrence using K-means clustering. While Missikoff et al. [39] combine a linguistic and statistics-based method to perform automated OntoLearn ontology generation tasks from texts.

Fig. 1. Our methodology for clustering medical RDF data applied to theme identification

The objective of this section is to propose an integrated approach that automatically identifies adequate themes or labels for the generated medical data clusters. The computation of these recommendations is described in the following steps:

• Definition 1: A Swb is defined as the union of one incoming edge and walks with a specific vertex vr by depth d using this formula:

Swb vr ð Þ ¼ [ fWalki ¼ ðvr þ i�1; er þ i; vr þ iÞ 2 Tj0 � i � dg

150 S. Eddamiri et al.

3.3 RDF Data Clustering Process

Where e represents the set of all types defined as:

4 Results and Discussion

152 S. Eddamiri et al.

The F-measure is based on basic knowledge recovery criteria, “precision” and the concept of “recall”:

For the entire clustering result, the F-measure is computed as following:

The overall value for purity is the weighted average of all values for purity. It is given by the following formula:

We evaluate the relevance of both the RDF graph clustering and theme identification algorithms on mesh dataset using a set of backward Walks defined in Sect. 3.1 and with the Doc2vec techniques and TF-IDF as a pre-processing block as the following steps:

• The RDF graphs are first transformed into sequences by constructing a Walks-with backward with the depth d = 2, 3.

The goal of these experiments is to show that even with missing links and entities we can identify and grouped vertices in the medical RDF graph.

4.4 Experimental Results

Table 4. Result of the Sequential Theme Identification Algorithm on the three datasets in terms of Purity and F-measure

Thus, the sequential approach also supported by TF-IDF to remove low frequency entities. Nevertheless, there is a little variation between the size of the cluster interest and the implications of the identification of themes. Many vertices do not include the type (rdf: type) or have more than one type leading to overlapping between the clusters. This allows analyzing and mapping existing relations between the two different semantic types as shown along with the disease symptom application scenario.

5 Conclusion

References

1. Studnicki, J., Berndt, D.J., Fisher, J.W.: Using information systems for public health administration chapter overview
2. Burgun, A., Bodenreider, O.: Accessing and integrating data and knowledge for biomedical research. Yearb. Med. Inform. 17, 91–101 (2008).
3. Weiner, M.G., Embi, P.J.: Toward reuse of clinical data for research and quality improvement: the end of the beginning? Ann. Intern. Med. 151, 359–360 (2009)
4. Maro, J.C., Platt, R., Holmes, J.H., Strom, B.L., Hennessy, S., Lazarus, R., Brown, J.S.: Design of a national distributed health data network. Ann. Intern. Med. 151, 341–344 (2009) 5. Prokosch, H.U., Ganslandt, T.: Perspectives for medical informatics - reusing the electronic medical record for clinical research. Methods Inf. Med. 48, 38–44 (2009).

156 S. Eddamiri et al.

14. Belleau, F., Nolin, M.A., Tourigny, N., Rigault, P., Morissette, J.: Bio2RDF: towards a mashup to build bioinformatics knowledge systems. J. Biomed. Inform. 41, 706–716 (2008) 15. Tudorache, T., Nyulas, C.I., Noy, N.F., Musen, M.A.: Using semantic web in ICD-11: three years down the road. In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), pp. 195–211 (2013) 16. Malone, J., Holloway, E., Adamusiak, T., Kapushesky, M., Zheng, J., Kolesnikov, N., Zhukova, A., Brazma, A., Parkinson, H.: Modeling sample variables with an experimental factor ontology. Bioinformatics 26, 1112–1118 (2010).
17. UniProt Consortium: Update on activities at the Universal Protein Resource (UniProt) in 2013. Nucl. Acids Res. (2012).
18. Croft, D., O’Kelly, G., Wu, G., Haw, R., Gillespie, M., Matthews, L., Caudy, M., Garapati, P., Gopinath, G., Jassal, B., Jupe, S.: Reactome: a database of reactions, pathways and biological processes. Nucl. Acids Res. 39, D691–D697 (2010). 19. Pence, H.E., Williams, A.: ChemSpider: an online chemical information resource (2010) 20. Kelder, T., Van Iersel, M.P., Hanspers, K., Kutmon, M., Conklin, B.R., Evelo, C.T., Pico, A.

26. Celermajer, D.S., Chow, C.K., Marijon, E., Anstey, N.M., Woo, K.S.: Cardiovascular disease in the developing world: prevalences, patterns, and the potential of early disease detection. J. Am. Coll. Cardiol. 60, 1207–1216 (2012).
27. Brixner, D., Rubin, D.T., Mease, P., Mittal, M., Liu, H., Davis, M., Ganguli, A., Mark Fendrick, A.: Patient support program increased medication adherence with lower total health care costs despite increased drug spending. J. Manag. Care Spec. Pharm. 25, 770–779 (2019).
28. Phua, C., Lee, V., Smith, K., Gayler, R.: A comprehensive survey of data mining-based fraud detection research (2010).
29. Hastie, T., Tibshirani, R., Friedman, J.: Overview of supervised learning. In: The Elements of Statistical Learning, pp. 9–41. Springer (2009)
30. Bundschus, M., Dejori, M., Stetter, M., Tresp, V., Kriegel, H.P.: Extraction of semantic biomedical relations from text using conditional random fields. BMC Bioinform. 9 (2008).

31. Shouman, M., Turner, T., Stocker, R.: Applying k-nearest neighbour in diagnosing heart disease patients. Int. J. Inf. Educ. Technol. 2, 220–223 (2012).
32. Chang, C.L., Chen, C.H.: Applying decision tree and neural network to increase quality of dermatologic diagnosis. Expert Syst. Appl. 36, 4035–4041 (2009)

39, 45–65 (2003)
45. Zhang, W., Yoshida, T., Tang, X.: A comparative study of TF* IDF, LSI and multi-words for text classification. Expert Syst. Appl. 38, 2758–2765 (2011)
46. Steinbach, M., Karypis, G., Kumar, V.: A comparison of document clustering techniques. In: KDD Work. Text Mining, pp. 1–20 (2000). 47. Zhao, Y., Karypis, G.: Empirical and theoretical comparisons of selected criterion functions for document clustering. Mach. Learn. 55, 311–331 (2004)

A Proposal for a Deep Learning Model to Enhance Student Guidance
and Reduce Dropout

1 Introduction

Every day, 2.5 trillions bytes of data are created in the world, to the point that 90% of the data in the world were created in the last two years. According to a study conducted in 2012 by the IDC for the EMC (IDC study for the EMC 2012), the data will be multiplied by 50 between 2005 and 2020 to reach 40 zettabytes1.

store large quantities of longitudinal data on students’ right down to very specific transactions and activities on learning and teaching. When students interact with learning technologies, they leave behind data trails that can reveal their sentiments, social connections, intentions and goals. Researchers can use such data to examine patterns of student performance over time from one semester to another or from one year to another (Daniel 2014), we can therefore make a sense to university’s data through collecting, storing and analyzing it with machine learning in order to improve student learning quality.

The flow of new students who have obtained their high school degrees and intend to join university has continued to grow. But, university dropout rates increases too, so how are these students directed towards a given specialty the university. Thus, and according to the National Evaluation Body INE2of the CSERFS3, guidance and tutoring are considered as the weak links at the origin of the losses, especially those occurring in the first year of the cycle of license.

2 INE: Instance Nationale d’Evaluation Marocain.

3 Conseil Supérieur de l’Education de la Formation et de la Recherche Scientifique Marocain.

Our idea is then to propose a deep learning model based on neural networks that allow to predict the results of the new students, based on a dataset containing the data of the old students and their results, i.e. the best pathway for each category of students, in order to optimize students’ choices of courses.

We first present the data used as a dataset and then the proposed model that predicts the results of the new students, hereafter the general outline of our system “cf. Fig. 1”.

In this first part, in perspective to can predict a result of every student, we followed the process of manipulating, processing, cleaning, and crunching data. Then we have collected the data present in the Oracle database of the university to use them as dataset“cf. Fig. 2”, these data concern several students belonging to several sectors, namely the gender, the place of birth, the series of the baccalaureate, the province of the baccalaureate, the academy, the diploma in which the student continues his studies, the result of the student. Generally, the dataset must have thousands of rows or more. These data will allow us to have an idea about the adequacy of the training to the types of students following the predictions that reflect if that a category of students could succeed in a pathway or not.

A good way to quickly check correlations among columns, correlation coefficient for each feature to every other feature, is by visualizing the correlation matrix as a heatmap, then, first we calculated a correlation matrix and after we used a heatmap and we obtained a result presented in “cf. Fig. 3”.

162 M. Sabri et al.

In this part, we will present a model of neural network that analyzes the data already collected to have a prediction on the success rate in a pathway for a category of students according to several criteria by using deep learning, more specifically the Tensorflow library which is an open source machine learning tool from Google.

The goal of our proposed model is to solve a problem of the supervised classifi-cation type, since the result “output” of our dataset will be a class, i.e. whether the student has succeeded or not, we have in the input layer “input” of our network 6 nodes“neurons” corresponding to 6 variables of the dataset while the last column of our model will be the output. In addition, the dataset will be divided into two parts, 80% training data and 20% testing data.

After, we obtain the activation, according to the function “ReLU” in the two hidden layers and with the “Sigmoid” in the output layer. The result obtained is compared with the real value of the dataset to obtain the accuracy rate and the loss rate, the weights are then corrected in the following iterations in order to increase the accuracy rate and decrease the loss rate using the gradient descent algorithm. We ran the algorithm 500 times/iterations.

A Proposal for a Deep Learning Model to Enhance Student Guidance 163

Fig. 5. Accuracy and loss values of our model

The success rate is 0.9231426131511529 and the loss rate is 0.2739, so we can say that the model trained well on the dataset to be able to give predictions concerning the orientation of the future students with a accuracy rate of 92%, Then we applied these results to a few students in 2019 to find out if they will succeed or not in their chosen pathways, and we have obtained the following results (Fig. 6):

5 Conclusion

Deep learning is a branch of machine learning which can process large data in order to solve problems of different types, in our case we have a supervised classification problem so we used deep learning to set up a new student guidance system based on data from previous years of the university, this new system predicts a best pathway in university for a student according to his category following an apprenticeship made on the data dataset. This new student guidance system can help to decrease the dropout that is increasing more and more. The problem we have dealt with is a nonlinear problem, we used a neural network with two hidden layers, the first with 12 nodes and the second with 15 nodes “cf. Fig. 3”. These two choices directly determine the number of weights to be estimated and therefore the complexity of the model. They participate in the search for a good compromise bias/variance, i.e. the balance between quality of learning and quality of prediction. Other parameters also affect this compromise as the maximum number of iterations (100 in our case).

IDC study for the EMC: The Digital Universe in 2020: big data, bigger digital shadows and biggest growth in the Far East. IDC (2012)
Seufert, S., Meier, C.: Big data in education: supporting learners in their role as reflective practitioners (Book Chapter). University of St. Gallen, St. Gallen, Switzerland (2018) Daniel, B.: Big Data and analytics in higher education: opportunities and challenges. Br. J. Educ.

Technol. (2014).
Instance Nationale d’Evaluation: Rapport sectoriel - l’Enseignement Supérieur au Maroc Efficacité efficience et défis du système universitaire à accès ouvert (2018)
Nathalie Beaupère: Sortir sans diplôme de l’université. Chargée d’Études CAR Céreq Bretagne, Faculté d’économie, Université de Rennes 1 (2009)
Binder, T.A., Sandmann, A.A., Sures, B.B., Friege, G.C., Theyssen, H.D., Schmiemann, P.A.: A note on the draft International Council for Harmonisation guidance on estimands and sensitivity analysis. Int. J. STEM Educ. (2019)
Menard, B.: Higher education dropout in the light of the capability approach. University of Toulouse Jean Jaurès, France, August 2018
Elbir, A., Gündüz, E., Diri, B.: Estimating the School Dropout Trend by Using Data Mining Methods. Bilgisayar Mühendisliǧi Bölümü, Yildiz Teknik Üniversitesi, Istanbul, Turkey, November 2018
Oladokun, V.O., Adebanjo, A.T., Charles-Owaba, O.E.: Predicting students’ academic performance using artificial neural network: a case study of an engineering course.

tarek.cse,sadekur.cse}@diu.edu.bd

Abstract. In the present time Chatbot is an essential tool used by many organizations to provide services to their targeted customers round the clock. This research focuses on a domain-specific Chatbot that can be helpful for educational institutes. This Chatbot will be a virtual (representation) to the admission seekers. It will provide answers regarding the university, its depart-ments, admission fees and other admission related FAQ. For the sake of the research, frequently asked questions of a university were collected and an unsupervised learning model along with natural language processing techniques was deployed to answer the questions of the admission candidates. Tokeniza-tion, stop words removal followed by vectorization were implemented for preprocessing the training data. User’s inputs were similarly processed and then tf-idf based cosine similarity applied to retrieve the best answer. Later, a user-centric evaluation metric was used to evaluate the model and as per the metric, our current model showed approximately 80% accuracy.

© Springer Nature Switzerland AG 2021
T. Masrour et al. (Eds.): A2IA 2020, LNNS 144, pp. 166–174, 2021.

EduBot: An Unsupervised Domain-Specific Chatbot for Educational Institutions 167

2 Related Work

Chatbot has a long history in the research world. Alan Turing in 1950 published a paper“Computing Machinery and Intelligence” which is considered as the benchmark work for computing intelligence of machines [2]. Based on that work Joseph Weizenbaum developed ELIZA [3] which used to check users’ input and extracted some keywords and based on those keywords processed a response. In 1995 ALICE (Artificial Lin-guistic Internet Computer Entity) become popular which used heuristic pattern matching rules to responds to users’ query(ies) [4]. In recent times IBM’s Watson [5, 6] Apple’s Siri [7] and Amazon’s Alexa [8] become very popular.

Ahmed Fadhil presented the challenges based on a question of whether a Chatbot can determine any one’s diet for a Chatbot application. This system was designed with some proper frequent steps to meet the goal. A study on BotAnalytics shows that only 60% of users are continuing their conversation for the second message and another 75% are continuing their conversation for further. Based on this study, an investigation has been done for meal recommendations and also for lifestyle promotion on the performance of Chatbots activity [12].

Kar and Haldar, presented their works on the application of Chatbot to the internet of things that specially described about the opportunities and architectural elements. Study shows that there are some chat interfaces those are being used in an Instant Messaging (IM) platforms like Slack, Facebook Messenger, Kik, Telegram etc. those are very popular and growing rapidly [13]. Moreover, this study shows how the top ten messaging platforms alone account for about 4 billion users. So, the worldwide demand for Chatbot accelerating the invention with the more user-friendly applications and specific purpose Chatbots.

Fig. 1. Research methodology

3.2 Data Preprocessing

As per the research method, all our data were processed though through a few steps which are described below.

Fig. 3. Sample list of stopwords

Similarity Calculation and Answer Selection
The question collected from the user is taken as input and using similarity algorithm (such as cosine-similarity, TF-IDF matching) matching the most similar question of the dataset is selected. Later corresponding answer to that question is selected to present as the answer to the user’s query.

Cosine similarity is calculated through the following equation:

closer are the texts.

One advantage of cosine similarity is that, if the distance of the two documents is far by Euclidean distance, it can still find the similarity between them.

Fig. 5. Frequency distribution of users’ feedback

Therefore,

EduBot: An Unsupervised Domain-Specific Chatbot for Educational Institutions 173

Fig. 6. Sample responses of the users

5 Conclusion

1. Shawar, B.A., Atwell, E.: Different measurements (measurement)metrics to evaluate a Chatbot system. In: Proceedings of the Workshop on Bridging the Gap: Academic and Industrial Research in Dialog Technologies, pp. 89–96. Association for Computational Linguistics, April 2007
2. Turing Test, Turing, A.: Computing machinery and intelligence. Mind 59, 433–460 (1950).

3. ELIZA, Weizenbaum, J.: Computer Power and Human Reason: From Judgment to Calculation, p. 2, 3, 6, 182, 189. W.H. Freeman and Company, New York (1976). ISBN 0- 7167-0464-1
4. AIML: AIML 2.0 draft specification released. . Accessed 16 Jan 2013 5. Jackson, J.: IBM Watson Vanquishes Human Jeopardy Foes. PC World. IDG News, 17 February 2011
6. Takahashi, D.: IBM researcher explains what Watson gets right and wrong. VentureBeat, 18 February 2011
7. Assefi, M., Liu, G., Wittie, M.P., Izurieta, C.: An experimental evaluation of apple siri and google speech recognition. In: Proceedings of the 2015 ISCA SEDE, pp. 1–6 (2015) 8. Kepuska, V., Bohouta, G.: Next-generation of virtual personal assistants (Microsoft Cortana, Apple Siri, Amazon Alexa and Google Home). In: 2018 IEEE 8th Annual Computing and Communication Workshop and Conference (CCWC), pp. 99–103. IEEE, January 2018 9. Liu, Q., Huang, J., Wu, L., Zhu, K., Ba, S.: CBET: design and evaluation of a domain- specific Chatbot for mobile learning. Universal Access in the Information Society, pp. 1–19 (2019)
10. Sinha, S., Basak, S., Dey, Y., Mondal, A.: An educational Chatbot for answering queries. In: Emerging Technology in Modelling and Graphics, pp. 55–60. Springer, Singapore (2020) 11. Ghose, S., Barua, J.J.: Toward the implementation of a topic-specific dialogue-based natural language Chatbot as an undergraduate advisor. In: 2013 International Conference on Informatics, Electronics and Vision (ICIEV), pp. 1–5. IEEE, May 2013
12. Fadhil, A.: Can a Chatbot determine my diet? Addressing (the)challenges of Chatbot application for meal recommendation. arXiv preprint (2018)
13. Kar, R., Haldar, R.: Applying Chatbots to the internet of things: Opportunities and architectural elements. arXiv preprint (2016)
14. Arruda, D., Marinho, M., Souza, E., Wanderley, F.: A Chatbot for goal-oriented requirements modeling. In: International Conference on Computational Science and Its Applications, pp. 506–519. Springer, Cham (2019)

Keywords: NLP � Embedding � RNN � LSTM � GRU � DataSet � WikiSQL �SQL

1 Introduction

from SQLNet approach [10]; in particular, we employ a sketch to generate a SQL query from naturel language. The sketch aligns naturally to the syntactical structure of a SQL query; Neural Networks are then used to predict the content for each slot in the sketch. Our approach can be viewed as a neural network alternative to the traditional sketch based program synthesis approaches [11, 12].

2 Related Work

Fig. 1. Query structure

Our approach can be viewed as a neural network alternative to the traditional sketch based program synthesis approaches, so we also follow the slot filling. The idea is to employ a sketch to generate a SQL query from naturel language. The sketch aligns naturally to the syntactical structure of a SQL query; neural networks are then used to predict the content for each slot in the sketch, rather than to predict the content and the SQL grammar. As showed in Fig. 1, slots that will be predicted are tokens started by‘$’, so our purposed pipeline can be broken down into five modules ($AGG, $SEL-LCOL, $CONDCOL, $OP, $VALUE), each one has a model that predict it. We present below the details of the five modules.

4.1 AGG Module

4.2 SELCOL Module

The main purpose of this module is to find the appropriate select COLUMN given the natural language utterance; it is treated also as a classification problem also. Given the user question and the schema table as inputs, the model choose one column from the schema table and return it. Tokens of the two inputs are converted to embedding then passed to a bidirectional GRU, then we extract the hidden states of the GRU and pass it to another GRU layer. After that, we concatenate the representations, and we pass it to a dense layer with a relu activation function. Lastly, we pass all to a dense layer in which a softmax function is applied to give a score (probability) between 0 and 1 of each column, and the model return the column with high probability. Figure 3 shows the visualization of the particular module.

Fig. 4. VALUE Modules architectures

5 Technical Details

6.1 Results

We show results of executions and accuracy of each model on used DataSet and WikiSQL DataSet (Table 1).

1. Hillestad, R., Bigelow, J., Bower, A., Girosi, F., Meili, R., Scoville, R., Taylor, R.: Can electronic medical record systems transform health care? potential health benefits, savings, and costs. Health Aff. 24(5), 1103–1117 (2005)
2. Beck, T., Demirgüç-Kunt, A., Levine, R.: A new database on the structure and development of the financial sector. World Bank Econ. Rev. 14(3), 597–605 (2000)
3. Ngai, E.W.T., Xiu, L., Chau, D.C.K.: Application of data mining techniques in customer relationship management: a literature review and classification. Expert Syst. Appl. 36(2), 2592–2602 (2009)
4. Androutsopoulos, I., Ritchie, G.D., Thanisch, P.: Natural language interfaces to databases– an introduction. Nat. Lang. Eng.(1995)
5. Popescu, A.-M., Etzioni, O., Kautz, H.: Towards a theory of natural language interfaces to databases. In: Proceedings of the 8th International Conference on Intelligent User Interfaces, pp. 149–157 (2003)
6. Bakliwal, A., Arora, P., Patil, A., et al.: Towards enhanced opinion classification using NLP techniques. In: Proceedings of the Workshop on Sentiment Analysis where AI Meets Psychology (SAAIP 2011), pp. 101–107 (2011)
7. Mikolov, T., Karafiát, M., Burget, L., et al.: Recurrent neural network based language model.

In: Eleventh Annual Conference of the International Speech Communication Association (2010)

EMNLP (2018)
21. Warren, D.H.D., Pereira, F.C.N.: An efficient easily adaptable system for interpreting natural language queries. Comput. Linguist. 8(3–4), 110–122 (1982)
22. Li, Y., Yang, H., Jagadish, H.V.: Constructing a generic natural language interface for an xml database. In: EDBT, vol. 3896, pp. 737–754. Springer (2006)
23. Giordani, A., Moschitti, A.: Translating questions to SQL queries with generative parsers discriminatively reranked. In: COLING (Posters), pp. 401–410 (2012)
24. Wang, C., Cheung, A., Bodik, R.: Synthesizing highly expressive SQL queries from input- output examples. In: Proceedings of the 38th ACM SIGPLAN Conference on Programming Language Design and Implementation, pp. 452–466 (2017)
25. Li, F., Jagadish, H.V.: Constructing an interactive natural language interface for relational databases. In: VLDB (2014)
26. Yaghmazadeh, N., Wang, Y., Dillig, I., Dillig, T.: SQLizer: query synthesis from natural language. In: Proceedings of the ACM Programming Languages, 1(OOPSLA), pp. 63:1– 63:26 (2017)
27. Popescu, A.-M., Etzioni, O., Kautz, H.: Towards a theory of natural language interfaces to databases. In: Proceedings of the 8th International Conference on Intelligent User Interfaces, pp. 149–157. ACM (2003)
28. Giordani, A., Moschitti, A.: Translating questions to SQL queries with generative parsers discriminatively reranked. In: COLING (2012)

SQL Generation from Natural Language Using Supervised Learning 183

Department of Computer Science, Mohammedia Engineering School (EMI), Mohammed V University, Rabat, Morocco
{yassinbendriss,youssefhamdaoui}@research.emi.ac.ma,

186 Y. Bendriss et al.

be significant, around 30% and these systems will be able to anticipate needs and situations and react to the environment around them. In this paper, we propose an dynamic distribution concept based on some parameters collected by smart sensors [6] and consumers classification for smart buildings through an smart main controller focusing on diverse requirement satisfaction, including response time, energy effi-ciency, real time meter measurement, continuous supply, reliability and end’s user needs. The smart power management system presented considers an smart distribution network that collect real-time data from a physical sensors and meters through a connected house controller that have a role of power switcher from on/off power use in order to satisfy comfort conditions of houses, saving energy and reduce costs.

The duration of interruptions might be from minutes to hours depending on the severity of the fault that occurred and costs of an outage are not recoverable called outage Costs. Outage detection is achieved by using islanding detection methods (IDM) as shown in Fig. 1.

The Cyber-Physical Energy Systems represent a new class of widely distributed and globally interconnected energy systems that integrate computation processes, com-munication processes and control processes [11]. For example when we are in sunny case we can detect rain with sensor and the CPS make a decision base on this context aware information to open all windows in building through controllers and actuators.

The CPS layered architecture should include an information-centric protocol stack to support data fusion for making the data into the network and converting to high-level information for applications as shown in Fig. 2 [11].

Table 1. CPS properties [14]

Smart building is environment where cooperation of objects is allowed (e.g., sensors, devices, appliances) and systems that have the capability to self-organize themselves. All connected smart components are controlled by a main controller as

An Intelligent Power Distribution Management 189

2.4 Smart Home Concept

The energy aware smart house is in interaction with internal and external environments. The external environment consists of all the entities belonging to the SB and the internal consists of all appliances and devices belonging to the smart home, which are centrally managed by a smart controller [15]. Smart Homes (SH) can significantly contribute not only integrating energy generation system but also properly scheduling tasks to maximize the consumption of locally-generated energy and reduce the energy demand to the grid during peak hours. Controller is a smart sophisticated component which is responsible for execution actions making by the CPS and executing decisions and centralize all information transmitted by the smart sensors, meters, devices and actuators inside house [16, 17] as shown in Fig. 4.

Fig. 4. House smart components

2.5 Hybrid Distributed Energy Renewable

Smart power balancing in building is based on integration of distributed resources and technologies. We propose a new building model to analyze and discuss our approach as shown in Fig. 5. The smart building contain a main switcher to allow balancing between main grid power and self building power production/stored, the decision depend on the building context (loads, events, level batteries charge, solar panel pro-duction, weather, consumptions..). The DERs are installed on building rooftop for to ensure security conditions and we have a hybrid installation with inverter. The CPSM is the decision maker and all decisions are based on existing power (stored or real production) and loads. CPSM can allow exporting surplus energy to the grid or use the main grid power, some decisions examples are cited bellow:

Case 1: When not using all power that the solar PV system is generating, then the system will ensure that any surplus energy is used to charge the battery.

192 Y. Bendriss et al.

Case 5: if we need to use more electricity, then the system is in the obligation to make decision to C0 to switch on for using main grid power.

The Distribution network contains different nodes as explain in the last section and each node has his weight into building. The selection construction model consists on the decisive parameters state in the selection layered tree, there is others parameters like house maturity, environment, weather context…. In this paper we base our approach on 4 parameters:

• Costumer Category (CC).

The classification is dynamic and depends on entity last declaration or the CPS parameters for public entities like common power area.

3.3 Entity Selection Scheme Algorithm

Figure 7 present an IEEE Bus model to explain our optimization strategy after building our islanded layered tree and to decrease the number of islanded powered loads.

The first step is search islanded group like O3 and O2 in Fig. 7(a) that present any loads between this two groups then we group O3 and O2 I one Tree T2.

Fig. 6. Islanding algorithm

An Intelligent Power Distribution Management 195

Our Approach is implemented with Java language, Mysql Database and based on house’s consumption generated by Gridlab-d simulator. To have a real time simulation we run n simulation with different emergency due to a random function existing in java in the Math class, which returns a random number value in the range of zero (inclusive) and n (exclusive). The category is determined on the basis of the random function return a.

• IF 0 � a < 0.2 THEN CC = 1
• IF 0.2 � a < 0.5 THEN CC = 2
• IF 0.5 � a < 1 THEN CC = 3
As shown in Table 2, simulation is run for a building contains 15 houses and some DGs. Costumer classification is generated by a random function and loads are con-figured and generated by Gridlab-D simulator to have real cases. The test system adopts reactive local compensation, all power loss is ignored. We suppose for this simulation a total DGs active power/kWh 300 kWh for one day power (PV/Storage Batteries) that can be ensured by a TESLA batteries installation.

According to the selection scheme algorithm proposed in the paper, the simulation steps and results are presented as follows.

The building contains 15 houses with different classification and contains some DERs, the algorithm search the entities with high priority and start looping on the query results, E1, E3, E14 are entities with classification CC1, for each one we verify if his obligatory demand minus his self production Cp is more than existing updated power P (total), if yes we add the entity to our vector, else we skip and loop for next entity. When we add an entity we update the P(total) = P(total) – P + Cp. After finish with entities with CC1, we look for entities with CC2, in our case we have E2, E4, E5; E7; E10; E13; E15 are entities with classification CC2, for each one we verify if his obligatory demand minus his self production (P – Cp) is more than existing updated power P(total), if yes we add the entity to our vector, else we skip and loop for next entity. When we add an entity we update the P(total) = P(total) – P + Cp. After finish with entities with UC2, we look for entities with CC3, in our case we have E6; E8; E9; E11; E12 are entities with classification CC3, for each one we verify if his obligatory demand minus his self production (P – Cp) is more than existing updated power P(total), if yes we add the entity to our vector, else we skip and loop for next entity. When we add an entity we update the P(total) = P(total) – P + Cp. After finish with last classification for obligatory demand, we verify the existing power depletion, in our case, after satisfy all selected entities p(total) = 300 − 221 = 79 kWh, then we look to power interrupted demand (Q) but we loop into our selected vector by order of selection, in our case we can satisfy E1; E3; E14; E2; E4; E5; E7; E10; E13; E15, because the sum of their Q is 79 kWh. The islanding algorithm exit when power depleted or when covers all entities.

This paper presents an efficient approach and a smart distribution selection in an islanded building operation and proposes a solution to outage problem in regular intervals. The objective is to ensure a continuous power supply during outage. The proposed approach is based on a smart dynamic power balancing with dynamic parameters to control and perform the distributed energy in islanded building through a smart CPS and achieve a 100% self-powering. A next step of work is to make a model/prototype with Gridlab-D to simulate real scenarios and take more parameters in our islanding selection like weathers, building events, prosumers context, building aware to predict future loads and based on this predication information we will make a real smart decision to avoid outage problem and satisfy all demands and ensure a continuous power.

References

8. Hamdaoui, Y., Maach, A.: Dynamic balancing of powers in islanded microgrid using distributed energy resources and prosumers for efficient energy management. In: 2017 IEEE Smart Energy Grid Engineering (SEGE). IEEE (2017)
9. Guerrero, J.M., Vasquez, J.C., Matas, J., de Vicuna, L.G., Castilla, M.: Hierarchical control of droop-controlled AC and DC microgrids—a general approach toward standardization.

IEEE Trans. Ind. Electron. 58, 158–172 (2011)
10. Hamdaoui, Y., Maach, A.: Smart islanding in smart grids. In: 2016 IEEE Smart Energy Grid Engineering (SEGE), pp. 175–180. IEEE (2016)
11. Ge, Y., Dong, Y., Zhao, H.: A cyber-physical energy system architecture for electric vehicles charging application. In: 2012 12th International Conference on Quality Software (QSIC), pp. 246–250. IEEE (2012)
12. Gurgen, L., Gunalp, O., Benazzouz, Y., Gallissot, M.: Self-aware cyber-physical systems and applications in smart buildings and cities. In: Proceedings of the Conference on Design, Automation and Test in Europe, pp. 1149–1154. EDA Consortium (2013)
13. Nallamothu, B.K., Selvam, C., Srinivas, K., Prabhakaran, S.: Study on energy savings by using efficient utilites in buildings. In: Communication, Control and Intelligent Systems (CCIS), 2015, pp. 477–481. IEEE (2015)
14. Hamdaoui Y., Maach A.: Energy efficiency approach for smart building in islanding mode based on distributed energy resources, advanced information technology, services and systems. In: Lecture Notes in Networks and Systems (LNNS), vol 25, pp 36–49. Springer, Cham (2018)
15. Komninos, N., Philippou, E., Pitsillides, A.: Survey in smart grid and smart home security: issues, challenges and countermeasures. IEEE Commun. Surv. Tutor. 16, 1933–1954 (2014) 16. Khanna, A.: Smart grid, smart controllers and home energy automation—creating the infrastructure for future. Smart Grid Renew. Energy 03, 165–174 (2012)
17. Hamdaoui, Y., Maach, A.: Ontology- based context agent for building energy management systems. In: Advanced Intelligent Systems for Sustainable Development, vol. 13. Springer, Cham (2018)
18. Hamdaoui, Y., Maach, A., El Hadri, A.: Autonomous power distribution system through smart dynamic selection model using islanded micro grid context parameters and based on renewable resources. Int. J. Mech. Eng. Technol. (IJMET) 9(11), 1755–1780 (2018) 19. Al-Sarawi, S., Anbar, M., Alieyan, K., Alzubaidi, M.: Internet of Things (IoT) communi- cation protocols. In: 2017 8th International Conference on Information Technology (ICIT), pp. 685–690. IEEE (2017)
20. Hamdaoui, Y., Maach, A.: An intelligent islanding selection algorithm for optimizing the distribution network based on emergency classification. In: 2017 International Conference on Wireless Technologies, Embedded and Intelligent Systems (WITS), pp. 1–7. IEEE (2017) 21. Ma, T., Yang, H., Lu, L.: A feasibility study of a stand-alone hybrid solar–wind–battery system for a remote island. Appl. Energy 121, 149–158 (2014)
22. Pang, C., Dutta, P., Kezunovic, M.: BEVs/PHEVs as dispersed energy storage for V2B uses in the smart grid. IEEE Trans. Smart Grid 3, 473–482 (2012)
23. Kiviluoma, J., Meibom, P.: Methodology for modelling plug-in electric vehicles in the power system and cost estimates for a system with either smart or dumb electric vehicles.

1 Introduction

Integration of Artificial Intelligence (AI) and the Internet of Things (IoT) technologies in the healthcare industry is becoming an effective practice [1]. IoT basically consists of internetworking of physical objects operating in locations not previously connected, enabling to collect and exchange significant amounts of data over time [2]. While AI includes machine learning (ML) and deep learning (DL) tools used for many core tasks, such as disease prediction, diagnostics, medical decision making, home health, and remote artificial doctor [3]. Particularly, combine AI into the IoT system plays a vital role in remote patient monitoring. As patient data increases, the proper management and analysis of these data by IoT devices are helpful for both physicians and healthcare industries. Indeed, AI components incorporate intelligent decision making and rea-soning algorithms with data from different sources enabling rapid processing and complex analysis of data [4]. However, this combination of different types of tech-nologies in smart IoT system requires to create architecture for the whole system by

• Reuse: There are architectures to support the reuse of components and frameworks. Thus, promote reusability of domain-specific architecture and patterns.

• Construction: An architectural description is a blueprint of a system. It indicates the major components and dependencies between them. For instance, a layered view of architecture typically documents abstraction boundaries between parts of a system’s implementation, identifies the internal system interfaces, and constrains the entire system function through connection with other components.

Specifically, in the purpose of patient telemonitoring, a smart IoT system must solve a set of tasks linked with recording signals characterizing the activity of the body systems, processing and analyzing biomedical information, assessing ongoing body status, identifying the dynamics of changes, and predicting disease exacerbations. The solution of this set of tasks clearly requires a multi-level structure, each level opti-mizing the solution of a particular task [4].

The aim of this work is to design a conceptual architecture of a knee rehabilitation exercises telemonitoring system. This architecture is the conceptual model that defines the structure, behavior, and more views of hardware, software, and networks hanging together to become a system.

Fig. 1. Typical IoT architecture-based healthcare system [7]

• The third level of the system is the server of the medical institution which collects the data from different sources including, health-related data from patients’ devices as well.

• The fourth level is a device of the physician with appropriate software for decision support in data analysis.

To ensure reliable long-term registration of biomedical signals, the wearable device should have a high level of operational autonomy. Therefore, wearable devices not capable of complicated signal processing that requires high-performance multipro-cessors and a large power supply. Thus, the use of intelligent algorithms for recording and transmitting biomedical signals at the next level of the system – signal receiver gateway – significantly decreases current consumption and increases device autonomy. The radio signal transmission power from the wearable device to the gateway must be low energy. IEEE 802.15.4 standard provides a solution for low-rate low-power

204 M. El Fezazi et al.

2.3 Cloud-Based Medical Institution Server

Cloud-based tools and technologies of the medical institution server present the third level of the system. This is the level at which detailed analysis of the dynamics of the functional status of the body systems and integral assessment of health status is per-formed. Also, the cloud database stores related information, including individual normal values of the patient, the criteria for disease diagnosis and prediction of exacerbations. Reports of functional changes in the body systems constituting threats to the patient’s health are generated for the treating doctor. The target functions of the third level of the system are to conduct long-term monitoring of patient health and predict the exacerbation of patients’ diseases [8, 11].

In the context of knee telerehabilitation, we propose a conceptual model that defines the structure, behavior, and more views of hardware, software, and networks hanging together to become a system. This model is designed premised upon the structure of the smart system for patient remote monitoring. It is suggested in order to ensure the telemonitoring function of the knee rehabilitation process [2]. The architectural model of the system environment consists of four main units: wearable measurement unit, gateway processing unit, data aggregation-analysis unit, and monitoring-visualization unit, as illustrated in Fig. 3.

206 M. El Fezazi et al.

the shank segment as one side of this angle and the longitudinal axis of the thigh segment as the other side, as shown in Fig. 4. Thus, thigh and shank segment angle should be obtained, and then the relative knee angle should be calculated using the following Eq. (1) [12]:

Fig. 4. Definition of joint angles and coordinates

3.2 Gateway Processing Unit

In order to compensate for the drift of the integrated angular velocity, a fusion of these sensors allows an improved estimate of the actual angle, where their purpose is a combination of rotational information from the accelerometer and with the integrated gyroscope measurements [13]. They are several methods to achieve this, such as complementary filer, adaptive-based algorithms, Kalman Filter-based algorithms [15].

For instance, we propose an implementation of the Kalman Filter (KF) technique in the gateway unit. Basically, KF-based technique consists of a loop that contains two steps. Prediction step, getting data from gyroscope reading, and then the correction step using the measurements from accelerometer [15]. The Fig. 5 below illustrates the process loop for IMU sensor fusion [14]. The obtained knee angle is transmitted via WLAN channel to the cloud server for aggregation and a detailed analysis of data.

The transmitted data is published and stored in a cloud server space. This level allows feedback, adherence, and physician supervision. It can integrate more analysis algo-rithm such as a classification algorithm for knee exercise recognition in order to evaluate the correctness of the rehabilitation actions.

3.4 Monitoring-Visualization Unit

1. Darwish, A., Hassanien, A.E., Elhoseny, M., et al.: The impact of the hybrid platform of internet of things and cloud computing on healthcare systems: opportunities, challenges, and open problems. J. Ambient. Intell. Human Comput. 10, 4151–4166 (2019).
2. Fezazi, M.E., Aqil, M., Jbari, A., Jilbab, A.: IoT-based knee rehabilitation system for inclusive smart city. In: Proceedings of the 4th International Conference on Smart City Applications (SCA 2019). ACM, New York (2019). Article 82, 6 p.
3. Maddox, T., Rumsfeld, J., Payne, P.: Questions for artificial intelligence in health care.

JAMA 321 (2018).
4. Adhikary, T., Jana, A.D., Chakrabarty, A., Jana, S.K.: The Internet of Things (IoT) aug- mentation in healthcare: an application analytics. In: Gunjan, V., Garcia Diaz, V., Cardona, M., Solanki, V., Sunitha, K. (eds.) ICICCT 2019 – System Reliability, Quality Control, Safety, Maintenance and Management. ICICCT 2019. Springer, Singapore (2020)

. Accessed 15 Jan 2020
15. Du, J., Gerdtman, C., Lindén, M.: Signal quality improvement algorithms for MEMS gyroscope-based human motion analysis systems: a systematic review. Sensors 18, 1123 (2018).

A WBAN Platform for Healthcare: Real-Time Remote Monitoring of Human Vital
Parameters

1 Introduction

In Morocco, hypertensive diseases and heart disease caused more than 20.3% of deaths in 2014 [1] and the elderly has become 9.4% of the population [2], yet the country still suffers from a shortage of medical staff and equipment [1]. In order to overcome this situation, several development and research programs including our research work are being launched to build intelligent health care systems that helps caregivers to monitor the health state of patients remotely [3, 4].

The aim of this research is to develop a WBAN platform using of four sensor nodes to measure physiological signals: cardiac electrical activity (ECG), heart rate, arterial oxygen saturation (SPO2), respiration rate and body temperature [8]. The sensor nodes produce simultaneously synchronous signals [3], which will be collected in a sink node and then sent to a gateway node through a wireless communication. Therefore, we designed each node to be able to acquire the physiological signal using a biomedical sensor, to convert and process the signal using a microcontroller and to communicate the signal data to the sink node and gateway node by means of radiofrequency module NRF24L01. Furthermore, the choice of the transceiver module and the network topology applied in this infrastructure represent a good ratio of flow rate to energy consumption.

First of all and in order to test this platform, we have built a prototype of four sensor nodes, then we have tested the signal acquired by the sensor nodes one by one (transmitter), and the transmission of the signal to the sink node (receiver). Secondly, we have extended the test platform to form a wireless network by putting the four nodes in service simultaneously and collecting the data in sink node, which retransmit it to the gateway node. Then, we display and record the received data using a graphical interface developed under Labview in PC monitoring.

212 E. Baba et al.

Mobility of Nodes: The big advantage of the WBAN is that the non-wired network, i.e. the patient and the doctor can move freely, but it poses enough challenges to beat at the level of power supply, transmission protocols errors and multi-hop routing protocols.

Safety: The worry about safety is evident with this sensitive data on the health status of patients. With the regulatory obligations, the measures to protect the patient’s pri-vacy are applied on the biomedical devices. Thus, an encryption of data circulating in the network is carried out to ensure data protection.

We propose in this paper a WBAN prototype platform contributing to the improvement of performance measures for the requirements mentioned above, this hardware and software infrastructure will be subject to detailed studies in the following sections.

addition, this platform allows the patient to move freely in an area of 100 m [14], so it can be integrated into a medical care system in a health monitoring center.

3.2.1 Processing Stage
We used an ATmega328P 16 MHz microcontroller for the data processing in each node [11]. The microcontroller is the core of the proposed design for all nodes, thus it is exploited to implement the program managing all units of the node. In addition, the microcontroller is equipped with an e-Health Sensor Shield V2.0 for multiplexing and converting the signals captured by biomedical sensors [13]. The microcontroller send the processed data via SPI (Serial Peripheral Interface) to the transceiver module.

Following figure (Fig. 3) shows the flow charts of the algorithms implemented in the microcontroller of each node. For the algorithms of the sensor nodes, it performs almost the same functions except the sensing part, which depends on sensor (a). However, the

Fig. 3. The flow charts of algorithm platform: sensor nodes (a) sink node (b) and gateway node.

The reason behind the selecting of biomedical sensors is the overwhelming importance of the provided information on a patient’s health state [4]. These infor-mation will certainly serve in the accurate diagnosis of diseases and will ensure the correct interpretation of the signals detected by the sensors. As well as, the doctors can be exploit these to prescribe a better treatment. In the following, we present the biomedical sensors experienced in this platform and their reason of choice:

Body Temperature Sensor: Body temperature measurement can help in the diagnosis and the treatment performed by the doctor, especially in the case of diseases accom-panied by characteristic changes in body temperature [9–13].

The module nRF24L01 uses NORDIC Semiconductor’s “ShockBurst” communi-cation protocol, which allows automatic packet assembly, synchronization, transmis-sion error management and automatic retransmission in case of non-response of the destination node [15]. That guarantees effectively a very low power consumption and a high reliability in this network.

3.2.4 Power Supply
For powering the wearable sensor nodes, we used a battery (11.1 V 2200 mAh) connected with processing stage to power the node units. The consumption of nodes is low, as we have seen; therefore, the batteries live time will be higher. However, we used the serial port of PC monitoring for powering the gateway node.

Fig. 4. Displaying of vital parameters on VI graphical interface.

The acquisition of the five signals is done continuously through the serial port of PC monitoring, which displayed the results in real time. Moreover, each change or inflection on a parameter appears clearly and immediately on the waveform corre-sponding to this parameter. The first box named SPO2 displays the arterial oxygen saturation, the second box named BPM displays the average heartbeat per minute. The third is in the form of a bar named Temperature displays the body temperature. The fourth is in the form of a graph called ECG displays the graphical representation of cardiac electrical activity and the last graph named Airflow signal, which is the tracing of breathing within the time. In additionally, we records the values of each physiologies parameter in a file.

The results obtained by this prototype of the WBAN platform are satisfactory, whether in terms of the measurements accuracy in real time, or in terms of reliability and continuity of remote monitoring. The objectives achieved are very encouraging because we have successfully monitored five vital parameters. In addition, the platform provide the possibility of capturing other physiological signals by addition of other sensors without any major changes in the design of the platform. In fact, an improved version of this platform will be very useful for implementation in a healthcare system of the patients, especially in a health-monitoring center.

Acknowledgements. This work was supported by the Moroccan Ministry of Research and the National Center for Scientific and Technical Research (CNRST) under the project “Réseaux de capteurs sans fil biomédicaux” Grant no. PPR/2015/45.

References

IEEE Pulse 7(1), 21–25 (2016)
11. Arduino (2017). . Accessed 22 Sept 2017 12. Shivwanshi, R.R., Ahuja, C., Abrol, P., Singh, B.: Design and development of wireless sensor network for biomedical application. In: International Conference on Advances in Computing, Communication & Automation (ICACCA) (Fall), 30 September 2016, pp. 1–6.

IEEE (2016)

Energy Management Strategy of a Fuel-Cell Electric Vehicle Based on Wavelet Transform

Nada Rifai1,2(&), Jalal Sabor1, and Chakib Alaoui3

The automobile industry is one of the major sources of environmental problems. The steady increase of automobiles number causes serious problems for the environment and the human life. Air pollution, global warming, and the rapid depletion of the Earth’s petroleum resources are considered as the major challenges for the actual and future generations. There has been numerous research and development (R&D) activities on electric vehicle development (EVs) over the past twenty years [1, 2], because this technology is promising for better energy conversion efficiencies and the reduction of greenhouse gasses emission, by displacing the energy demands from crude oil to electricity. Electrochemical batteries, especially sealed lead acid at first [3] and lithium-ion more recently [4], have drawn a lot of attention as well for their possibility to feed the vehicle. Moreover, hydrogen power delivered by the fuel cells vehicles seems to be an additional alternative for the internal combustion engines [5].

© Springer Nature Switzerland AG 2021
T. Masrour et al. (Eds.): A2IA 2020, LNNS 144, pp. 220–235, 2021.

This paper is organized as follows. Section 2 discusses the need to couple a fuel-cell based energy source with a complementary power source in order to improve the performance of the power source and presents the energy sources models used in simulation. Section 3 presents a solution based on wavelets transform in order to control a hybrid power source that is composed of a fuel-cell and a super-capacitor for a vehicular application. Section 4 presents simulation results of the proposed algorithm. This paper ends with a conclusion.

2 Hybridization of the Fuel Cell

However, the operating conditions are the most affecting fuel cell lifetime. About 56.5% of fuel cell degradation are caused by load changing, 33% are caused by start-stop conditions, 5.8% by high power and 4.7% by idling conditions [6].

222 N. Rifai et al.

Figure 1 shows a model of the hybrid energy system used in this study. It is consisted of a fuel cell connected to a unidirectional DC/DC converter, an ultraca-pacitor connected to a bidirectional converter. An optional bidirectional DC/DC con-verter between the DC bus and the load is used to keep the DC load voltage stable, the wavelet Algorithm which determines the required power of the fuel cell and ultraca-pacitor, and a controller to control the converters for power tacking and voltage reg-ulation. The next section focuses on the integration of a load distribution algorithm based on wavelet decomposition to ensure an efficient power flow control strategy.

In the following section, the model of the fuel cell as well as that of the superca-pacitor will be presented.

In many studies, the fuel cell is modelled by an electrochemical model that describes its static behavior only [7, 17]. In this model, the output voltage of a single cell, which can be defined as follows:

Uconc ¼ B � ln 1 � Ifc=iL� � ð2Þ

Activation Zone: This zone causes thermodynamic losses due to water resistance in the pores of the material. The role of the catalyst is to accelerate the reaction. The activation voltage drop is expressed as the Eq. (3).

224 N. Rifai et al.

The fuel cell will be represented by a parallel RC circuit. There is a first order delay in the activation and the concentration voltage components, which are represented by the resistances Rconc and Ract Fig. 2. This delay is caused by the charge double layer effect. It acts just like an electrical capacitor. Then, when there is an increase in the FC current, there is a delay until the FC voltage decreases. The ohmic over-potential is not affected by the charge double layer effect as it is directly related to the current. Based on this model, the output voltage of the fuel cell is shown in Eq. (5), (6) and (7).

Fig. 2. Equivalent electric circuit of the fuel cell

Fig. 4. Polarization curve of the fuel cell

226 N. Rifai et al.

Fig. 6. Electric model of supercapacitor

3.1 Fuzzy Logic

Energy management based on fuzzy logic supervisor is one of the most used strategies especially for a hybrid system composed of fuel cell, battery and ultracapacitor. This energy system is very complex and having several variables. The fuzzy logic is suitable for managing this type of system [10]. A. Augusto Ferreira proposed in [11] a new topology for the energy system using a multiple-input power electronic converter (MIPEC). This solution has low cost and more efficiency and flexibility. D. André in [15] used a topology where the fuel cell supplies the power to the load via the uni-directional DC/DC converter, the ultra-capacitor pack is connected to the dc voltage bus via the bidirectional DC/DC converter and the battery pack is connected directly to the DC voltage bus. Both of them worked on a supervisor based on fuzzy logic. In [15], a rule based algorithm was implemented.

Fig. 9. Load power profile corresponding to Artemis Drive cycle and its time/frequency representation

3.2 Artificial Neural Network

In his study [14], Z. Jiang presented an adaptive control strategy to balance the power flow between the fuel cell and the battery. The goal of this approach is to achieve maximum efficiency of the fuel cell while taking into consideration the state of charge of the battery in order to avoid its rapid depletion. The main idea is to regulate the fuel cell current while limiting the voltage of the battery, since there is a direct relationship between the SOC and the voltage of the battery. Three regulation modes were proposed in this paper: fuel cell current limit, battery voltage limit, and battery current limit. A state machine was implemented in Matlab to define the different regulation modes and the conditions to swipe from a state to another. A proportional-integral controller was used to regulate the currents and voltages.

H. Alloui presented in [20] a power management strategy based on frequencies separation. Two PI controllers were implemented in this strategy. One is to regulate the DC bus voltage, and to give the DC bus voltage and to give the DC bus current reference. A low pass filter is then used to obtain the fuel cell current. The second PI controller is for fuel cell current; it gives the control voltage of the boost converter. A PWM is applied on this control voltage to have the signal pulses sent to the IGBT of the boost chopper.

230 N. Rifai et al.

The load power profile in automobile environment is a nonstationary fluctuating and transient signal. The wavelet transform is suitable for analyzing and evaluating such signals. Figure 9 present the load power profile corresponding to urban Artemis driving cycle, as shown in Fig. 8 for a Citroen Berlingo which characteristics are given in the Table 1. Artemis cycles are based on a statistical study carried out in Europe. They are not used for pollutant or fuel consumption certification; however, car man-ufacturers use this type of cycle to better understand real driving conditions and to evaluate the real performance of their vehicles.

Unlike Xi.zhang [18] and M. Uzunoglu [19] who used he Haar wavelet. In this paper, we worked on the Mexican Hat wavelet. One of the most used wavelet function for detecting and localizing transients is the normalized negative of the second derivative of a Gaussian function known as Mexican hat or Ricker wavelet. Its expression is given in (10).

A diode at the output of the fuel cell stack is necessary to prevent the negative current going into the stack. Due to the negative current, it is possible that the cell

Energy Management Strategy of a Fuel-Cell 231

Figure 13 shows that the decomposed components of the load profile during a drive cycle are distributed to the FC, and ultracapacitor, compatible with their respective characteristics in order to obtain better performance. In fact, the FC system satisfies the smoothed load demand, which is extracted from the total power demand using the wavelet-based algorithm and control system and the ultracapacitor satisfied the tran-sient load demand.

Table 1. Hybrid Citroen Berlingo characteristics

Fig. 11. The reconstructed signal from the wavelet Transform using the Ricker wavelet

The transient and high power demand are the main factors in the degradation of the fuel cell. To prevent the fuel cell from operating at a frequency outside its comfort zone, the power management strategy proposed in this article is based on a spectral study of the load power demand. It determine an optimal power distribution between the fuel cell and the secondary source according to their dynamic characteristics.

1. Alaoui, C., Salameh, Z.M.: On-board diagnostic and rejuvenation system for electric vehicles. In: International Symposium on Power Electronics, Electrical Drives, Automation and Motion, SPEEDAM 2006. IEEE (2006)
2. Alaoui, C., Salameh, Z.M.: A novel system-on-chip system for diagnostic & rejuvenation for electric & hybrid vehicles. In: 2003 IEEE 58th Vehicular Technology Conference, VTC 2003-Fall (IEEE Cat. No. 03CH37484), vol. 5. IEEE (2003)
3. Alaoui, C., Salameh, Z.M., Lynch, W.A.: Sealed lead acid electric vehicle batteries: a performance comparison study. Int. J. Power Energy Syst 21(3), 174–179 (2001)
4. Alaoui, C., Salameh, Z.M.: Electric vehicle diagnostic and rejuvenation system (EVDRS).

Int. J. Power Energy Syst. 27(2), 151 (2007)
5. Thounthong, P., Raël, S., Davat, B.: Energy management of fuel cell/battery/supercapacitor hybrid power source for vehicle applications. J. Power Sources 193(1), 376–385 (2009) 6. Yan, X., et al.: The study on transient characteristic of proton exchange membrane fuel cell stack during dynamic loading. J. Power Sources 163(2), 966–970 (2007)
7. Gong, W., Cai, Z.: Accelerating parameter identification of proton exchange membrane fuel cell model with ranking-based differential evolution. Energy 59, 356–364 (2013)
8. Erdinc, O., Vural, B., Uzunoglu, M.: A wavelet-fuzzy logic based energy management strategy for a fuel cell/battery/ultra-capacitor hybrid vehicular power system. J. Power Sources 194(1), 369–380 (2009)
9. Zhang, X., Mi, C.C., Masrur, A., Daniszewski, D.: Wavelet-transform-based power management of hybrid vehicles with multiple on-board energy sources including fuel cell, battery and ultracapacitor. J. Power Sources 185(2), 1533–1543 (2008)
10. Gao, D., Jin, Z., Lu, Q.: Energy management strategy based on fuzzy logic for a fuel cell hybrid bus. J. Power Sources 185(1), 311–317 (2008)
11. Ferreira, A.A., Pomilio, J.A., Spiazzi, G., de Araujo Silva, L.: Energy management fuzzy logic supervisory for electric vehicle power supplies system. IEEE Trans. Power Electron. 23 (1), 107–115 (2008)
12. Xie, C., Quan, S., Chen, Q.: Control strategy of hybrid power system for fuel cell electric vehicle based on neural network optimization. IEEE Xplore (2008)
13. Moreno, J., Ortúzar, M.E., Dixon, J.W.: Energy-management system for a hybrid electric vehicle, using ultracapacitors and neural networks. IEEE Trans. Ind. Electron. 53(2), 614– 623 (2006)
14. Jiang, Z., Gao, L., Dougal, R.A.: Adaptive control strategy for active power sharing in hybrid fuel cell/ battery power sources. IEEE Trans. Energy Convers. 22(2), 507–515 (2007) 15. Rodatz, P., Paganelli, G., Sciarretta, A., Guzzella, L.: Optimal power management of an experimental fuel cell/supercapacitor-powered hybrid vehicle. Control Eng. Pract. 13, 41–53 (2005)
16. Njoya Motapon, S., Dessaint, L.A., Al-Haddad, K.: A comparative study of energy management schemes for a fuel-cell hybrid emergency power system of more-electric aircraft. IEEE Trans. Ind. Electron. 61(3), 1320–1334 (2014)
17. Thounthong, P., Pierfederici, S., Martin, J.P., Hinaje, M., Davat, B.: Modeling and control of fuel cell/supercapacitor hybrid source based on differential flatness control. IEEE Trans.

Natural Language Processing: Challenges

and Future Directions

Abstract. Natural language processing (NLP) is a well-known sub-field

of artificial intelligence that is having huge success and attention in recent

NLP, then we dive into the different challenges that are facing it, finally,

we conclude by presenting recent trends and future research directions

NLP is having big momentum and huge progress in recent years, mostly thanks to deep learning [41], this progress was driven primarily by the big tech companies who are collecting a huge amount of structured and unstructured text data, that data needs to be processed, analyzed and exploited, in the last two years, these big tech corporations advanced the state-of-the-art on numerous NLP tasks by some very sophisticated models.

Despite the enormous progress of NLP in recent years, there are still some tremendous challenges and issues, some of these challenges are related to deep learning and some are specific to NLP, presenting and describing these challenges is the core of this paper.

Research in NLP began more than six decades ago and can be divided into 3 waves: rationalism, empiricism, and the current deep learning wave [41], it’s this last wave that is behind the recent remarkable success of NLP, in the last two years alone, deep learning-based NLP models like BERT [7], RoBERTa [21] and ERNIE [44] that were released especially from the big tech companies made impressive progress in the state-of-the-art of many NLP tasks.

Furthermore, the big number of NLP’s commercial applications like personal voice assistants, chatbots, automatic online translation etc. . . is also a driving force behind the recent explosion of research in NLP. The field of NLP can be further divided into sub-fields like syntax, semantics, dialogue systems, speech recognition among others, each one of these sub-fields hosts a number of tasks, a list of major tasks in NLP grouped by category is shown in Fig. 1. The different NLP tasks shown in Fig. 1 all have a wide range of real-world applications, some of the major ones are:

– Social media monitoring: here, NLP is used in brand reputation moni-toring, product reviews processing and analyses, clients sentiment analyses etc...

– Search engines: search engines are using NLP tasks like question answering to provide better direct responses to how/why question, NLP is also used in question auto-completion and question spelling correction.

NLP

Fig. 1. NLP major tasks

– Extracting semantic meanings from text and capturing the linguistic or the semantic relationships between pairs of vocabulary terms.

– The difficulty of deriving context
– Word sense disambiguation: identifying the meaning of a word according to its context
– Coreference resolution: identifying all expressions that refer to an entity in a given text, for example: “Zakaria sent a letter to Youssef telling him that. . . ” him refer to whom.

3.2 Natural Language Generation Related Challenges

Natural language generation (NLG) is the task concerned with generating mean-ingful phrases in the form of natural language. Considerable progress has been made in NLG by using the different deep learning architectures like recurrent neural network (RNN), long short-term memory (LSTM) and more recently transformers, one of the most famous examples of the progress that has been made is the GPT-2 language model [31].

240 Z. Kaddari et al.

the use of Positive-Unlabeled Learning (PU-Learning) for learning word embed-dings [15], the use of data augmentation techniques [33] and the use of meta-learning [12].

Most machine learning/deep learning models provide no explanations for their predictions, the same goes for deep learning-based NLP; this is why these deep learning models are considered as “Black boxes”, the impossibility for these mod-els and techniques to provide clear explanations for their predictions is currently a big issue, especially in some sensitive areas like medical NLP, job resumes processing among others.

Several techniques are being investigated to overcome this issue like sensitiv-ity analyses (SA) [5,36], layer-wise relevance propagation (LRP) [4], embedding of fuzzy logic systems [23,24,26] into deep learning networks or the use of a generative explanation framework [20].

Current NLP systems are predominantly unimodal, meaning that they are only capable of processing and analyzing linguistic inputs, humans on the other side

Natural Language Processing: Challenges and Future Directions 241

A dialogue system, also called conversational agent is an NLP system designed to have a conversation with a human being, there are two types of dialogue systems: open-domain dialogue systems like general chatbots or personal voice assistants and task-oriented dialogue systems like customer support or restaurant reservation dialogue systems.

Due to their wide range of commercial applications like personal assistants and chatbots, dialogue systems have attracted a big number of researchers in academia and especially in the industry, therefore, this sub-field of NLP pro-gressed immensely.

– Incorporating emotion and sentiment analyses.

– Improving open-domain dialogue systems.

– Multiple-intent identification in one question: dialogue systems can detect and classify the intent of a question, but are still failing when it came to multiple-intent identification.

4 Future Research Directions

The focus of current research trends is mainly tackling the challenges described in Sect. 3, as for future trends, several research directions are being speculated:

1 .

Natural Language Processing: Challenges and Future Directions 243

Quantization: quantization as a technique that compresses neural networks to a reasonable size, while also achieving high-performance accuracy, this technique is a requirement for models turning on edge-devices where the memory size and the computing power are limited. One of the future research directions in NLP is also applying quantization to NLP models.

Neural Symbolic Architectures: also referred to as hybrid neural networks, is an old approach that is gaining strong traction recently, this approach tries to combine elements of symbolic computation and fuzzy logic [9,28] rule based techniques with artificial neural networks into one model, the goal here is to benefit from advantages of both paradigms while avoiding their shortcomings. A recent success of this approach is being able to perform elaborated tasks in mathematics, such as symbolic integration and solving differential equations [17], this hybrid approach is one of the most promising future research directions in NLP.

244 Z. Kaddari et al.

many shortcomings and limitations, thus the need for a new generation of deep learning models for NLP or maybe experimenting with hybrid models where deep learning is only one component of two or many more.

Natural Language Processing: Challenges and Future Directions 245

14. Gururangan, S., Swayamdipta, S., Levy, O., Schwartz, R., Bowman, S., Smith, N.A.: Annotation artifacts in natural language inference data. In: Proceedings of the 2018 Conference of the North American Chapter of the Association for Com- putational Linguistics: Human Language Technologies, Volume 2 (Short Papers), New Orleans, Louisiana, June 2018, pp. 107–112. Association for Computational Linguistics (2018).
15. Jiang, C., Yu, H.F., Hsieh, C.J., Chang, K.W.: Learning word embeddings for low- resource languages by PU learning. In: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), New Orleans, Louisiana, June 2018, pp. 1024–1034. Association for Computational Linguistics (2018).
16. Kordjamshidi, P., Rahgooy, T., Manzoor, U.: Spatial language understanding with multimodal graphs using declarative learning based programming. In: Proceedings of the 2nd Workshop on Structured Prediction for Natural Language Processing, Copenhagen, Denmark, September 2017, pp. 33–43. Association for Computational Linguistics (2017).
17. Lample, G., Charton, F.: Deep learning for symbolic mathematics (2019) 18. Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., Soricut, R.: ALBERT: A lite BERT for self-supervised learning of language representations (2019) 19. Liu, C.W., Lowe, R., Serban, I., Noseworthy, M., Charlin, L., Pineau, J.: How NOT to evaluate your dialogue system: an empirical study of unsupervised evaluation metrics for dialogue response generation. In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, Austin, Texas, November 2016, pp. 2122–2132. Association for Computational Linguistics (2016).
20. Liu, H., Yin, Q., Wang, W.Y.: Towards explainable NLP: a generative explanation framework for text classification. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, July 2019, pp. 5570– 5581. Association for Computational Linguistics (2019).
21. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: RoBERTa: a robustly optimized BERT pretraining approach (2019)
22. McCoy, T., Pavlick, E., Linzen, T.: Right for the wrong reasons: diagnosing syn- tactic heuristics in natural language inference. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, July 2019, pp. 3428–3448. Association for Computational Linguistics (2019).
23. Murinov´a, P., Nov´ak, V.: A formal theory of generalized intermediate syllogisms.

A Genetic Algorithm Based Optimal Sizing Strategy for PV/Battery/Hydrogen
Hybrid System

Nouhaila Lazaar1,2(&), Eyman Fakhri1, Mahmoud Barakat1, Jalal Sabor2, and Hamid Gualous1

1 Introduction

At present, the use of gray energy produced by the combustion of fossil fuels such as oil, gas or uranium (mainly in France) is still prevalent. These sources are by nature non-renewable and require millions of years to replicate [1]. While the share of renewable energies, in the form of wind (8.2%), tidal (less than 4.4%) or solar (3.2%) energy in final electricity consumption in 2017 remains low [2]. Given the environ-mental impact and rising costs of fossil fuels, the use of renewables as an alternative solution is given great importance. Although these technologies are promising, their costs compared to conventional sources are high. Photovoltaic panels are regarded as the best generators in remote areas for harnessing the solar source [3, 4], Nevertheless, the main challenge when integrating this source into a microgrid is its intermittency.

In this paper, the optimal sizing and energy management of a standalone hybrid system are investigated. The PV arrays are considered as the primary energy source to supply a residential island load. The batteries and hydrogen storage systems are inte-grated into the microgrid in order to supply the load if an energy shortage arises. The optimal sizing is implemented in Matlab software, using GA, and performed for var-ious ELFmax values as problem constraints [19]. Furthermore, the impact of the hydrogen system cost on the sizing problem is studied.

The PV system is connected to the DC bus via a DC/DC converter. PV arrays output power can be determined as follows, depending on the solar radiation and ambient temperature [13]:

calculated by (2).

2.2 Battery Storage System

The battery state of charge (SOC) is controlled by the power generated by the PV arrays and the load demand. If the energy produced by the source far exceeds the load needs, the surplus energy is used to charge the battery until it’s fully charged. The excess will be transferred to the electrolyzer. The battery energy at any given moment t can be determined by (3) [13, 20].

Where r is the hourly rate of battery self-discharge, which depends on both the battery SOC and the battery state of health (SOH) [21], EL is the energy demand, ginv, gch band gdch b are, the inverter efficiency, the battery charging efficiency and the battery dis-charging efficiency, respectively, which are equal to 80%. Epv represents the energy produced by the PV arrays. In order to avoid the battery aging or damage, its SOC has to remain within its defined range.

Pelz ¼ Nelz:Ielz:Velz ð5Þ

Where Nelz, Ielz, Velz are the number of electrolyzers, the electrolyzer current (A) and its voltage (V), respectively.

The sizing aims to find the system optimal configuration with minimum cost (number of photovoltaic arrays Npv, batteries Nb, fuel cells Nfc, electrolyzers Nelz, and tanks Nh), considering ELF as a reliability criterion. For this purpose, the genetic algorithm is implemented in Matlab software.

3.1 Objective Function

order to convert the replacement cost of a component at the end of its lifetime to present

cost, Kj is used, it can be calculated by (9).

3.2 Reliability Criterion

The reliability of feeding the energy to the load is a primary factor in sizing standalone

fuel cells and Ebmin is the minimal energy stored in batteries.

1 � Nb � Nbmax ð14Þ

1 � Nelz � Nelzmax ð15Þ

Table 1. The GA parameters.

The optimal sizing of the proposed hybrid system was performed to guarantee a sus-tainable energy supply to residential loads. The hourly load consumption during a year (8760 h) is depicted in Fig. 4. The GA has been used to find the best configuration of the PV/battery/hydrogen system, considering NPC as an objective function and ELF as a reliability criterion. The control variables of the optimization problem are the number of PV arrays with 500 kW rated power, the number of batteries with 250 kWh rated energy, the number of electrolyzers and fuel cells with 1.5 MW rated power, and the number of 10 kg tanks. An energy management approach has been integrated into the optimization program to maintain a balance between energy generation and load consumption. The specifications of the system components considered in this study are summarized in Table 2 [24, 25].

The simulations were conducted for three different values of ELFmax (1%, 5% and 10%). The optimal sizing results obtained by GA are presented in Table 3. From the results, we can clearly see that the reliability index has a notable impact on both the NPC and the number of system components. As the ELFmax increases, the system’s overall cost and size decrease. Moreover, the amount of hydrogen stored in the storage tanks is reduced. The system NPC for 1%, 5% and 10% is equal to $28,8 MM, $27,7

Fig. 4. Household annual load profile.

Fig. 6. PV/battery/hydrogen system power profiles for one week.

Table 2. Technico-economical parameters of the system components.

One of the most daunting concerns of hybrid power systems is finding the optimal configuration with minimum cost. In this paper, the optimal sizing of a standalone PV/battery/hydrogen system is presented, using the genetic algorithm (GA). The hydrogen storage system is composed of an electrolyzer, a fuel cell and a hydrogen tank. The net present cost (NPC) has been chosen as an objective function, considering the equivalent loss factor (ELF) as a reliability index. The proposed algorithm has been performed for different values of ELFmax. The results showed that the increase in the reliability constraint results in a lower number of system components. Hence, the total cost is reduced. Also, the results cleared that the hydrogen system has a significant impact on the NPC. By reducing its components cost, considerable savings can be made.

2. Coltier, Y., Plouhinec, C.: Key figures for renewable energies, p. 92 (2019). (in French) 3. Fendri, D., Chaabene, M.: PV/batteries sizing and energy dispatching using Continuous Petri Net. Appl. Sol. Energy 53, 92–99 (2017). 4. Castañeda, M., Cano, A., Jurado, F., Sánchez, H., Fernández, L.M.: Sizing optimization, dynamic modeling and energy management strategies of a stand-alone PV/hydrogen/battery- based hybrid system. Int. J. Hydrogen Energy 38, 3830–3845 (2013).

5. Pavan Kumar, Y.V., Bhimasingu, R.: Renewable energy based microgrid system sizing and energy management for green buildings. J. Mod. Power Syst. Clean Energy 3, 1–13 (2015).

10. Durão, B., Joyce, A., Mendes, J.F.: Optimization of a seasonal storage solar system using Genetic Algorithms. Sol. Energy 101, 160–166 (2014).

11. Mahesh, A., Sandhu, K.S.: A genetic algorithm based improved optimal sizing strategy for solar-wind-battery hybrid system using energy filter algorithm. Front. Energy 14(1), 139– 151 (2017).
12. Mahesh, A., Sandhu, K.S.: Optimal sizing of a grid-connected PV/wind/battery system using particle swarm optimization. Iran. J. Sci. Technol. Trans. Electr. Eng. 43(1), 107–121 (2018).
13. Abdelshafy, A.M., Hassan, H., Jurasz, J.: Optimal design of a grid-connected desalination plant powered by renewable energy resources using a hybrid PSO–GWO approach. Energy Convers. Manag. 173, 331–347 (2018). 14. Rousis, A.O., Konstantelos, I., Strbac, G.: A planning model for a hybrid AC–DC microgrid using a novel GA/AC OPF algorithm. IEEE Trans. Power Syst. 35, 227–237 (2020).
15. Abdelkader, A., Rabeh, A., Mohamed Ali, D., Mohamed, J.: Multi-objective genetic algorithm based sizing optimization of a stand-alone wind/PV power supply system with enhanced battery/supercapacitor hybrid energy storage. Energy 163, 351–363 (2018).
16. Tegani, I., Aboubou, A., Becherif, M., Ayad, M.Y., Kraa, O., Bahri, M., Akhrif, O.: Optimal sizing study of hybrid wind/PV/diesel power generation unit using genetic algorithm. In: 4th International Conference on Power Engineering, Energy and Electrical Drives, Istanbul, Turkey, pp. 134–140. IEEE (2013). 17. Hadidian Moghaddam, M.J., Kalam, A., Nowdeh, S.A., Ahmadi, A., Babanezhad, M., Saha, S.: Optimal sizing and energy management of standalone hybrid photovoltaic/wind system based on hydrogen storage considering LOEE and LOLE reliability indices using flower pollination algorithm. Renew. Energy 135, 1412–1434 (2019).

Data Mining Model for Student Internship Placement Using Modified Case Based
Reasoning

Ferddie Quiroz Canlas(&)

1 Introduction

Internship also called practicum or on the job training enables students to acquire skills, which cannot be learned within the four walls of the classroom and to augment lab-oratory and practical activities provided by the course [1]. This endeavor poses potential benefits to both students seeking for internship placement and the organiza-tion providing the internship program. Many studies show that students who underwent internship programs have a high probability of finding employment after graduation and employers prefer to hire successful interns than those non-interns for much obvious reason for which cost of training is one of them [2]. Organizations face challenges in identifying qualified students for company immersion. These challenges were sum-marized into: (1) gathering and analysing data; (2) collating findings; and (3) synthe-sizing recommendations [3].

Combining CBR with custom SQL scripts makes it more flexible than static rules embedded in code.

The proposed model aims to implement a community based, open-source platform for Higher Education Institutions, Business Firms, and students. The model is cloud-based and can be accessed using web browsers and mobile phones.

262 F. Q. Canlas

a resume downloader and the data normalizer. The former uses a web crawler that downloads resumes from the internet thus, it does not require applicants to upload their profile on the company’s website and the later parses unstructured data produced by the Text Miner into a structured format suitable for database storage. The studies above did not give weight to the importance of the selection criteria.

Case Based Reasoning has four major processes: Retrieve; Reuse; Revise and Retain. These processes go into a cycle as the basis on how a system learns. Figure 1 shows the cycles [15].

Local Similarity. The attributes/features of the new case/problem are identified and compared to those existing problems stored in the knowledge base using the formula:

• For Discrete Values:

a is the new feature
b is the previous features; and range is the value of the
difference between the upper and lower boundary of the set:

Global Similarity. It is the build-up of all the local similarity functions to find sim-ilarities between cases and usually generalized using rank (the proposed model will use percentages). To determine the global similarity the formula to be used is:

The proposed implementation model is a cloud-based, collaborative architecture that provides access to students, higher education institutions and industries for both applying, placing and providing internship programs.

4.1 The Architecture

Users access the system via mobile and web applications connected to the cloud. Students can create and update their profile and account using their mobile phones. In addition, the mobile application provides search, apply and receive notification about the approval status of the internship placement application. Higher education institu-tions can input students’ academic information thru a web portal and establish com-munication with organizations providing internship programs for students. The platform provides industries/business organizations the ability to query baseline information about internship applications and to shortlist candidates based on a set of pre-specified criteria. The CBR system provides individual percentages of similarity of each potential candidate vis-a-vis the selection criteria.

The CBR knowledge base is stored in a MySQL database server, an open source software known for many features such as data security, on-demand scalability, high performance, round the clock uptime (using clusters) and comprehensive transactional support [19] best suited for the architecture. It works directly with PHP; a cross platform web scripting language has built-in database support and strong library [20] suitable for handling CBR processes and querying the database for similar cases and performing local and global similarity computations.

Fig. 3. CBR knowledge base

Table 1. Internship placement criteria

Querying the Knowledge Base. The CBR knowledge base in the proposed model is composed of various information collected from different students and higher education institutions. Using the web interface customized SQL statement is generated based on selection made, the knowledge base is queried and returns records shown in Table 2.

Data Mining Model for Student Internship Placement 267

Table 3. Local Similarities (LS) for course & gender attributes

Using the formula for continuous values (2), the sample LS for Age and GPA of Student A (Table 4):

CBR Computation for Global Similarity. Using the formula (3), Student A is taken as an example to elaborate the global similarity computation. The complete results for the rest of the cases is show in Table 5:

The proposed model can be pictured as a three-tier implementation that includes a client and server architecture allowing various devices to access the CBR system. The knowledge base resides in the MySQL server; while the Retrieval process is being performed, using PHP scripts and results are returned to the mobile application using JSON arrays.

Similarity results are displayed in percentage format allowing organizations to draw decisions using quantitative and empirical bases, and with certain degree of flexibility and not merely on static rule base approach.

Int. J. Eng. Res. Technol. 5(11), 468–471 (2016)
11. Dhende, S., Pashankar, A., Pawar, S., Salave, A., Salunkhe, S.: Candidate hiring through CV analysis. Int. Res. J. Eng. Technol. 5(5), 3148–3149 (2018)
12. Pombo, L.: Landing on the right job: a machine learning approach to match candidates with jobs applying semantic embeddings. Master’s thesis, Universidade Nova de Lisboa (2019) 13. Burkhard, H.D.: Case-Based Reasoning Technology from Foundations to Applications.

Springer, Heidelberg (1998)
14. Kolodner, J.L.: Case-Based Reasoning. Morgan Kaufmann, San Francisco (1993) 15. Aamodt, A., Plaza, E.: Case-based reasoning: foundational issues, methodological variations, and system approaches. AI Commun. 7(1), 39–59 (1993)
16. Chen, L., Ou, D., Yao, H.: Research on fault diagnosis of vehicle equipment for high-speed railway based on case-based reasoning. In: 3rd International Conference on Electrical and Information Technologies for Rail Transportation (EIRT) 2017, Changsha, China (2017) 17. Pal, S.K., Shiu, S.C.: Foundation of Soft-Case. Wiley, Hoboken (2004)
18. Von Wangenheim, C.G.: Case-Based Reasoning – A Short Introduction. Universidade do Vale do Itajaí (2000)
19. Branson T.: Datamation (2016).
20. Davis P.: Advantages and Disadvantages of PHP (2019).
21. Sharma R.: JSON – its advantages and disadvantages (2019).

Abstract. Nowadays, Electric Vehicles have gained a lot of interest among academic researches and the industrial actors, however, for a vast adoption of these tools, tasks such as their autonomy prolongation as well as ensuring their battery security are of great importance. Tasks that are accomplished via a survey of their battery state of charge (SOC) and state of health (SOH). In this paper, we present two advanced algorithms; Artificial Neural Network and Adaptive Gain Sliding Mode Observer (AGSMO), based on a combined battery equivalent circuit model (CBECM) to estimate the SOC of a lithium-ion battery for electric vehicles. To verify the effectiveness of each algorithm, PIL (Pro-cessor In the Loop) tests were implemented using an STM32F429ZI discovery board. The experimental results prove that both algorithms have good perfor-mance in battery SOC estimation, with a slight edge of AGSMO over the ANN due to the limitation of training data.

Keywords: Adaptive sliding mode observer � Electric vehicle � State of charge � Artificial neural network � Lithium-ion battery � Processor in the loop

arranged in three different categories were developed to estimate SOC [1]. The direct methods such as EletroMotive Force, Open Circuit Voltage and impedance measure-ment methods all have several limitations due to their dependence on some physical properties and parameters of the battery. The second is the indirect or book-keeping method which is based on coulometric measurement [2], making it an open-loop system that takes in consideration capacity loss, self-discharging and others internal parameters, despite its implementation simplicity it suffers from initial SOC estimation uncertainty and accumulation of measurement error caused by the integrative system. The last one is adaptive system methods like Artificial Neural Network (ANN) [3], fuzzy logic [4], Kalman filter [5] …etc., these approaches can self-adjust in case of any system change. Because of the non-linearity of the SOC and the chemical factors that affect the battery, these methods provide a good solution for SOC estimation.

In this paper, an artificial neural network and adaptive sliding mode observer are chosen as a result of their high accuracy, performance, and independence on the selected battery model. For battery modeling as a necessary step to approach the real battery, Thevenin electrical model was implemented in this paper. A comparative study of the two methods is presented based on a Processor-In-the Loop (PIL) testing on an STM32F429ZI discovery board.

272 Y. Mazzi et al.

battery of 2.6 Ah capacity is used to provide experimental data for the algorithm.

Fig. 3. Variation of Voc SOC ð Þ

Table 1. Internal parameters of the battery

Where k and m are determined in every 10% of SOC, so we can represent the time derivation in each 10%:

For Vt in (2), the observer can be designed as follows:

Where

The discontinuous function sgn eVt ð Þ is replaced with a continuous function eVt=ð eVt define the adjusting speed of the adaptive gain j þ dÞ to reduce chattering effects. d is defined as a small positive scalar. a . is set a positive constant with

ð9Þ

ð11Þ

Where eVt ¼ Vt �bVt, eVoc ¼ Voc SOC Þ � Voc�SOC d�¼ keSOC, eVp ¼ Vp �c Vp and

y ¼ f�Xn i¼0wizi þ b� ð12Þ

The topology was trained using several functioning scenarios of the EV, and is presented by the following Fig. 4:

Following the design of the model and the controller, the PIL blocs (green blocs) of the AGSMO and ANN controller blocs (red blocs) were generated automatically by using C-code as illustrated in Fig. 5 and Fig. 6. Simulation results are presented in Figs. 7 and 8.

276 Y. Mazzi et al.

Fig. 7. SOC estimation curves

5 Conclusion

In this paper, PIL implementations of AGSMO and ANN methods are realized using an STM32F429ZI discovery board in the goal to estimate SOC. From the results obtained, it can be deduced that both of strategies perform well the SOC estimation in terms of precision, the AGSMO is robust against the modeling errors with less chattering ripples but it suffers from high time convergence to the true SOC, Contrarily to ANN which is converge rapidly and truck very well the SOC with an error that takes the maximal value in instantaneous transition of current. besides, the ANN implementation is simpler than AGSMO.

Enhancing the Energy Performance of Passive Building Through the Internet of Things

El Mehdi El Khattabi1(&), Omar Diouri1, Mohamed Mharzi2, and Mohammed Ouazzani Jamil1

Buildings account for almost 40% of total energy consumption [1, 2]. Heating and cooling accounts for 85% of this figure, and electricity, especially lighting, for 15% [3]. The potential for energy optimization is enormous. Here, building automation [4–6] plays a decisive role, as does thermal insulation [7] and the use of energy-saving appliances [8]. The indoor climate is an essential factor for well-being, performance and health. One of the aims of room automation [6, 9] is therefore to create a pleasant indoor climate for the user and to ensure a comfortable use of the room. This dual objective should be achieved with the lowest possible energy consumption. When rooms are used for different purposes, optimizing their use according to circumstances helps to reduce energy costs, but also to protect the environment and conserve energy resources.

Our work aims at optimizing the conditions of summer/winter thermal comfort in the building by parametrically studying passive elements such as glazing, building orientation, ventilation and the thermo-physical properties of materials [10–12], taking into consideration the external conditions.

They can adapt the user interface to the user’s needs and preferences. This adap-tation is achieved by transferring parameters between different operating scenarios. It can be said that the use of SBC models to express and process user needs in an intelligent environment will lead to a fully adaptive space, thus meeting the IoT [13–15] objective of ensuring good control and management of energy and making daily life easier.

2 Case Studies

In order to properly monitor the thermal behavior of this building includes this innovative mechanism, we have installed an IoT system that ensures a positive energy gain in terms of consumption and achieved thermal comfort inside the building by relying on various factors such as Reviewing and optimizing the real needs of the building (type of activities, occupation schedule, temperature, humidity, etc.) and ensuring that the heating, ventilation and air conditioning systems meet these needs. This intelligent system can also allow for better integration of new energy sources, especially renewable (solar energy).

Enhancing the Energy Performance of Passive Building 281

A building is only considered intelligent when its HVAC, ventilation and lighting equipment, as well as other installations, are combined in a network. This networking is the key to energy efficiency and helps to optimize the operating costs of a building.

4 Description of an Automation System

Since the heart on its own can only function if there are sensors and actuators. For this reason, the need to add further expansion modules to the PLC is mandatory. The communication between the PLC and these modules is done through the existing protocol in the PLC which is the RS-485 protocol using the S-bus via port 1 of the PLC. These modules allow the acquisition of information from switches and push buttons (inputs for the PLC) to act on the lamps (outputs for the PLC).

Port 2 of the PLC is connected to the fans through the RS-485 protocol using Modbus. These fans are silent (no noise) and also they consume less energy. We have used two fans of this type, one is an extractor and the other is a vacuum cleaner. The particularity of this type of fans is that they contain a speed regulator, this regulator is of PID type which is used to stabilize the speed of the fan.

Enhancing the Energy Performance of Passive Building 283

Fig. 4. Communication between the PLC and the wireless switch

This study reflects its interest in controlling one of the most complex sectors in terms of energy control and management. The study showed how to reach the comfort zone in a very effective way, without using active systems that generate energy loss, but through a low-consumption ventilation system, which allows compensation for various losses in order to ensure a balance between building standards through a management system called the Internet of Things.

References

15. Singh, N., Kumar, S., Kanaujia, B.K.: HC Choi et KW Kim, «Energy-Efficient System Design for Internet of Things (IoT) Devices», pp. 49–74. Studies in Systems, Decis. Control (2019)
16. Presse Dualsun. (Communiqués Septembre 2018). L’offre heliopacsystem+® développée par DualSun et HELIOPAC obtient un Titre V Système Dynamique

Modelling and Optimization of Stirling Engine for Waste Heat Recovery from Cement Plant Based on Adiabatic Model
and Genetics Algorithms

The depletion of fossil resources and their negative impact on the environment is an obligation today for looking for other opportunities. Renewable energies are the main alternatives that can be used thanks to its sustainability and renewable process.

For several years, great effort has been devoted to the study of Stirling engines under its different types namely Alpha, Beta and Gamma and their renewable energy supplies especially solar energy and biomass. Kadri et al. [3], proposed a small scale power plant using a dish Stirling generator system suitable for electrification in Tunisian remote areas and they developed their system by adding a storage battery. Their results showed that solar dish Stirling-generator system reachs the objectives of power stability and system autonomy.

Damirchi et al. [4] manufactured and optimized a Gamma type Stirling engine with 580 cc total volume. They demonstrated the ability of Stirling engines to reach biomass gases. Their results showed that the thermal efficiency of the engine is equal to 16% and the maximum brake power was 96.7 W at 550 °C hot temperature, 700 rpm rotational speed and 10 bar pressure.

2.1 Description of the Stirling Cycle

The Stirling engine is an externally heat engine operating on the compression and expansion of a working gas such as nitrogen, helium, air or hydrogen at different temperatures [7]. Stirling engines can be classified into three major configurations that are distinguished by their mechanical construction [8].

– Process 4-1: cooling of working gas at constant volume and release of heat to the regenerator in order to reduce the amount of heat introduced by the external source.

290 K. Laazaar and N. Boutammachte

The mass equations of adiabatic model are calculated in each SE component:

The net work output is calculated by the next equation:

dW ¼ PðdVe þ dVcÞ ð9Þ

The physical and geometrical specifications used in the present study are sum-marized in Table 1.

Table 1. Specifications of the Stirling engine.

292 K. Laazaar and N. Boutammachte

Figure 3 presents the expansion and compression volumes variations (Pressure-Volume diagram) during an engine cycle at 1200 rpm rotational speed. From the graph, it is clear that the expansion work is larger than the compression work which will have

The effect of cooler temperature on heat input and efficiency of SE is displayed in Fig. 5. The power output and efficiency decreases with the increase of the cooler temperature and reaches a maximum value of 8.64 KW and 68.7% respectively at 280 K. This is theoretically explained by the fact that is must be a large temperature difference between cold and hot sources of Stirling device to maximize its output performance.

The effect of phase angle variation between the hot and cold cylinders is displayed in Fig. 7. An increase in work can be seen while increasing the phase angle up to 90 degrees. After this value, a diminution in total work is noted and varies from 402.85 J to 393.09 J. It is shown also a slight increase in efficiency due to the decrease in heat input, so the heat loss is minimized when a rise in phase angle is done.

294 K. Laazaar and N. Boutammachte

Figure 8 depicts the effect of gas mass on heat input and output power of SE. The augmentation of fluid mass increases significantly the output power because of the increase of charge pressure and then the increase of output work. However, a large augmentation of mass leads to an addition of heat amount in the SE heater achieving to 808.21 J at 1.261 g gas mass.

Modelling and Optimization of Stirling Engine for Waste Heat Recovery 295

• 80 <= Alpha <= 110°
• 1000 <= N <= 2500 rpm
• 0.0224 <= Zc <= 0.042 m
• 278 <= Tk <= 300 K
• 1.0123 <= M <= 1.2460 g

In the present paper, a thermodynamic modelling of an Alpha type Stirling engine was conducted using adiabatic model. The engine is powered by the heat dissipated from clinker cooling during the cement fabrication for electricity production purpose.

First, a sensitivity analysis was carried out using MATLAB software in order to investigate the effects of several operating conditions and geometrical parameters such as phase angle, rotational speed, piston stroke, cooler temperature and gas mass. The results showed that the output power reaches 23.52 KW while increasing the engine speed to 3000 rpm. By rising the stroke of SE from 0.02 m to 0.045 m, an

References

1. Ipci, D., Karabulut, H.: Thermodynamic and dynamic analysis of an alpha type Stirling engine and numerical treatment. Energy Convers. Manag. 169, 34–44 (2018)
2. Chen, W., Wong, K.: A numerical study on the effects of moving regenerator to the performance of a b-type Stirling engine. Int. J. Heat Mass Transf. 83, 499–508 (2015) 3. Kadri, Y., Hadj, H.: Performance evaluation of a stand-alone solar dish Stirling system for power generation suitable for off-grid rural electrification. Energy Convers. Manag. 129, 140–156 (2016)
4. Damirrchi, H., Najafi, G.: Micro combined heat and power to provide heat and electrical power using biomass and Gamma-type Stirling engine. Appl. Therm. Eng. 103, 1460–1469 (2016)
5. Cascella, F., Gaboury, S., Sorin, M.: Proof of concept to recover thermal wastes from aluminum electrolysis cells using Stirling engines. Energy Convers. Manag. 172, 497–506 (2018)
6. Ferreira, A., Nunes, M.: Thermodynamic and economic optimization of a solar-powered Stirling engine for micro-cogeneration purposes. Energy 111, 1–17 (2016)
7. Coria, M., Cobas, V.: Perspectives of Stirling engines use for distributed generation in Brazil. Energy Policy 34, 3402–3408 (2006)
8. Kongtragool, B., Wongwises, S.: A review of solar-powered Stirling engines and low temperature differential Stirling engines. Renew. Sustain. Energy Rev. 7, 131–154 (2003) 9. Ahmadi, M., Ahmadi, A., Pourfayaz, F.: Thermal models for analysis of performance of Stirling engine: a review. Renew. Sustain. Energy Rev. 68, 168–184 (2016)
10. Almajri, A., Mahmoud, S., Al-dadah, R.: Modelling and parametric study of an efficient Alpha type Stirling engine performance based on 3D CFD analysis. Energy Convers.

Abstract. Melanoma is the most dangerous form of skin cancer that often looks like moles. Dermatologists often recommend a regular examination of the skin to identify eliminate the melanoma in its early stages. To facilitate this process, we present an embedded skin cancer classification on a Raspberry Pi. This system provides a real-time classification of lesion taken by an embedded pi camera. The classification model used is deployed using a dataset issued from the ISIC2017 challenge. The model was first created on a computer and seri-alized to the raspberry pi. Features used are those used by dermatologist based on skin and lesion color and texture information. SVM is used as the classifi-cation algorithm. Experimentation results show the effectiveness of our pro-posed classification implementation.

Keywords: Skin cancer � Melanoma � Features extraction � Classification �SVM � Raspberry pi

© Springer Nature Switzerland AG 2021
T. Masrour et al. (Eds.): A2IA 2020, LNNS 144, pp. 297–306, 2021.

298 K. Ibrahim et al.

Fig. 1. Image database for Non-melanoma and Melanoma skin cancer

should extract the features from the lesion and from the skin. Many approaches used the features used by dermatologists in their clinical routine. The ABCD rules (Asymmetry, Bordure, Color and diameter). Some research used to extract features as the Gray Level Co-occurrence Matrix; Histogram of Oriented Gradient and Local Binary Pattern to extract textural features. Features using a skeleton representation of the lesion; that gives good result comparing with the literature is proposed in [5]. A features selection is an optimal steps; its role is to reduce the features vector by selecting the only the relevant ones to got a high accuracy rate in the classification step [6–8].

There are currently several machine learning methods for classifying skin cancer. Logistic Regression, Decision Trees, K-Nearest Neighbors and different kernel Support Vector Machine are the most used classifier in the case of skin cancer classification.

Fig. 2. System hardware block diagram

Fig. 3. Raspberry Pi board (Model B+).

3.3 Camera Interface

Fig. 4. Camera Raspberry Pi

3.4 Skin Lesion Detection and Classification

• Offline classification on the ISIC2017 challenge dataset:

1. Features extraction
2. Classify the features matrix using the SVM classifier 3. Serialize the classification model
4. Copying the model to the Raspberry Pi

The features used to classify skin cancer are color and texture that show from our previous work the effectiveness to classify skin cancer.

In pattern recognition, features are the measurable properties of an observed physical phenomenon. The extraction of discriminant features is a fundamental step in the recognition process, prior to classification.

• Gray Level Co-Occurrence Matrix (GLCM)

In our work, GLCM is used to better extract textural features from the images of skin lesions. It characterizes the spatial distribution within a limited surface and the gray levels.

Embedded Skin Cancer Detection and Classification on Raspberry Pi 303

The principle of the descriptor histogram of oriented gradients is that the position and shape of the local object within an image can be represented by the distribution of gradients of intensity or edge directions.

In this part, the given results of preprocessing, and classification are depicted and discussed.

A total of 2000 images were collected from the ISIC challenge in three classes of digital dermoscopy images used for experimentation:

In Table 1 we present a comparative study using 3 kernels that are linear, quadratic and Gaussian kernels of the support vector machine. We will use the Sensitivity (TP rate), Specificity (TN rate) and Accuracy measures to evaluate the classification results.

The quadratic kernels give the best accuracy compared with linear and Gaussian kernel with 84.00 accuracy rate (Fig. 6).

TP: number of true positives
TN: number of false positives
FP: number of true negatives
FN: number of false negatives

Table 1. Average result of different kernel of SVM

Table 2. The confusion matrix of the SVM of quadratic kernel

1_ load image: to load an image we already have implied on the raspi
2_capture: to take a picture directly from the raspi camera
3_ Classify: to analyze the image and provide the result, either melanoma or non-melanoma.

5 Conclusion

306 K. Ibrahim et al.

9. El Khoukhi, H., Filali, Y., Yahyaouy, A., Sabri, M.A., Aarab, A.: A hardware implementation of OTSU thresholding method for skin cancer image segmentation. In: 2019 International Conference on Wireless Technologies, Embedded and Intelligent Systems (WITS), pp. 1–5. IEEE, April 2019
10. Marot, J.: Raspberry Pi for image processing education, July 2020
11. Engineering, C., Shivkumar, H., Scientist, S.: Implementation of image processing on, vol.

Abstract. This research is a part of a work which addresses the issue of reconfiguration of production systems that aims to extend the assembly line life and to offer solutions to the introduction of a new product or the ability to change the existent one or the capacity regulation. Through this article, we propose a new reconfiguration process, which is defined by structured recon-figuration strategies divided to several operations. Each strategy follows specific steps in response to a reconfiguration issue. The process solves the issue by taking into account the impact limitation, which is to respond as locally as possible before considering cooperative or global solutions. The definition of these reconfiguration strategies allows us to implement a concatenation of steps and actions that characterize each point of decision-making. We propose then a model that defines each operation and the relation-ship between them and the production system entities, taking into account the various parameters of the system. We use the example of the automotive assembly lines to cover the analysis process.

Keywords: Reconfiguration process � Production system reconfiguration �Modelling process � Decision making � Assembly lines

Reconfiguration is the modification process of line architecture or operations/ resources allocation for addressing industrial issues like new product introduction on an existing assembly line [1].

The line configuration has to be changed according to existing resources and new performance objectives. Typically, the reconfiguration problem is similar to a first design process with more constraints resulting from the existing layout and resources.

In order to structure the approach, a modelling solution is proposed. Thus, the environment, the reconfiguration operations and strategies are synthesized in a meta-model that defines the entities of each part of the reconfiguration process. An industrial case illustrates the approach steps.

2 Reconfiguration Process

This approach plays a significant role in reducing the design time, which favors the parallel development of the products and the future lines. As shown in (Fig. 2) the design and development step of the following product starts before the termination phase of the previous one, taking into account that the line will still be existing. Thus, the time to launch the new product with a suitable line is considerably reduced.

As interpreted in (Fig. 3), when the introduction of a new vehicle occurs, the line can deal with this task using several solutions. The final configuration is the one that has been chosen after an evaluation, taking into consideration the initial constraints.

This line evolution brings a new identity to the production system. The new process may need new tasks or more preparation time or may even need a new unit with more operators that have a specific training (Fig. 3). Thus, the reconfiguration has a lot of constraints to deal with. Therefore, the reconfiguration method should take into account both line situations: the initial configuration and the intended one.

310 Z. Moukkar et al.

As usually in every process development, we deal with the reconfiguration problem using the following steps:
Analysis-Design-Selection-Implementation-Evaluation, as already discussed in a previous paper [5].

The Analysis step is considered as the data provider, it consists in defining the existing environment and selecting all the information which can help achieve the reconfiguration process.

– A method to select the overall characteristics of the first assembly line.

– A set of rules to compare the new product to the elder one already available on line.

Design Approach for Assembly Lines Reconfiguration Process 311

In our approach, we describe the environment entities in a top-down logic. In an automotive assembly context, the description can go from the site to the workstation.

Taking into account the existing constraints and the structure of each part (Fig. 5) and using the same classification as the robot’s relevance levels (Fig. 4), some units are basic and do not need modification, others are specific and they require at least a workstation update.

In this paper, we propose operations and strategies that arrange the previous con-figuration in a way to respond to the new configuration requirements.

For the Selection step, other methods proposed a reconfiguration assessment that gives the ability to add, delete or rearrange in a timely and cost effective time, com-ponents and functions. This process can provide a set of solutions [6, 7].

Repair planning approach is generally used to manage disruptions and minimize their impact on the initial schedule [13]; in the same way, the reconfiguration process manages the introduction of a new product/capacity and minimizes its impact on the initial configuration. If we consider that the disruption element can be the introduction of a new product/capacity, the only difference between the two processes is linked to the optimized parameter: it is “time” for the “repair planning” and “space and resources” for the “reconfiguration process”.

This approach leads us to define the reconfiguration operations in a context of optimizing space and resources, resulting from the analogy with the temporal context of the “planning recovery” (Table 1).

– Total surface of the workshop (TS);
– Ceiling height (CH);
– Number of units (UN);

Consequently, we can know the exact position and surface of each unit and its workstations. The data can be structured as coordinates which makes us define a unit in the space by five dimensional coordinates (Xi, Xf, Yi, Yf, S)
Where: (Xi, Xf) are the abscissa coordinates.

(Yi, Yf) are the ordinate coordinates. Figure 7 (S) is the surface that can eventually be calculated using the formula:

The unit can be described as the main entity that changes in response to the reconfiguration call. Based on the analogy with the “repair planning”, the unit is considered as a task in a spatial context. Therefore, we define it using several attributes, in order to have all the elements needed to express the constraints and define the reconfiguration process.

Table 2. Assembly unit attributes (type, details, data type).

The sum of all the unit surfaces cannot be bigger than the workshop total surface. This obvious definition can help us choose automatically the best solution in the case of adding space in order to fix a reconfiguration disorder. Other definitions are used to lead or limit the reconfiguration operations and strategies, for example:

Spatial Constraint Example
In a workshop that has a height H, every unit i 2 [1, UN] must have a height less than that.

4 Reconfiguration Process Definition

The reconfiguration process is characterized by strategies that contain operations. The strategy is the way the process deals with the reconfiguration issue. It is a sequence of actions and paths. Consequently, the operation is considered as the elementary entity of the process definition. It acts on every attribute of the assembly line environment by changing the initial parameters.

of operations, the local ones that affect only one unit and the global ones that can modify more than one unit. Both types can have the same results, it only depends on how many units are affected.

• Decomposition operation

This operation allows to break the unit into two or more sub-units to meet the need of reconfiguration. This includes a task decomposition (Example bellow).

Design Approach for Assembly Lines Reconfiguration Process 317

It can be a local or cooperative operation.

Local: The used surface already belongs to the unit (it does not exceed the original boundaries of the unit).

This operation moves the entire unit without affecting its surface. It is local if it does not affect any other unit and does respect the workshop limits. Otherwise it is considered as cooperative.

4.1.2 Capacity Operations
This type of operation changes the capacity supported by the unit, either increasing or decreasing it. The increase operation is limited by the maximum capacity of the unit (Table 2).

318 Z. Moukkar et al.

These operations act directly on the unit and do not involve others. Either by adding resources or removing them, it can also act as a replacement operation. They are local actions.

Initial configuration

4.2 Reconfiguration Strategies

Reconfiguration strategies are the background of the process, they reply to a recon-figuration trigger (for example: capacity modification, adding activity, new product introduction). They use a set of reconfiguration operations previously defined.

Fig. 9. Limitation of the impact of a disturbance: from local to global processing

320 Z. Moukkar et al.

If the unit can address the problem described in this message, the problem is solved by a local strategy. If not, a request message depending on the reconfiguration trigger (for example: need for extra space) is sent to the nearest units (the previous and the next ones on the assembly process). If they answer positively, global operations are per-formed and then the problem is solved by a cooperative strategy. If not, the request message is transferred to a higher level. This process lasts until the request is answered. The process stops in one condition, when all the units answer negatively. Then the conclusion is that the reconfiguration solution has to be performed outside the work-shop process (for instance: adding new resources).

5 Conclusion

Abbreviations Table

.

CIRP Ann. Manuf. Technol. 51(1), 9 (2002)
4. Pedrazzoli, P., Urbani, A., Molinari Tosatti, L.: Flexibility and reconfigurability: an analytical approach and some examples. In: CIRP 1st International Conference on Agile Reconfigurable Manufacturing (2001)
5. Feno, M.R., Cauvin, A., Ferrarini. A.: Verification of design rules as an evaluation method during the reconfiguration process of production systems (2015)
6. Farid, A.M.: Reconfigurability measurement in automated manufacturing systems (2007) 7. Karl, F., Reinhart, G.: Strategic planning of reconfigurations on manufacturing resources.

Procedia CIRP 3, 608–613 (2012)
8. Bryan, A., Ko, J.: Co-evolution of product families and assembly systems. CIRP Ann.

J. Prod. Res. 46, 967–992 (2008)
12. Feno, M.R., Cauvin, A., Ferrarini, A.: A production system reconfiguration model based on repair approach (2014)
13. Cauvin, A., Fournier, S: Disruption management in distributed organizations: a cooperative repair approach for reactive planning and scheduling. Prog. Econ. Res., 23 (2011). Nova Science Publishers
14. Feno, M.R., Cauvin, A., Ferrarini, A.: An approach for production system reconfiguration in the case of introduction of a new product (2014)
15. Cerezuela, C.: Contribution to method and tool development in order to help design a concurrent engineering perspective. The case of the electric wiring. Aix-Marseille University, Marseille (1996)
16. Feno, M.R.: Analyse et évaluation de la reconfiguration d’une ligne de production – application à l’assemblage tôlerie automobile, Ph.D. Ecole nationale supérieure d’arts et métiers, Aix-en-Provence, France (2016)

Author Index

B
Baba, Elhoussaine, 210
Barakat, Mahmoud, 247
Belhaouari, Samir Brahim, 110
Belkasmi, Mohammed Ghaouth, 175, 236 Ben Sassi, Hicham, 270
Benabdellah, Mohammed, 61
Bendriss, Yassin, 184
Benghabrit, Asmaa, 145
Beniysa, Mohsin, 47
Benlahbib, Abdessamad, 135
Berrich, Jamal, 236
Bouajaj, Adel, 47
Bouchentouf, Toumi, 175, 236
Boumhidi, Achraf, 135
Bourouhou, A., 78
Boutammachte, Noureddine, 287
Britel, Mohammed Réda, 47

C
Canlas, Ferddie Quiroz, 260
Cauvin, Aline, 307
Chevalier, Thibaud, 1
Chkouri, Mohamed Yassin, 158

H
Habib, Md. Tarek, 166 Hajji, Tarik, 61
Hamdaoui, Youssef, 184 Hammouch, Ahmed, 210

T. Masrour et al. (Eds.): A2IA 2020, LNNS 144, pp. 323–324, 2021.

J
Jamali, Abdellah, 99
Jamil, Mohammed Ouazzani, 279 Jbari, Atman, 200
Jilbab, Abdelilah, 78, 200, 210 Jolimaitre, Elsa, 1

K
Kaddari, Zakaria, 236
Kalim, Ruhullah Bin, 166 Khattabi, El Mehdi El, 279

L
Laazaar, Kaoutar, 287
Lahoussine, Hanane Ait, 91 Largo, S., 78
Lazaar, Nouhaila, 247

Sabor, Jalal, 220, 247
Sabri, Mouhcine, 158
Sabri, My Abdelouahed, 297 Senhaji, Saloua, 119
Smail, Zaki, 126

Moreaud, Maxime, 1