Journal of Automation, Mobile Robotics and Intelligent Systems, vol. 14, no. 3 (2020) by Ł-PIAP

VOLUME 14, N° 3, 2020 www.jamris.org

Indexed in SCOPUS

Journal of Automation, Mobile Robotics and Intelligent Systems pISSN 1897-8649 (PRINT) / eISSN 2080-2145

(ONLINE)

WWW.JAMRIS.ORG • pISSN 1897-8649 (PRINT) / eISSN 2080-2145 (ONLINE) • VOLUME 14, N° 3, 2020

logo podstawowe skrót

Publisher: ŁUKASIEWICZ Research Network – Industrial Research Institute for Automation and Measurements PIAP

logo podstawowe skrót

ŁUKASIEWICZ Research Network – Industrial Research Institute for Automation and Measurements PIAP

Journal of Automation, Mobile Robotics and Intelligent Systems A peer-reviewed quarterly focusing on new achievements in the following fields: • Fundamentals of automation and robotics • Applied automatics • Mobile robots control • Distributed systems • Navigation • Mechatronic systems in robotics • Sensors and actuators • Data transmission • Biomechatronics • Mobile computing Editor-in-Chief

Typesetting

Janusz Kacprzyk (Polish Academy of Sciences, Łukasiewicz-PIAP, Poland)

PanDawer, www.pandawer.pl

Advisory Board

Webmaster

Dimitar Filev (Research & Advenced Engineering, Ford Motor Company, USA) Piotr Ryszawa (Łukasiewicz-PIAP, Poland) Ministry of Science Kaoru Hirota (Japan Society for the Promotion of Science, Beijing Office) and Higher Education Editorial Office Witold Pedrycz (ECERF, University of Alberta, Canada) Republic of Poland ŁUKASIEWICZ Research Network Co-Editors – Industrial Research Institute for Automation and Measurements PIAP Roman Szewczyk (Łukasiewicz-PIAP, Warsaw University of Technology, Poland) Al. Jerozolimskie 202, 02-486 Warsaw, Poland (www.jamris.org) Oscar Castillo (Tijuana Institute of Technology, Mexico) tel. +48-22-8740109, e-mail: office@jamris.org Marek Zaremba (University of Quebec, Canada) Ministry The reference version of the journal is e-version. Printed in 100 copies.

Executive Editor

Ministry of Science ´ Katarzyna Rzeplinska-Rykała, e-mail: office@jamris.org Poland) and Higher(Łukasiewicz-PIAP, Education

Associate Editor

Republic of Poland

´ (Poznan ´ University of Technology, Poland) Piotr Skrzypczynski

of Science

Articles are reviewed, excluding advertisements and descriptions of products. and Higher

Education If in doubt about the proper edition of contributions, for copyright and reprint Republic of Poland permissions please contact the Executive Editor. P ublishing of “Journal of Automation, Mobile Robotics and Intelligent Systems” – the task Ministry of Science financed under contract 907/P-DUN/2019 and Higher Education from funds of the Ministry of Science and Republic of Poland Higher Education of the Republic of Poland allocated to science dissemination activities.

Statistical Editor ´ Małgorzata Kaliczynska (Łukasiewicz-PIAP, Poland)

Ministry of Science and Higher Education Republic of Poland

Editorial Board:

Mark Last (Ben-Gurion University, Israel) Anthony Maciejewski (Colorado State University, USA) Krzysztof Malinowski (Warsaw University of Technology, Poland) Andrzej Masłowski (Warsaw University of Technology, Poland) Patricia Melin (Tijuana Institute of Technology, Mexico) Fazel Naghdy (University of Wollongong, Australia) Zbigniew Nahorski (Polish Academy of Sciences, Poland) Nadia Nedjah (State University of Rio de Janeiro, Brazil) Dmitry A. Novikov (Institute of Control Sciences, Russian Academy of Sciences, Russia) Duc Truong Pham (Birmingham University, UK) Lech Polkowski (University of Warmia and Mazury, Poland) Alain Pruski (University of Metz, France) Rita Ribeiro (UNINOVA, Instituto de Desenvolvimento de Novas Tecnologias, Portugal) Imre Rudas (Óbuda University, Hungary) Leszek Rutkowski (Czestochowa University of Technology, Poland) Alessandro Saffiotti (Örebro University, Sweden) Klaus Schilling (Julius-Maximilians-University Wuerzburg, Germany) Vassil Sgurev (Bulgarian Academy of Sciences, Department of Intelligent Systems, Bulgaria) Helena Szczerbicka (Leibniz Universität, Germany) Ryszard Tadeusiewicz (AGH University of Science and Technology, Poland) Stanisław Tarasiewicz (University of Laval, Canada) Piotr Tatjewski (Warsaw University of Technology, Poland) Rene Wamkeue (University of Quebec, Canada) Janusz Zalewski (Florida Gulf Coast University, USA) ´ Teresa Zielinska (Warsaw University of Technology, Poland)

Chairman – Janusz Kacprzyk (Polish Academy of Sciences, Łukasiewicz-PIAP, Poland) Plamen Angelov (Lancaster University, UK) Adam Borkowski (Polish Academy of Sciences, Poland) Wolfgang Borutzky (Fachhochschule Bonn-Rhein-Sieg, Germany) Bice Cavallo (University of Naples Federico II, Italy) Chin Chen Chang (Feng Chia University, Taiwan) Jorge Manuel Miranda Dias (University of Coimbra, Portugal) Andries Engelbrecht (University of Pretoria, Republic of South Africa) Pablo Estévez (University of Chile) Bogdan Gabrys (Bournemouth University, UK) Fernando Gomide (University of Campinas, Brazil) Aboul Ella Hassanien (Cairo University, Egypt) Joachim Hertzberg (Osnabrück University, Germany) Evangelos V. Hristoforou (National Technical University of Athens, Greece) Ryszard Jachowicz (Warsaw University of Technology, Poland) Tadeusz Kaczorek (Białystok University of Technology, Poland) Nikola Kasabov (Auckland University of Technology, New Zealand) ´ Marian P. Kazmierkowski (Warsaw University of Technology, Poland) Laszlo T. Kóczy (Szechenyi Istvan University, Gyor and Budapest University of Technology and Economics, Hungary) Józef Korbicz (University of Zielona Góra, Poland) ´ University of Technology, Poland) Krzysztof Kozłowski (Poznan Eckart Kramer (Fachhochschule Eberswalde, Germany) Rudolf Kruse (Otto-von-Guericke-Universität, Germany) Ching-Teng Lin (National Chiao-Tung University, Taiwan) Piotr Kulczycki (AGH University of Science and Technology, Poland) Andrew Kusiak (University of Iowa, USA)

Publisher: ŁUKASIEWICZ Research Network – Industrial Research Institute for Automation and Measurements PIAP

Articles

Journal of Automation, Mobile Robotics and Intelligent Systems Volume 14, N° 3, 2020 DOI: 10.14313/JAMRIS/3-2020

Contents 48

Synthesis of an Intelligent UAV Control System Based on Fuzzy Logic in External Disturbance Conditions Igor Korobiichuk, Dmytro Shevchuk, Iryna Prokhorenko, Nataliia Tymoshenko, Yaroslav Smityuh, Regina Boyko DOI: 10.14313/JAMRIS/3-2020/26

Preface to Special Issue on Modern Intelligent Systems Concepts Abdellah Idrissi DOI: 10.14313/JAMRIS/3-2020/33

The More You See Me the More You Like Me. Influencing the Negative Attitude Towards Interactions With Robots Paweł Łupkowski, Filip Jański-Mały DOI: 10.14313/JAMRIS/3-2020/27

Comparative Study of CNN and LSTM for Opinion Mining in Long Text Siham Yousfi, Maryem Rhanoui, Mounia Mikram DOI: 10.14313/JAMRIS/3-2020/34

Preface to Special Issue on Recent Advances in Information Technology III Piotr A. Kowalski, Szymon Łukasik DOI: 10.14313/JAMRIS/3-2020/28 20

Modelling and Control of Discrete-Event Systems Using Petri Nets and Arduino Microcontrollers Erik Kučera, Oto Haffner, Roman Leskovský DOI: 10.14313/JAMRIS/3-2020/29 28

Two Cascaded and Extended Kalman Filters Combined with Sliding Mode Control for Sustainable Management of Marine Fish Stocks Katharina Benz, Claus Rech, Paolo Mercorelli, Oleg Sergiyenko DOI: 10.14313/JAMRIS/3-2020/30 36

Proposal of IoT Devices Control using Mixed Reality and QR Codes Erich Stark, Erik Kučera, Oto Haffner DOI: 10.14313/JAMRIS/3-2020/31 42

Online Control Education Using 3D Holographic Visualisation Jakub Matišák, Matej Rábek, Katarína Žáková DOI: 10.14313/JAMRIS/3-2020/32 2

Articles

Fuzzy Multi Agent System for Automatic Classification and Negotiation of QOS in Cloud Computing Zineb Bakraouy, Wissam Abbass, Amine Baina, Mostafa Bellafkih DOI: 10.14313/JAMRIS/3-2020/35 65

Modeling of a Dynamic and Intelligent Simulator at the Infrastructure Level of Cloud Services Faouzia Zegrari, Abdellah Idrissi DOI: 10.14313/JAMRIS/3-2020/36 71

Unit Load Devices (ULD) Demand Forecasting in the Air Cargo for Optimal Cost Management Mounia Mikram, Maryem Rhanoui, Siham Yousfi, Houda Briwa DOI: 10.14313/JAMRIS/3-2020/37

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

Synthesis of an Intelligent UAV Control System Based on Fuzzy Logic in External Disturbance Conditions Submitted: 20th January 2020; accepted: 31st July 2020

Igor Korobiichuk, Dmytro Shevchuk, Iryna Prokhorenko, Nataliia Tymoshenko, Yaroslav Smityuh, Regina Boyko DOI: 10.14313/JAMRIS/3-2020/26 Abstract: To ensure reliable execution of flight tasks in the presence of both external perturbations and internal parametric perturbations, deterioration of the characteristics of the sensors, a control system structure based on intelligent technologies is proposed. The process of forming a “knowledge base” of a fuzzy controller is considered. The results of mathematical modeling of the longitudinal UAV control channel with a PID-controller and a fuzzy controller in the control loop are presented. Keywords: Unmanned Aerial Vehicle (UAV), Fuzzy controller, Longitudinal control channel, Height stabilization, Automatic control system, Pilot-navigation complex

1. Introduction The modern development of society requires the use of UAVs to solve a wide range of problems of varying complexity, which requires the improvement and creation of new UAVs, as well as their pilot-navigation complexes with the use of intellectual technologies. UAVs are extremely relevant today, trusting the investigation of large detainees found with the safety and life of people [1]. Unmanned aerial vehicles do not require any pilot on board and can be operated autonomously or remotely operated by the pilot [2, 3]. For today information technologies have radically changed the concept of UAVs and expanded their military and civilian use [4, 5]. Drones are indispensable in solving of a number of civilian tasks, such as: finding, detecting and identifying objects; disaster monitoring and control [6]; monitoring of oil and gas pipelines; fire detection [7]; search and rescue [8]; observing of public events [9]; observation of land and sea traffic [10]; ecological control and monitoring of plant growing [11]; terrestrial display and photographing [12]; meteorological observation; transportation of cargo; aero-photography; traffic monitoring and control [13, 14]. UAVs are indispensable during military missions. Thus, for the promising UAVs, it is possible to outline such basic tasks as: reconnaissance of above-ground,

air and naval targets, terrain exploration; radiation, chemical and biological investigation; installation of radio interference; fire management and targeting of ground, air and marine firearms; evaluation of the results of blows on the enemy [15, 16]. It should be mentioned that the development of UAVs for various purposes requires a number of topical tasks, including the development of an onboard control system and flight stabilization in the conditions of external disturbances [17, 18], as well as onboard high-precision navigation system. For example, a UAV flight control system must satisfy a number of conflicting requirements: reliability, simplicity of design, low cost, light weight and power consumption of actuators on the one hand, on the other hand: the accuracy of flight control in the conditions of external UAV perturbations. A compromise between different options can be achieved by use in the production of onboard UAV control system of modern intelligent control methods [19, 20]: artificial neural [21]; fuzzy logic [22]; genetic algorithms. As follows, the relevance of the above studies is to develop an onboard UAV control system based on modern intellectual technologies, which will improve the quality and accuracy of stabilization of its motion parameters in the conditions of external disturbances.

2. Materials and Methods of Research In Fig. 1. Typical UAVs are presented which were developed at the National Aviation University (Ukraine, Kyiv) [23]. Those UAV’s are physically constructions [23]. As a typical UAV is considered M-7D “Heavenly patrol” (Fig. 2d). This UAV is a twin-engine aircraft of the normal scheme with a high wing. There is an opening under the gondola for mounting the bottom / front camera. Main UAV technical characteristics: starting mass, kg – up to 150; payload weight, kg – up to 50; top speed, km / h. – 190; maximum flight altitude, m – up to 5000; maximum flight duration, h. – up to 10; method of start and landing – by plane; control modes – automatic / semi-automatic. UAVs can be used for patrolling linear objects, mapping and aerial photography, real-time video surveillance, etc. [23]. The numerical indicators for the control system with actuators of the M-7D “Heavenly patrol”[23] is 2 kg.

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

V SPS

ADC

SNS

b) ϑ ϑ ψ ψ

Fig. 1. Typical UAVs which were developed in NAU: a) M-6-3 “Zhaivir”; b) M-7V5 “Heavenly patrol”; c) BTS M-10 “Eye 2”; d) M-7D “Heavenly patrol”

Structural and functional diagram of the pilot-navigation complex M-7D “Heavenly patrol” with the use of intellectual technologies is presented in the following form in Fig. 2. The UAV pilot-navigation complex is structured as three modules of main units (trajectory control system, intelligent flight control system and navigation system), which are connected by cables. The UAV navigation system consists of a satellite navigation system (SNS), which operates offline and is designed to determine the location and speed of UAV flight. SNS receiver – is a combined module of GPS receiver and antenna. Performs UAV coordinates and transmits current information to the navigation system. The satellite navigation system provides information in a geocentric coordinate system linked to the center of the Articles

R ,V

System trajectory control

GIK

Trajectory control system

FPS

2020

Earth. But used system is a topographic coordinate system, so it is necessary to translate the information into a topographic system. The air-data computer (ADC) includes static pressure sensor (SPS) and full pressure sensor (FPS) and an ambient temperature sensor (ATS). The trajectory of the control system consists of a coordinate converter (R – UAV’s position vector; V – airspeed), a trajectory control system and a flight path setter that stores points of space (altitude, latitude, and longitude in the geographical coordinate system). Navigation system

N° 3

Fuzzy logic controller actuator (longitudinal δer canal) Fuzzy logic controller (lateral channel)

δa

actuator

Trajectory generator

Intelligent flight control system

Fig. 1. Structure of the UAV M-7D Heavenly Patrol PilotNavigation Complex Using Intelligent Technologies, where: FPS – full pressure sensor; SPS – static pressure sensor; TS – temperature sensor; ADC – air-data computer; GIC – gyro-induction compass; SNS – satellite navigation system; СC – coordinate converter (R – plane position vector; V – airspeed);intelligent flight control system; height H; СS – command signal ψ, θ – yaw angle and pitch angle The intelligent flight control system consists of two main modules: fuzzy UAV longitudinal motion controller and fuzzy UAV lateral motion controller. Fuzzy controllers take data from the navigation system and produce control effects based on bases of laws of management in the form “IF (flight situation) THEN (required control effect)”, and then submit them in the form of signals to the UAVs (steering heights and ailerons). So, the longitudinal control channel is presented as: height H, pitch angle ϑ and . its derivative θ, cs – a command signal that switches the mode circuit. For the longitudinal control channel, these are modes such as dial mode and height stabilization. In the presented UAV model, the control plane – elevation rudder δer (Fig. 2). Side angle . feed: angle of lightning ψ and its derivative ψ; cs – a command signal that switches the mode circuit. For the lateral control channel, these are modes such as reversal mode, setpoint mode, setpoint stabilization mode. In the presented UAV model, the control

Journal of Automation, Mobile Robotics and Intelligent Systems

plane – ailerons δa (Fig. 2). In order to ensure the competitiveness and efficiency of the use of UAVs, it is necessary for them to have a low cost and large weight of useful weight.For the purpose to provide these requirements a minimum number of low-cost sensors are set on board of the UAV, which means that the available measurements contain noise, which makes the synthesis of an efficient UAV motion control system much more difficult, and the use of standard control laws becomes impossible or inefficient. Taking into account all these features, as well as to avoid the use of slow-acting expensive adaptive systems, there is a need to synthesize control systems based on intelligent technologies that allow maintaining the controllability and stability of UAVs in the conditions of external disturbances. Consider forming a “base of laws” of control in a fuzzy UAV controller. Let’s represent UAV as an object of control with mn inputs and one output which corresponds to the controlling influence on the executive bodies: p = f (m1, m2,...mn), (1) where: f – output variable (angle of deflection of rudder of height, direction, ailerons, etc.); m1, m2,...mn – input variables (speed, height, angles of attack, sliding, pitch, roll, yaw, it derivatives etc). Variables m1, m2,...mn and p are quantitative, so the known limits of their change are assumed:

= Mi [m = i min , mi max ], i 1, n; (2) P=[ pmin , pmax ], (3) where: mi min (mi max) – the minimum (maximum) value of the input variable mi , i = 1, n ; P = [ pmin , pmax ] – the minimum (maximum) value of the output variable p. The task is to vector M* = [m1*,m2*,...,mn*] of fixed values of input variables mi* ∈ Mi , i = 1, n UAV to determine the required action of the automatic control system with a fuzzy controller p*∈P, that is, a clear value of the regulated value is submitted to the UAV executive bodies. A necessary condition for the formal solution of such problem is the presence of dependency (1). For example, to determine this dependency in a longitudinal channel UAV control, we’ll consider the input variables xi ∈ X= [V , α , H , ϑ , ω z ]T ,=i 1,6 , where V – flight speed, α – angle of attack, H – flight altitude, θ – pitch angle, ωZ – the angular velocity of change of the pitch angle and the output variable p = δer, where δer – rudder as linguistic variables given by universal sets (2) (3). To evaluate linguistic variables mi, i = 1, n і p we’ll l use fuzzy sets: Si = [ si1 , si2 ,..., si ] – fuzzy variable set mi, i = 1, n ; V = [v1,v2,...vr] – fuzzy variable set p, where siq − q – is a fuzzy set of linguistic variables mi, = q 1,= li , i 1, n ; vj – j – is a fuzzy set of linguistic variables p; r – number of different solutions in this field. In general case l1 ≠ l2 ≠ ... ≠ ln . l Names of individual terms si1 , si2 ,..., si si1,si2,...,sili may also differ from each other for different linguistic q variables mi, i = 1, n . Linguistic sets si ∈ Si і vj∈V i

VOLUME 14,

N° 3

2020

= q 1,= li , i 1,= n, j 1, r we’ll consider fuzzy sets given by

universal sets Mi and P, defined by the ratios (2) (3). Fuzzy sets siq and vj we’ll define the ratios: = siq

v= j

mi max

∫

mi min

µ si (mi )/ mi ;

pmax

∫

pmin

(4)

µ j ( p)/ p , (5)

where: siq (mi ) – function of belonging of the input variable mi∈[mi min,mi max] to the set siq ∈ Si , q = 1, li , i = 1, n ; v µ ( y ) – function of belonging of the output variable p∈[pmin,pmax]– to the solution vj ∈V, j = 1, r . The available expert data, presented in the form of a knowledge matrix, establishes a connection between the set of input parameters which characterize the current state of the UAV m1–mn and appropriate management influences vj j = 1, r to UAV’s executive bodies, in the form of logical statements of the type “IF (flight situation) THEN (control influence required)” in the following form: j

n

   (m=i

q 1= = i 1

 sijq ) → p= v j , = j 1, r . 

(6)

where: ∪ (or), ∩ (and); v j ( j = 1, r ) – linguistic evaluation of the output variable p, determined from a fuzzy set P; sijq – linguistic evaluation of the input variable mi in q-th row of the j-th disjunction selected from the corresponding fuzzy set S= 1,= n, j 1,= r , q 1, k j ; kj – i, i the number of rules that determine the value of the output variable of the controller on the UAV actuators. In conclusion, to resolve the issue of formalizing the process of making a managerial existence based on fuzzy logic, note the following:first, the representation of the input parameters (altitude and flight speed, pitch angle, etc) of the fuzzy UAV controller in the form of linguistic variables with fuzzy sets (negative large, negative small, zero, positive large, and so on) allows to describe the cause-effect relationships “input parameters – control effects” in natural language using fuzzy logical statements.

3. The Results of Research Analysis of the dynamic characteristics of UAV, allows us to formulate the basic requirements for flight control system (FCS) in the following form: – FCS should provide an acceptable quality of transients under the influence of deterministic control signals, as well as stochastic disturbances of the environment (horizontal and vertical gusts of wind).For example, for a typical UAV (Fig. 1.d), the time for the establishment of a transient process by a command to increase (decrease) the height by 50 m should be no more than 10 seconds, overshoot%, the oscillations should be damped in no more than 1 period. Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

where: V – actual airspeed (AAS); α – angle of attack, θ – pitch angle; ωZ – relative angular velocity; h – UAV flight altitude. The control vector in longitudinal motion this is the deflection of the steering wheel height, namely:

Fig. 3. Structural scheme of modeling of longitudinal motion of UAV in MATLAB environment, where: block 1 – UAV drive model; block 2 – UAV longitudinal motion model represented in state space; block 3 – UAV autopilot model – to ensure the necessary level of adaptability of FSC, these requirements must be maintained for parametric disturbed UAV motion models. In addition, it is also necessary to further reduce the region of uncertainty associated with the influence of the Reynold’s number on the stability of aerodynamic parameters. Thus, FSC must satisfy such conflicting requirements as minimizing the UAV motion stabilization error along given flight path under the action of deterministic and stochastic external disturbances and ensuring the adaptability of the system with respect to internal parametric disturbances or in the presence of uncertainty in the parameters of the UAV mathematical model. All this can be achieved by using an autopilot based on intelligent technologies in the UAV control loop. Let us assume, that investigated UAV, will be described in the space of states of the following equation:

x = Ax + Bu + w; (7) y Cx + v; =

where: x – state vector; A, B – state and control matrix; u – vector of control effects; y –observation vector; C – observation matrix; w, v – state noises and observations. The vector of measured coordinates in longitudinal motion has the following form: y=[h,θ,ωZ]T.

(8)

x=[V,α,θ,ωZ,h]T,

(9)

u=[δer]T (10)

where: u – control vector; δer – elevation rudder. In Fig. 3 is a structural diagram of UAV motion simulation in a longitudinal control channel. For example, let’s depict the fuzzy logic condoler (longitudinal canal) with two input variables (x1,x2). For the variable x1 7 terms are used, for the variable x2 – 7 terms, for the variable x3 – 7 terms (Fig. 4.). An example of such a network can be a system consisting of the following types of rules: Example1: IF x1 is NB AND x2 is NB THEN y is NB, Example 2: IF x1 is NB AND x2 is NM THEN y is NM, Example 3: IF x1 is NM AND x2 is NS THEN y is NM,

where: x1, x2, – input variables, y – outputvariable, NB (negative big), NS (negative medium), NS (negative small) certain sets with membership functions of the triangle type. Analyzing the main architecture of the fuzzy logic controller its generalized version can be seen (Fig. 4.).

Fig. 4. Generalized architecture of a neuro-fuzzy system with a logical conclusion subsystems Fragment of the database of a fuzzy autopilot (block 3 of Fig. 3) in the form of “IF...THEN” of the control rules connecting two input variables “height mismatch error” and “rate of change of error mismatch in height” with the output variable “angle of rotation of steering wheel height” is presented in Fig. 5.

The vector of the UAV state in longitudinal motion in the state space is represented in the following form:

Articles

Fig. 5. Display of a fragment of the “knowledge base” of the fuzzy controller in the form of “IF ... THEN” control rules

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

The UAV control surface is shown in Fig. 6 which connects two input parameters (height mismatch error and rate of change of error mismatch in height) with the output parameter (angle of rotation of steering wheel height).

a) 4 1

de (deg)

0 2

-2 -4

Fig. 6. UAV control surface connecting input and output variable

-6 -8

The process of processing the “database” of the fuzzy controller and the formation of the control action on the elevator is shown in Fig. 7.

t (s)

b) 4 1

de (deg)

0 2

-2 -4 -6 -8

t (s)

theta (deg)

15 10

Fig. 7. The process of processing the “database” of the fuzzy controller

0 -5

In Fig. 8 the results of mathematical modeling of UAV longitudinal motion control is presented.In the simulation process the PID-controller is compared with the control law u=k1+k2/p+k3p, where k1=0,12, k2=2,1, k3=3,45 – gain factors, as well as autopilot with a base of “control laws” (Fig. 5).

t (s)

d) Fig. 8. Transient management in longitudinal motionwhere:1 – PID-controller, 2 – fuzzy controller: a) by height; b) by the angular velocity of the pitch; c) by an elevation steering angle; d) by the pitch angle

Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

A comparative analysis of the results shows that the control loop of a fuzzy controller with a base of “control laws” is used it provides an increase in the of the speed of stabilization of altitude in comparison with the PID-controller under the conditions of wind gust 10 m/s by 15%, and also reduces energy costs on the steering wheel deviation by 18%. Controllers Schneider M340 or Schneider M540 and Matlab software can be used for implementing this intelligent control system based on fuzzy logic. The results mathematical modelling can not be applied to other classes of UAVs, because in mathematical model used specific values for UAV M-7D Heavenly Patrol

4. Conclusion For the UAV class which is considered, an important task is reducing of the number of sensors which is used if to take into account the reduce of cost of production. By reducing the number of sensors, it is possible to reduce the cost of both the control system and the UAV in general, which in turn makes the synthesized control system based on intelligent technologies accessible to a wider range of users with the desired characteristics of transients. Also, it should be noted that in the conventional methods of UAV control under uncertainty, the generalized control rule is set by a single law, and the fuzzy control uses a large number of partial laws. Each law operates in a given area of information space, which reflects an aerodynamic properties of this type of UAV, as well as the change of the external environment, which allows to provide the properties of adaptation to deterministic and stochastic external perturbations.

AUTHORS Igor Korobiichuk* – ŁUKASIEWICZ Research Network – Industrial Research Institute for Automation and Measurements PIAP, Jerozolimskie 202, 02-486 Warsaw, Poland, e-mail: igor.korobiichuk@piap.lukasiewicz.gov.pl. Dmytro Shevchuk – National Aviation University, Lyubomyr Huzar Avenue, 1, 03058, Kyiv, Ukraine, e-mail: dmitroshevchuk@gmail.com. Iryna Prokhorenko – National Aviation University, Lyubomyr Huzar Avenue, 1, 03058, Kyiv, Ukraine, e-mail: i.prokhorenko@nau.edu.ua.

Nataliia Tymoshenko – National Aviation University, Lyubomyr Huzar Avenue, 1, 03058, Kyiv, Ukraine, e-mail: n.tymoshenko@nau.edu.ua.

Yaroslav Smityuh – Ukrainian State University of Food Technologies, 68 Volodymyrska Street, 01033, Kyiv, Ukraine, e-mail: Smityuh1@gmail.com. Articles

VOLUME 14,

N° 3

2020

Regina Boyko – Ukrainian State University of Food Technologies, 68 Volodymyrska Street, 01033, Kyiv, Ukraine, e-mail: rela@ukr.net. *Corresponding author

REFERENCES [1] R. Austin, Unmanned Aircraft Systems: UAV Design, Development and Deployment, WileyBlackwell, 2010. [2] J. Bishop, “The Role of Affective Computing for Improving Situation Awareness in Unmanned Aerial Vehicle Operations: A US Perspective”. In: J. Vallverdú (eds.), Handbook of Research on Synthesizing Human Emotion in Intelligent Systems and Robotics, 2015, 10.4018/978-1-4666-7278-9.ch020. [3] “Cir 328 AN/190 Unmanned Aircraft Systems (UAS)”. International Civil Aviation Organization, https://www.icao.int/Meetings/UAS/Docu ments/Circular%20328_en.pdf. Accessed on: 2020-12-10. [4] I. Korobiichuk, Y. Danik, O. Samchyshyn, S. Dupelich and M. Kachniarz, “The estimation algorithm of operative capabilities of complex countermeasures to resist UAVs”, SIMULATION, vol. 95, no. 6, 2019, 569–573, 10.1177/0037549718791264. [5] I. Korobiichuk, M. Nowicki, Y. G. Danik, S. Dupelich and S. Oleksyj, “The Selection Methods for Multisensor System Elements of Drone Detection”. In: R. Szewczyk and M. Kaliczyńska (eds.), Recent Advances in Systems, Control and Information Technology, 2017, 20–26, 10.1007/978-3-319-48923-0_3. [6] H. Choi, M. Geeves, B. Alsalam and F. Gonzalez, “Open source computer-vision based guidance system for UAVs on-board decision making”. In: 2016 IEEE Aerospace Conference, 2016, 10.1109/AERO.2016.7500600. [7] M. Chen, Q. Hu, C. Mackin, J. F. Fisac and C. J. Tomlin, “Safe platooning of unmanned aerial vehicles via reachability”. In: 2015 54th IEEE Conference on Decision and Control (CDC), 2015, 4695–4701, 10.1109/CDC.2015.7402951. [8] N. Rupasinghe, A. S. Ibrahim and I. Guvenc, “Optimum Hovering Locations with Angular Domain User Separation for Cooperative UAV Networks”. In: 2016 IEEE Global Communications Conference (GLOBECOM), 2016, 10.1109/GLOCOM.2016.7842113. [9] M. Naphade, G. Banavar, C. Harrison, J. Paraszczak and R. Morris, “Smarter Cities and Their Innovation Challenges”, Computer, vol. 44, no. 6, 2011, 32–39, 10.1109/MC.2011.187. [10] T. J. Zajkowski, “Unmanned aerial vehicles: Remote sensing technology for the USDA For-

Journal of Automation, Mobile Robotics and Intelligent Systems

est Service”, Project Report RSAC-1507-RPT1, Remote Sens. Application Center, Salt Lake City, Utah, 2003. [11] C. Brodbeck, E. Sikora, D. Delaney, G. Pate and J. Johnson, “Using Unmanned Aircraft Systems for Early Detection of Soybean Diseases”, Advances in Animal Biosciences, vol. 8, no. 2, 2017, 802–806, 10.1017/S2040470017001315. [12] U. E. Franke, “Civilian Drones: Fixing an Image Problem?”, https://isnblog.ethz.ch/security/ civilian-drones-fixing-an-image-problem. Accessed on: 2020-12-10. [13] M. Erdelj and E. Natalizio, “UAV-assisted disaster management: Applications and open issues”. In: 2016 International Conference on Computing, Networking and Communications (ICNC), 2016, 10.1109/ICCNC.2016.7440563. [14] A. C. Watts, V. G. Ambrosia and E. A. Hinkley, “Unmanned Aircraft Systems in Remote Sensing and Scientific Research: Classification and Considerations of Use”, Remote Sensing, vol. 4, no. 6, 2012, 1671–1692, 10.3390/rs4061671. [15] S. G. Gupta, M. M. Ghonge and P. M. Jawandhiya, “Review of Unmanned Aircraft System (UAS)”, International Journal of Advanced Research in Computer Engineering & Technology (IJARCET), vol. 2, no. 4, 2013, 1646–1658. [16] M. Hassanalian and A. Abdelkefi, “Classifications, applications, and design challenges of drones: A review”, Progress in Aerospace Sciences, vol. 91, 2017, 99–131, 10.1016/j.paerosci.2017.04.003. [17] I. Korobiichuk, V. Karachun and V. Mel’nick, “Stochastic Structure of Inciting Factors of Trivial Gyrostabilized Platform”. In: R. Szewczyk, J. Krejsa, M. Nowicki and A. OstaszewskaLiżewska (eds.), Mechatronics 2019: Recent Advances Towards Industry 4.0, 2020, 36–44, 10.1007/978-3-030-29993-4_5. [18] I. Korobiichuk, “Experimental Investigations of a Precision Sensor for an Automatic Weapons Stabilizer System”, Sensors, vol. 17, no. 1, 2017, 10.3390/s17010023. [19] T. A. Johansen, “Stability, robustness, and performance of fuzzy model based control”. In: Proceedings of 35th IEEE Conference on Decision and Control, 1996, 604–609, 10.1109/CDC.1996.574390. [20] K. M. Passino and S. Yurkovich, Fuzzy Control, Addison-Wesley, 1997. [21] V. Tregub, I. Korobiichuk, O. Klymenko, A. Byrchenko and K. Rzeplińska-Rykała, “Neural Network Control Systems for Objects of Periodic Action with Non-linear Time Programs”. In: R. Szewczyk, C. Zieliński and M. Kaliczyńska (eds.), Automation 2019, 2020, 155–164, 10.1007/978-3-030-13273-6_16. [22] I. Korobiichuk, V. Tregub, O. Klymenko, I. Elperin, V. Sidletskyi, Y. Smityuh and M. Chorno-

VOLUME 14,

N° 3

2020

van, “Development of Logical Control System for the Purification Department at Molasses Production”. In: R. Szewczyk, J. Krejsa, M. Nowicki and A. Ostaszewska-Liżewska (eds.), Mechatronics 2019: Recent Advances Towards Industry 4.0, 2020, 206–213, 10.1007/978-3-030-29993-4_26. [23] “Наукові розробки (in Ukrainian)”. National Aviation University, https://nau.edu.ua/ua/ menu/science/naukovi-rozrobki/. Accessed on: 2020-12-10.

Articles

Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME N°33 2020 2020 VOLUME 14,14, N°

THE MORE YOU SEE ME THE MORE YOU LIKE ME. INFLUENCING THE NEGATIVE ATTITUDE TOWARDS INTERACTIONS WITH ROBOTS Submitted: 26th June 2019; accepted: 25th March 2020

Paweł Łupkowski, Filip Jański‑Mały DOI: 10.14313/JAMRIS/3‐2020/27 Abstract: The main aim of this paper is to present the study de‐ signed to check whether negative attitudes towards in‐ teractions with robots may be influenced by demonstra‐ ting videos presenting advanced modern robots. The atti‐ tude was measured with the use of the Negative Attitude toward Interactions with Robots questionnaire (NATIR). 66 subjects participated in the study divided into the pre‐ test, a video presentation and the post‐test. Our main fin‐ dings are the following. There is a significant difference between pre‐test and post‐test NATIR scores—the atti‐ tude towards interactions with robots improved after our subjects watched a video. We also observe an effect of participants gender on NATIR results—men had more po‐ sitive attitude than women. Keywords: Human‐Robot Interaction, Human unique‐ ness, Acceptance of robots, Negative attitude towards ro‐ bots, NATIR, BHNU

1. Introduction

10 10

In this paper we are presenting the study of the at‑ titude towards robots. Our aim was to check whether this attitude may be in�luenced by displaying videos presenting modern day advanced humanoid robots. It is worth to point out that the issues of human‑robot in‑ teraction are considered in this paper from a cognitive and psychological perspective. For a survey of appro‑ aches that are more focused on technological aspects and hardware and software implementations, we refer to, for instance, [6]. The issue of a positive attitude towards robots is becoming more and more important nowadays. This is due to the fact that we encounter real robots more often in a regular day situations, like e.g. vacuum cle‑ aners or autonomous cars. As it is pointed out in [16, p. 3–4] “the International Federation of Robotics has estimated that by 2019 more than 42 million robots have been sold for personal use; meaning, they are quickly becoming an unavoidable part of our social ecosystem”. Robots are also present in our common imagination due to famous movie productions (like “AUTOMATA” (2014), “Chappie” (2015), “Ghost in the Shell” (2017), “Blade Runner 2049” (2017)), TV series (“Westworld” (2016), “Altered Carbon” (2018)) and video games (e.g. “Detroit. Become Human” (2018)). What is more, robots are often a subject of a popu‑ lar media reports, see e.g. widely discussed 2017 in‑ terviews with the Sophia robot for the Good Morning Britain show (ITV) and CNBC; series of articles con‑ Articles

cerning autonomous cars and robo‑ethics (e.g. “Self‑ driving cars will kill people. Who decides who dies?” in Wired 09.21.2017; “How to punish a robot who committed a crime” (in Polish) for Gazeta Wybor‑ cza 10.16.2018) or recent discussion about sex ro‑ bots (e.g. “Prediction: Sex Robots Are The Most Dis‑ ruptive Technology We Didn’t See Coming” in Forbes 09.25.2018; “Sex robots and us” for BBC3, 04.08.2018 or the �et��i� 2018 documentary series “Watch us: Sex robots”). The issue of attitudes towards robots is also widely studied in the �ield of human‑robot interaction (HRI). As we read in [19, p. 18]: “…designing robots with human‑like traits can enhance their interactive and so‑ cial pro�iciency. Also, different degrees of human like‑ ness seem to impact differently potential user’s expec‑ tations and behavior”. Thus one of important contexts of HRI studies is the Uncanny Valley Hypothesis [12]. The hypothesis—stating that we will observe a decre‑ ase of af�inity for almost human‑like robots—is explo‑ red for real robots (see e.g. study of interaction with Geminoid HI‑1 humanoid robot presented in [1]). Un‑ canny Valley is also observed and studied for compu‑ ter generated characters in games and animated pro‑ ductions (see [4], [9], [11], [22]). HRI studies of pe‑ ople’s attitudes towards robots address also social is‑ sues related to robots. Authors of [23] investigate at‑ titudes towards service robots among German citi‑ zens, while [7] presents analysis of EU citizens’ atti‑ tudes towards robots in caring for the elderly. Also cross‑cultural studies of attitudes towards robots are conducted—see e.g. [15], where people’s acceptance of humanoid robots among UK and Japanese citizens is analyzed. Researchers are also interested how pre‑ vious contacts with robots in�luence aforementioned attitudes (see [17], [23]). There are also attempts of in‑ �luencing attitudes towards robots, e.g. Reich‑Stiebert and Eyssel [24] report the positive effect of subjects’ participation in prototyping process on their attitude towards robots in educational contexts. As a tool for the study described in this paper we have selected the well tested questionnaire called The Negative Attitude Toward Robots Scale (NARS) [13, 14]. NARS is designed to measure “psychological reactions evoked in humans by humanlike and non‑ humanlike robots” [17, p. 94]. NARS is widely used for studies addressing human‑robot interactions. Sy‑ rdal et al. [25] use NARS to explain participants’ eva‑ luations of real robot behavior styles; Ciechanowski et al. [2] employ NARS into a wide study of human‑ chatbot interaction; Dinet and Vivian in [3] describe

Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

results of a study of an attitude towards assistive ro‑ bots among French citizens; authors of [10] discuss the usage of NARS in the context of the uncanny val‑ ley effect for computer generated robots. Another inte‑ resting and important study is the one presented in [8, Chapter 3.3], in which the relationships between nega‑ tive attitudes, anxiety and an actual behavior toward robots were analyzed on the basis of interaction with Robovie (the humanoid robot). What is also impor‑ tant from our perspective is that the tool was success‑ fully adapted into Polish (NARS‑PL) [20]. Our study was conducted in Polish, thus we have decided to use the NARS‑PL. As authors of this adaptation claim: “[...] NARS‑PL is a useful tool to predict human responses to social robots in HRI studies in Poland.” [20, p. 70]. In the aforementioned adaptation two sub‑scales were identi�ied on the basis of obtained study results: The Negative Attitude toward Interactions with Ro‑ bots (NATIR) which aims at measuring the attitude to‑ wards interactions with robots and The Negative At‑ titude toward Robots with Human Traits (NARHT)— which “captures the responses to robots that display human traits like emotions, language, and agency” [20, p. 70]. NATIR items are listed in Section 2.2. Exemplary NARHT items are the following: ‑ I would feel uneasy if robots really had emotions. ‑ I would hate the idea that robots or arti�icial intelli‑ gences were making judgments about things.

‑ Something bad might happen if robots developed into living beings. As in our study we wanted to focus only on the as‑ pect of potential interactions with a robot, so we have decided to use only the NATIR sub‑scale. We have also decided to use The Belief in Hu‑ man Nature Uniqueness (BHNU) questionnaire. BHNU aims at capturing the “the extent to which humans re‑ serve human nature for their own group and deny the possibility of a human essence to robots” [20, p. 67]. We treat BHNU score as a useful information about our subjects—one may expect that when someone belie‑ ves that humans are unique s(he) will hold more nega‑ tive attitude towards such machines. (It is worth to no‑ tice that this relates to Turing’s intuitions concerning arti�icial intelligence in general. See ‘The Theological Objection’ and ‘The Heads in the Sand Objection’ dis‑ cussed in [26].) BHNU items are listed in Section 2.1. The paper is structured as follows. In the second section we present details about tools, the study de‑ sign and procedure. We also describe our subjects and research hypotheses. Section three contains results. The last—fourth—section covers summary and dis‑ cussion.

2. Methods 2.1. Tools

Tools used in the study were BHNU and NATIR questionnaires. As the study was conducted in Polish we have used their Polish versions presented in [20]. Below we present these tools (with items in the origi‑ nal English formulation).

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

1) NATIR 2) BHNU

Pre-test

Min. 2 weeks

Post-test

1) Video (Atlas) 2) NATIR

1) Video (Asimo) 2) NATIR

Group A

Group B

Fig. 1. The schema of the study BHNU questionnaire [20, p. 69]. Even if ultra-sophisticated... 1) a robot will never be considered as human being; 2) a robot will never feel the same emotions as a human being, 3) a robot will never use language in the same way as a human being; 4) a robot will always be a mechanical imitation of the human being; 5) a robot will never have consciousness; 6) a robot will never have morality. NATIR questionnaire [20, p. 69]. 1) I would feel uneasy if I was given a job where I had to use robots. 2) I would feel nervous operating a robot in front of other people 3) I would feel very nervous just standing in front of a robot. 4) I feel that if I depend on robots too much, something bad might happen. 5) I would feel paranoid talking with robot. 6) I am concerned that robots would be a bad influence on children. In both questionnaires participants responded on a 7‑point scale (1 – totally disagree to 7 – totally agree). The score of an individual at NATIR and BHNUS is cal‑ culated by summing up the scores of all the items in‑ cluded in the scale (see [13] and [20]).

2.2. Procedure The study was conducted with the use of online questionnaires (Google Forms). It consisted of two parts separated by at least two weeks break. The schema of the study is presented in Figure 1. Before each part of the study participants were informed about the aim of the study and their right to resign at any point without consequences. They were also in‑ formed that the gathered data will be processed only for scienti�ic purposes. Before each part, participants gave their consent to take part in the study. After com‑ pleting each part participants were thanked for their contribution. In the �irst part we asked our participants to �ill NATIR and BHNU questionnaires followed by ques‑ tions concerning their socio‑demographic data: age, gender, education and gaming habits. Articles

11 11

Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

The second part consisted of a short video presen‑ ting an advanced robot and its actions. The video was followed by NATIR questionnaire. [19, p. 19] provide an overview of previous studies indicating that the use of video materials for HRI studies proved to be a va‑ lid method. Participants �irst watched the whole video (it was embedded in the questionnaire, so participants would not leave Google Forms) and then they could proceed to the NATIR questionnaire (by clicking the “next” button). For this part we have used two videos (for groups A and B—see the study schema in Figure 1). Vi‑ deo for the group A presented Atlas (by Boston Dynamics) robot performing advanced movements, like traversing outdoor terrain or avoiding obstacles (https://youtu.be/hSjKoEva5bg, the video lasted for 00:01:00). Video for the group B presented Asimo (by Honda) robot singing and performing a dance mo‑ ves (https://youtu.be/gi71uXqCkvU, the video las‑ ted for 00:01:22). The robots are presented in Figures 2 and 3. There were no additional narration in both videos. The choice of robots for the study was arbi‑ trary but motivated by previous research, see e.g. [19] and [16], which suggest that Asimo should be evalua‑ ted as more friendly and likable than Atlas. For the second part of the study we have deci‑ ded not to include the BHNU questionnaire. Firstly, to keep this part relatively short in order to encourage more subjects to take a part. Secondly, as BHNU me‑ asures a general believes concerning human beings we would not expect to in�luence them by our short video stimuli. As we mention in the introduction we treat BHNU score as the valuable information about our participants. The �irst part of the study was carried out from the 22nd of December 2018 until the 4th of January 2019. The second stage started at the 18th January 2019 (to ensure at least 2 weeks break between parts for each participant) until the 23rd January 2019.

VOLUME 2020 VOLUME 14,14, N°N°33 2020

Fig. 2. Atlas robot (source https://commons.wikimedia.org). This robot was presented in the video for the group A.

2.3. Subjects

12 12

Participants were recruited from cognitive science students at the Institute of Psychology AMU (they re‑ ceived extra credits for participation) and via private communication as well as popular social networks. In the �irst part of the study 66 subjects took part— 40 women and 26 men aged from 17 to 45 (mean = 22.92, sd = 5.97, median =20). 23 declared higher edu‑ cation, 42 declared holding of a high school diploma, 1 was before graduating from a high school. Out of 66 subjects who took part in the study, 50% were the afo‑ rementioned cognitive science students. We have also collected the data addressing gaming habits of our participants. The reason for this is that vi‑ deo games are the most common way in which we may get into some form of interaction with (virtual) robotic characters—that is why we wanted to control this va‑ riable. The group characteristics is the following. 31% of subjects declared that they do not play video games at all; 41% play games once a month or less than once a month. 20% declare that they play few times a week and 8% that they play every day. We also asked about Articles

Fig. 3. Asimo robot (source https://commons.wikimedia.org). This robot was presented in the video for the group B. titles of the most played games by our subjects. The mostly repeated titles were not related to robots and robotic themes—these were ‘League of Legends’, ‘The Sims’, ‘Witcher’ and ‘Fifa’. We may say that our group of subjects was balanced when it comes to players and non‑players. What is more, our participants were not exposed to games, which are directly related to HRI is‑

Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

sues. (We are aware that gathering more data about our subjects would useful, but we wanted to keep our study reasonably short in order to ensure that most of participants will be willing to take part in the post‑ test.) The mean NATIR score for the �irst part of the study was 17.83 and the BHNU score was 29.09. The detailed characteristics of these results is presented in Table 1. For the second part of the study we have divi‑ ded the initial group into two balanced sub‑groups (in what follows we refer to them as A and B). For this part we have sent separate invitations to group A and to group B. While dividing our subjects we have taken into account the following factors: gender, age, BHNU and NATIR results from the �irst part. Group A con‑ sisted of 33 subjects—20 women and 13 men with average age of 23.18 (sd = 6.68, median 20). Group B also consisted of 33 subjects, of which 20 were wo‑ men and 13 were men. The mean age for this group was 22.67 (sd = 5.25, median 20). As for the mean BHNU scores for the groups they were following: 17.91 (sd = 8.46, median 17.00) and 17.76 (sd = 6.03, me‑ dian 18.00). The t‑test showed no statistically signi�i‑ cant difference between these scores (p = 0.9335). For mean NATIR results we got: 27.82 (sd = 9.29, median 28) and respectively 30.36 (sd = 7.24, median 29). The difference in results was not statistically signi�icant (t‑ test p = 0.2192). Finally, in the second part of the study 50 subjects took part (group A: 28 subjects; group B: 22), so 16 participants from the initial group have not accepted our invitation. The demographic characteristics are the following: 33 women and 17 men, mean age = 22.14 (sd = 6.34, median =19); 11 subjects with higher education and 39 with high school diploma. 2.4. Hypotheses

Our research hypotheses were the following. (H1) We will observe a positive correlation between BHNU and NATIR results for the �irst part of the study.

(H2) We will observe differences in results for women and men. Women should have higher BHNU and NA‑ TIR results than men. (H3) There will be a difference in NATIR results in group A and group B in the second part of the study.

(H4) NATIR results in the �irst part of the study and in the second part should differ. Results from the second part should be lower. As for (H1) and (H2) they are derived from the re‑ sults reported in [20], [5] and [19]. As BHNU score tells us to which extent humans reserve human na‑ ture to human beings and deny such a nature for ro‑ bots one may expect that it should correlate with the NATIR results. The more reluctant a subject is in as‑ cribing human characteristics to robots, the more re‑ luctant (s)he will be when it comes to interacting with them. Pochwatko et al. [20, p. 70] report a signi�icant effect on participants gender on NATIR results—men had more positive attitude than women. Authors of

VOLUME 2020 VOLUME 14,14, N°N°3 3 2020

Tab. 1. BHNU and NATIR scores for the first part of the study Score BHNU NATIR

N 66 66

Min 6 6

Max 42 41

Mean 29.09 17.83

SD 8.36 7.29

Median 28.50 17.50

[18] (see also [19]) suggest that the possible expla‑ nation for such a result is that male and female par‑ ticipants associate robots with different contexts, in which they may potentially get into interaction with them (industrial vs. domestic robots; help with unem‑ ployment vs. help at home). We expected similar ten‑ dency for BHNU results. Especially that our subject were informed at the beginning of the study, that it concerns human‑robot interactions, so they �illed out BHNU questionnaire with this information in mind. In the second part of the study participants of groups A and B were presented with two different vi‑ deos. They presented two different humanoid robots Atlas and Asimo. The design of these robots is different when it comes to revealing its construction elements (Atlas has more elements which are visible, as joints, cables and sensors)—see Figures 2 and 3. The are also differences when it comes to actions performed by robots. Atlas video presents an agile machine coping with dif�icult environment. It is more about physical activities. Asimo sings a song coordinated with dance hand movements—presenting higher‑level cognitive functions. We expected that different look and actions performed by robots will evoke different reactions of subjects in group A and B (H3). For the (H4) our expectation was that in the �irst part our participants used different ideas concerning robots that they have developed on the basis of their experience and knowledge. When asked NATIR ques‑ tions in the pre‑test they were not pointed at any spe‑ ci�ic robots, thus our participants use the aforementi‑ oned general ideas. Post‑test video should make them focused on recent robotics developments and abilities of modern robots and thus we expect that it will in�lu‑ ence their NATIR answers.

3. Results

For the data analysis we used R statistical software ( [21]; version 3.5.1). Reliability of the BHNU and NATIR questionnaires for the �irst part of the study (N = 66) is satisfactory— Cronbach’s alpha coef�icients are respectively 0.85 and 0.81. In the second part only NATIR questionnaire was used. Its reliability is also satisfactory (Cronbach’s al‑ pha 0.85 for N = 50). 3.1. BHNU and NATIR Scores

We have checked whether we observe correlation between BHNU and NATIR results for the �irst part of the study. Detailed BHNU and NATIR scores are pre‑ sented in Table 1. The distribution of NATIR and BHNU results was normal (as indicated by the Shapiro‑Wilk normality test) thus we use the Pearson’s test for correlation Articles

13 13

R = 0.36 , p = 0.0031

VOLUME N°33 2020 2020 VOLUME 14,14, N°

Journal Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

NATIR

Group A 10

BHNU

Fig. 4. BHNU and NATIR scores correlation plot

R = 0.85 , p = 6.5e−15

NATIR2

BHNU

Fig. 5. BHNU and NATIR2 (post‐test) scores correlation plot check. We observe a weak positive relationship—the higher the BHNU result is, the higher NATIR results are (r = 0.36; p = 0.0031). This result is presented in Figure 4. Interestingly the observed correlation becomes stronger when we take BHNU score and NATIR2 sco‑ res (for the post‑test). For this comparison we have taken 50 participants who participated in both parts of the study (see discussion in Section 3.4). Distribu‑ tions of BHNU scores for the selected sub‑group and NATIR2 results in the post‑test were normal (as indi‑ cated by the Shapiro‑Wilk normality test), so we have used the Pearson’s correlation check—the result is r = 0.85; p = 6.5e − 15 (Figure 5). 3.2. Gender and Attitude Towards Robots

14 14

In the �irst part of the study 40 women and 26 men took part—they �illed out two questionnaires� BHNU and NATIR. As for BHNU the mean result for women Articles

Group B

Fig. 6. NATIR results comparison for group A and group B was 31.32 (sd = 7.26, median 30.50) and for men it was much lower 15.81 (sd = 5.25, median 15.50). The difference is statistically signi�icant as the t‑test results show (p < 2.2e − 16). The mean result in NATIR for women’s group was 19.15 (sd = 8.04, median 18.00) while (similarly as for BHNU) for men’s group it was lower 15.81 (sd = 5.25, median 15.50). This difference is statistically signi�i‑ cant as the t‑test results show (p = 0.0495). This ten‑ dency in results is also observable in the second part of the study. In this part the mean result in NATIR for women’s group (N = 33) was 21.27 (sd = 8.71, me‑ dian 20.00) while for men’s group (N = 17) it was lo‑ wer 16.88 (sd = 4.92, median 16.00). This difference is statistically signi�icant as the t‑test results show (p = 0.0275). 3.3. Groups A and B Comparison

In the second part of the study 50 subjects took part—28 in group A and 22 in group B. For the com‑ parison of NATIR scores between groups we have se‑ lected a random sub‑sample of 22 results from group A. The distribution of NATIR results is normal in both groups (as indicated by the Shapiro‑Wilk normality test). Mean NATIR result for A is 22.27 (sd = 8.67, me‑ dian 21) while the result for group B is lower as the mean score equals 18.36 (sd = 6.28, median 17.5)— see Figure 6. However, this difference is not statisti‑ cally signi�icant (t‑test, p = 0.0949). 3.4. Pre‐test and Post‐test Natir Scores

As no signi�icant differences were observed bet‑ ween groups A and B for NATIR scores in the second part of the study, we have decided to take all the re‑ sults together. We will report pre‑test results (from the �irst part of the study) as NATIR1 and post‑test re‑ sults (from the second part) as NATIR2. For the com‑ parison 50 subjects who participated in both parts of the study were taken into account. The distribution

Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

with the one reported in [10]) con�irms that the Polish adaptation provides a proper tool for HRI research on Polish participants.

Pre−test

Post−test

Fig. 7. Pre‐test versus post‐test NATIR results of results in NATIR1 and NATIR2 was normal (as in‑ dicated by the Shapiro‑Wilk normality test), thus we have used the paired t‑test to establish the pre‑test and post‑test differences. A statistically signi�icant (p = 3.585e − 09) decrease in NATIR results is observable between NATIR1 and NATIR2 (from 29.54, sd = 8.42 to 19.90, sd = 7.84). The difference between NATIR1 and NATIR2 is presented in Figure 7. Summary of our �indings is the following: 1) Reliability of BHNU (for the �irst part of the study) and NATIR (for both parts) questionnaires is satis‑ factory. 2) There is a weak positive correlation between BHNU and NATIR scores for the �irst part of the study. (H1) is con�irmed. 3) In our subjects’ group we observe that women are less inclined to interact with robots and have stron‑ ger believes concerning human uniqueness than men. (H2) is con�irmed.

4) We observe no statistically signi�icant differences in NATIR scores between groups A and B in the second part of the study. Video manipulation was not successful in terms of modifying attitudes dif‑ ferently in these groups. (H3) is not con�irmed.

5) There is an observable difference between pre‑ test (�irst part) and post‑test (second part) NATIR results—post‑test scores are lower that these in pre‑test. The difference is statistically signi�icant. (H4) is con�irmed. (We should however stress that this effect is observed for our relatively small rese‑ arch group and may be a direct effect of the video, as such will not not last for long—we discuss these issues in the following section.)

4. Conclusion

Results reported in the previous section show that BHNU and NATIR questionnaires (in their Polish adap‑ tations) have a good internal consistency (as mea‑ sured with Cronbach’s alpha). This result (together

For our research group we also observe (weak) correlation between BHNU and NATIR scores. This is in line with previous results reported in [20] and [5] (as we read in [5]: “[...] the stronger the belief in hu‑ man nature uniqueness, the less positive the attitude towards interactions with robots [...]”). Also the obser‑ ved effect of gender on NATIR score—men had more positive attitude than women is in line with previ‑ ously reported results. What is worth stressing this tendency holds for pre‑test and post‑test results in our study. We observe analogous results for human‑ uniqueness scale: men had weaker belief in human‑ uniqueness than women. As BHNU and NATIR sco‑ res correlate we may explain this observation along the same lines, suggested in [18, 19] and discussed in Section 2.4. As the most important �inding of our study we con‑ sider the effect of in�luencing NATIR scores by pre‑ senting a short video to our subjects. Our subjects sig‑ ni�icantly lowered their negative attitude towards ro‑ bots after (at least) two weeks break from pre‑test. It is worth to stress that videos were rather simple pre‑ sentation of capabilities of modern day robots. The re‑ sult needs further investigations as we discuss below in the context of limitations of our study. However it may be used at least in two ways. First of all, similarly as [24], it suggest that we may in�luence the attitudes towards robots without interactions with real devices. We may use videos, games or computer simulations. Secondly, the results indicate the need for pre‑test and post‑test scheme while using NATIR preceded by a vi‑ deo presentation of robots in order to control the in‑ �luence of the used material. The procedure using vi‑ deo presentation of robots (but without pre‑testing) was used in the Polish adaptation study reported in [20]—a subject was presented with one of three pre‑ pared videos about robots, which was then followed by NATIR questionnaire. One of the main limitations of our study is visible in the lack of con�irmation of (H3). We would under‑ stand better the in�luence of the presented video ma‑ terial when the difference between two groups in the post‑test would be observed. However, no such signi‑ �icant difference appeared. One of possible explanati‑ ons of this fact may be that differences between pre‑ sented robots and their actions were too small for our subjects. The drawback of the study is that we have not collected any additional qualitative data from par‑ ticipants, e.g. concerning their motivations and rea‑ sons for choosing answers to NATIR questions. Such data would certainly shed light on the obtained re‑ sults. Employing the Anthropomorphism Scale used in the Polish adaptation of NARS would be also bene�i‑ cial. For future studies we will also consider a diffe‑ rent video stimuli design. The wider range of robots may be used (e.g. ranging from non‑anthropomorphic military robots to very human‑like ones like Sophia). In our opinion it is also worth considering the use of Articles

15 15

Journal Journal of ofAutomation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

videos presenting explicit human‑robot interactions which would evoke more emotional reactions (like e.g. widely commented videos from Boston Dynamics pre‑ senting an employee interrupting actions performed by Atlas—see an overview ‘Boston Dynamics New Vi‑ deo Is Just Another Reason Robots Will Hate Us One Day’ in The Washington Post 02.21.2018). In our opinion further study would be required on a larger group of subjects (with more variety when it comes to age and education). We would like to add additional questions concerning previous experien‑ ces with robots and such (possible) experiences du‑ ring the break period between pre‑testing and post‑ testing. It would be also bene�icial to add an additio‑ nal group of subjects. This group would not be presen‑ ted with any video material, but simply �illed our NA‑ TIR questionnaire. This would allow to test whether a simple repetition of the same questionnaire may so‑ mehow in�luence its results. What is more, for the pre‑ test and post‑test plan of our study a re‑test should be added in order to control how long the observed atti‑ tude change would last.

AUTHORS

Paweł Łupkowski∗ – Reasoning Research Group, Adam Mickiewicz University, Szamar‑ zewskiego 89/AB, 60–568 Poznań , Poland, e‑mail: Pawel.Lupkowski@amu.edu.pl, www: http://reasoning.amu.edu.pl/. Filip Jański‑Mały – Reasoning Research Group, Adam Mickiewicz University, Szamarzew‑ skiego 89/AB, 60–568 Poznań , Poland, www: http://reasoning.amu.edu.pl/. ∗

Corresponding author

REFERENCES

[1] C. Becker‑Asano, K. Ogawa, S. Nishio, and H. Ishi‑ guro, “Exploring the uncanny valley with gemi‑ noid hi‑1 in a real‑world application”. In: Procee‑ dings of IADIS International conference interfa‑ ces and human computer interaction, 2010, 121– 128.

[2] L. Ciechanowski, A. Przegalinska, and K. Weg‑ ner, “The necessity of new paradigms in measu‑ ring human‑chatbot interaction”. In: Internatio‑ nal Conference on Applied Human Factors and Er‑ gonomics, 2017, 205–214, 10.1007/978‑3‑319‑ 60747‑4_19. [3] J. Dinet and R. Vivian, “Exploratory investigation of attitudes towards assistive robots for future users”, Le travail humain, vol. 77, no. 2, 2014, 105–125, 10.3917/th.772.0105.

[4] T. Geller, “Overcoming the uncanny valley”, IEEE computer graphics and applications, vol. 28, no. 4, 2008, 11–17, 10.1109/MCG.2008.79.

16 16

[5] J.‑C. Giger, D. Moura, N. Almeida, and N. Piçarra, “Attitudes towards Social Robots: The Role of Be‑ lief in Human Nature Uniqueness, Religiousness Articles

VOLUME 14, N° N°33 2020 VOLUME 14,

and Taste for Science Fiction”. In: Proceedings of II International Congress Interdisciplinarity in So‑ cial and Human Sciences, 2017.

[6] M. A. Goodrich, A. C. Schultz, et al., “Human–robot interaction: a survey”, Foundations and Trends® in Human–Computer Interaction, vol. 1, no. 3, 2008, 203–275, 10.1561/1100000005. [7] J. Hudson, M. Orviska, and J. Hunady, “People’s at‑ titudes to robots in caring for the elderly”, Inter‑ national Journal of Social Robotics, vol. 9, no. 2, 2017, 199–210, 10.1007/s12369‑016‑0384‑5.

[8] T. Kanda and H. Ishiguro, Human‑robot inte‑ raction in social robotics, CRC Press: Boca Raton, London, New York, 2016. [9] J. Kä tsyri, M. Mä kä rä inen, and T. Takala, “Testing the ‘uncanny valley’ hypothesis in semirealistic computer‑animated �ilm characters: An empiri‑ cal evaluation of natural �ilm stimuli”, Internati‑ onal Journal of Human‑Computer Studies, vol. 97, 2017, 149–161, 10.1016/j.ijhcs.2016.09.010.

[10] P. Łupkowski and M. Gierszewska, “Attitude to‑ wards humanoid robots and the uncanny val‑ ley hypothesis”, Foundations of Computing and Decision Sciences, vol. 44, no. 1, 2019, 101–119, 10.2478/fcds‑2019‑0006. [11] P. Łupkowski, M. Rybka, D. Dziedzic, and W. Wło‑ darczyk, “The background context condition for the uncanny valley hypothesis”, International Journal of Social Robotics, vol. 11, no. 1, 2019, 25– 33, 10.1007/s12369‑018‑0490‑7.

[12] M. Mori, K. F. MacDorman, and N. Kageki, “The uncanny valley [from the �ield]”, IEEE Robotics & Automation Magazine, vol. 19, no. 2, 2012, 98–100, 10.1109/MRA.2012.2192811, (Original work published in 1970 in Japaneese). [13] T. Nomura, T. Kanda, and T. Suzuki, “Experi‑ mental investigation into in�luence of negative attitudes toward robots on human–robot inte‑ raction”, AI & Society, vol. 20, no. 2, 2006, 138– 150, 10.1007/s00146‑005‑0012‑7.

[14] T. Nomura, T. Kanda, T. Suzuki, and K. Kato. “Ex‑ ploratory investigation into in�luence of negative attitudes toward robots on human‑robot inte‑ raction”. In: Mobile Robots: towards New Appli‑ cations. InTech, 2006. [15] T. T. Nomura, D. S. Syrdal, and K. Dautenhahn, “Differences on social acceptance of humanoid robots between Japan and the UK”. In: Procs 4th Int Symposium on New Frontiers in Human‑Robot Interaction, 2015. [16] J. Palomä ki, A. Kunnari, M. Drosinou, M. Koverola, N. Lehtonen, J. Halonen, M. Repo, and M. Laak‑ asuo, “Evaluating the replicability of the un‑ canny valley effect”, Heliyon, vol. 4, no. 11, 2018, e00939, 10.1016/j.heliyon.2018.e00939.

[17] N. Piçarra, J.‑C. Giger, G. Pochwatko, and G. Gonçalves, “Validation of the portuguese version of the negative attitudes towards

Journal of Automation, Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

robots scale”, European Review of Applied Psychology, vol. 65, no. 2, 2015, 93–104, 10.1016/j.erap.2014.11.002.

[18] N. Piçarra, J.‑C. Giger, G. Pochwatko, and G. Gonçalves, “Making sense of social robots: A structural analysis of the layperson’s social representation of robots”, European Review of Applied Psychology, vol. 66, no. 6, 2016, 277–289, 10.1016/j.erap.2016.07.001.

[19] N. Piçarra, J.‑C. Giger, G. Pochwatko, and J. Moż a‑ ryn, “Designing social robots for interaction at work: socio‑cognitive factors underlying inten‑ tion to work with social robots”, Journal of Au‑ tomation Mobile Robotics and Intelligent Systems, vol. 10, no. 4, 2016, 17–26, 10.14313/JAMRIS_4‑ 2016/28. [20] G. Pochwatko, J.‑C. Giger, M. Ró ż ań ska‑Walczuk, J. S� widrak, K. Kukie�ka, J. Moż aryn, and N. Piçarra, “Polish version of the negative attitude toward robots scale (NARS‑PL)”, Journal of Automation Mobile Robotics and Intelligent Systems, vol. 9, 2015, 10.14313/JAMRIS_3‑2015/25.

[21] R Core Team. R: A language and environment for statistical computing. R Foundation for Statisti‑ cal Computing, Vienna, Austria, 2013. Accessed on: 2020‑11‑10.

[22] D. Ratajczyk, M. Jukiewicz, and P. Lupkowski, “Evaluation of the uncanny valley hypothesis ba‑ sed on declared emotional response and psy‑ chophysiological reaction”, Bio‑Algorithms and Med‑Systems, vol. 15, no. 2, 2019, 10.1515/bams‑ 2019‑0008.

[23] N. Reich and F. Eyssel, “Attitudes towards ser‑ vice robots in domestic environments: The role of personality characteristics, individual intere‑ sts, and demographic variables”, Paladyn, Journal of Behavioral Robotics, vol. 4, no. 2, 2013, 123– 130, 10.2478/pjbr‑2013‑0014. [24] N. Reich‑Stiebert, F. Eyssel, and C. Hohnemann, “Involve the user! Changing attitudes toward ro‑ bots by user participation in a robot prototyping process”, Computers in Human Behavior, vol. 91, 2019, 290–296, 10.1016/j.chb.2018.09.041. [25] D. Syrdal, K. Dautenhahn, K. Koay, and M. Wal‑ ters. The Negative Attitudes Towards Robots Scale and Reactions to Robot Behaviour in a Live Human‑Robot Interaction Study, 109–115. SSAISB, 4 2009.

[26] A. M. Turing, “Computing Machinery and Intel‑ ligence”, Mind, vol. LIX, no. 236, 1950, 443–455, 10.1093/mind/LIX.236.433.

Articles

17 17

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

Preface to Special Issue on Recent Advances in Information Technology III

DOI: 10.14313/JAMRIS/3-2020/28 This issue of the Journal of Automation, Mobile Robotics and Intelligent Systems is devoted to selected aspects of current studies in the area of Information Technology - as presented by young talented contributors working in this field of research. This special issue is already the third edition of this series. Among included papers, one can find contributions dealing with the modelling and controlling of discrete hybrid systems, microcontrollers, the Internet of Things, 3D visualization technology, sliding mode control and finally Extended Kalman Filters. The idea of creating this special issue was born as a result of broad and interesting discussions during the Sixth Doctoral Symposium on Recent Advances in Information Technology (DS-RAIT 2019), held in Leipzig (Germany) on September 1-4, 2019 as a satellite event of the Federated Conference on Computer Science and Information Systems (FedCSIS 2019). The aim of this meeting was to provide a platform for the exchange of ideas between early-stage researchers in Computer Science (PhD students in particular). Furthermore, the Symposium was to provide all participants an opportunity to obtain feedback on their ideas and explorations from the vastly experienced members of the IT research community who had been invited to chair all DS-RAIT thematic sessions. Therefore, submission of research proposals with limited preliminary results was strongly encouraged. Here, we would like to individually mention the contributions entitled “Sustainable Management of Marine Fish Stocks by Means of Sliding Mode Control” written by Katharina Benz, Claus Rech, and Paolo Mercorelli (Leuphana University of Lueneburg) and the paper “Proposal of Mechatronic Devices Control using Mixed Reality” by Erich Stark and Erik Kučera, Peter Drahoš, and Oto Haffner (Slovak University of Technology in Bratislava). These contributions have received the Best Paper Award at DS-RAIT 2019. This issue contains the following DS-RAIT papers in their special, extended versions.

The first paper, entitled Modelling and Control of Discrete-Event Systems Using Petri Nets and Arduino Microcontrollers, and authored by Erik Kučera, Oto Haffner and Roman. Leskovský, describes a design of a new software system for modelling and control of discrete-event and hybrid systems. For this purpose, Arduino family microcontroller and Petri net technology were applied. The developed software tool was successfully verified for control of laboratory systems. The proposed system offers a graphical way for designing control algorithms for hybrid and mainly discrete-event systems. In the opinion of editors and reviewers, this work, apart from its practical values, might have large influence on the development of contemporary hybrid control systems. Erich Stark, Erik Kučera and Oto Haffner, in their work entitled Proposal of IoT Devices Control using Mixed Reality and QR Codes, address problems related to establishing appropriate ways of connecting two areas: the Internet of Things (IoT) and mixed reality, where it is possible to control and monitor mechatronic devices using a mobile device with augmented/mixed reality support. The described proposal of interconnecting IoT and mixed reality can bring about a new form of Human Machine Interface that can save time for users or companies (the Industrial Internet of Things (IIoT)). In the opinion of the members of the Program Committee of DS-RAIT, this work, because of its great presentation of applicational aspects and promising results, was awarded the Best Paper of the event.

The paper entitled Online Control Education Using 3D Holographic Visualisation, was written by the team consisting of Jakub Matišák, Matej Rábek, Katarína Žáková. In this paper, the Authors take under consideration some aspects of Interactive 3D visualization technology. This contribution to the application for interactive teaching of control theory was well-described. This tool could be applied to simulating a holographic model of a selected mechatronic system that is a digital visualization of the real device.

Finally, Katharina Benz, Claus Rech, Paolo Mercorelli and Oleg Sergiyenko, provide a paper entitled Two Cascaded and Extended Kalman Filters and Sliding Mode Control for Sustainable Management of Marine Fish Stocks. The paper deals with the problem of controlling Marine Fish Stocks by the way of implementing a model described by the Lotka-Volterra equations with sliding mode control techniques, and includes some interesting methods for identifying investigated system parameters such as utilizing Extended Kalman Filters. Again in the case of

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

this contribution, the members of the Program Committee of DS-RAIT awarded the Best Paper of the event to this work – because of its modern aspect, great presentation and promising results.

We would like to thank all those who participated in, and contributed to the Symposium program, as well as all the authors who have submitted their papers. We also wish to thank all our colleagues, the members of the Program Committee, both for their hard work during the review process and for their cordiality and outstanding local organization of the Conference. Editors: Piotr A. Kowalski Systems Research Institute, Polish Academy of Sciences and Faculty of Physics and Applied Computer Science, AGH University of Science and Technology

Szymon Łukasik Systems Research Institute, Polish Academy of Sciences and Faculty of Physics and Applied Computer Science, AGH University of Science and Technology

Articles

Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME N°33 2020 2020 VOLUME 14,14, N°

MODELLING AND CONTROL OF DISCRETE‐EVENT SYSTEMS USING PETRI NETS AND ARDUINO MICROCONTROLLERS Submitted: 17th March 2020; accepted: 10th August 2020

Erik Kučera, Oto Haffner, Roman Leskovský DOI: 10.14313/JAMRIS/3‐2020/29 Abstract: The main aim of proposed article is the design of new software system for modelling and control of discrete‐ event and hybrid systems using Arduino and similar mi‐ crocontrollers. In this paper we propose a new tool. It is based on Petri nets and it is called PN2ARDUINO. It offers a capability of communication with the microcontroller. Communication with the microcontroller is based on mo‐ dified Firmata protocol so control algorithm can be im‐ plemented on all microcontrollers that support this type of protocol. The developed software tool was successfully verified for control of laboratory systems. It can also be used for education and also for research purposes as it offers a graphical way for designing control algorithm for hybrid and mainly discrete‐event systems. Proposed tool can enrich education and practice in the field of cyber‐ physical systems (Industry 4.0). Keywords: Hybrid systems, Petri‐nets, Microcomputer‐ based control, Motor, Discrete‐event dynamic systems

1. Introduction

20 20

Development of various systems is a complex dis‑ cipline that includes many activities, e.g. system de‑ sign, a speci�ication of required properties, implemen‑ tation, testing and further development of the system. As these operations are challenging and important for the �inal product, it is appropriate and necessary to create a model of the system. Development of cont‑ rol methods of discrete‑event and hybrid systems be‑ longs to the modern trends in automation and mecha‑ tronics. Hybrid system is a combination of continuous and discrete event systems. Control of such systems brings new challenges because it is necessary to join control methods of discrete event systems (where for‑ malism of Petri nets can be helpful) and classic cont‑ rol methods of continuous systems. With good metho‑ dology and software module, these approaches can be synergistically combined. This will give us an appro‑ priate and unique control system that allows harmo‑ nizing discrete event control methods with the met‑ hods of control of continuous systems (e.g. PID algo‑ rithms). Effective cooperation of these approaches al‑ lows to control hybrid system. This method would be useful in systems where it is necessary to use different control algorithms (for example PID controllers with different parameters) according to the state of the sy‑ stem. The concept of Petri nets is capable of covering a management of these control rules in a very ef�icient, robust and well‑arranged (graphical) way. This paper Articles

is aimed to present new Petri Net tools for modelling and control of discrete‑event and hybrid systems. Case studies for control of laboratory �ire alarm system and DC motor are also presented. In papers [1] and [2] authors deal with usage of hy‑ brid and colour Petri nets for modelling of crossroads and traf�ic on highways. �rom these authors, there are also interesting projects from the �ield of manufactu‑ ring systems [3] and [4]. Unfortunately, it is not men‑ tioned whether the results are only theoretical models or have been simulated using a SW tool or deployed in practice. In [5] author developed an interesting software tool that supports hybrid Petri nets named Visual Ob‑ ject Net++. There a lot of papers (mainly from Roma‑ nian author [6] and [7]) that describes capabilities of Visual Object Net++. This tool is not open‑source and it is not further developed. Software tool Snoopy [8] offers modelling using many classes of Petri nets like stochastic, hybrid, co‑ lour, music Petri nets, etc. Using this tool, many ty‑ pes of research in the �ield of biology and chemistry are being solved. Unfortunately, the source code is not available. In [9] and [10] Coloured Petri nets are used for control of computer model of automated storage and retrieval system. As an interesting way of research, a Modelica lan‑ guage and open‑source tool OpenModelica appeared. There is a library that supports modelling by Petri nets in this tool. One of the advantages of OpenModelica is that PN model can be connected with other com‑ ponents of Modelica. The �irst Petri net toolbox was introduced in [11]. An extension of this toolbox was described in [12]. The greater addition to the tool‑ box was made by the German author who enriches it by a support of extended hybrid Petri nets for model‑ ling of processes in biological organisms [13] and [14]. This tool was developed primarily for commercial tool Dymola and not for OpenModelica, so applicability in scienti�ic research and extensibility is limited. During 2015 the team that developed PNlib published modi‑ �ied version of PNlib that partially worked in OpenMo‑ delica. Unfortunately, it was not possible to use Open‑ Modelica for control purposes using microcontrollers because of lack of COM port communication support. Wolfram SystemModeler is easy‑to‑use modelling and simulation program for cyber‑physical systems [15]. It is based on Modelica language, too, but it is a proprietary software. Great advantage of Wolf‑ ram SystemModeler is a support of communication

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

with serial port. There a free extension ModelPlug that provides connection of simulation with microcon‑ trollers that supports Firmata protocol [16]. Unfortu‑ nately Wolfram SystemModeler supports only State‑ Graph [17] and not High‑level Petri nets (PNlib). The solution can be an integration of PNlib library to Wolf‑ ram SystemModeler. Attempts of this integration ap‑ peared in [18]. Unfortunately these attempts were not successful due to the various complications with Sy‑ stemModeler’s interpreter from Modelica to C++ lan‑ guage.

2. Description of PN2ARDUINO

Developed

Tool

As it was realized that there is no complex SW solu‑ tion to support control of discrete event and hybrid sy‑ stem by microcontrollers using High‑level Petri nets, it was necessary to develop it. As a basis for such soft‑ ware, PNEditor [19] was chosen. This tool is open‑ source. The developed extension of this tool is named PN2ARDUINO and is fully tested in [20] and [21]. The main topic of this paper is an introduction to this deve‑ loped software that can be used for control of discrete event and hybrid systems and its veri�ication on labo‑ ratory discrete‑event and hybrid system. There are more concepts of control using Petri nets. Petri net as a control logic is necessary to con‑ nect with the controlled system (e.g. using microcon‑ troller). One of the main aspects of the control system design is the question whether the Petri net’s logic should be stored in the microcontroller or into the PC (which can communicate with microcontroller). Both approaches have their advantages and disadvantages. If the Petri net’s logic is stored in the microcontrol‑ ler, the main advantage is the independence of con‑ trol unit from the software application (program on PC). The Petri net logic is modelled using PC, and then the Petri net is translated into program code which is loaded into the microcontroller. Then PC and mi‑ crocontroller can be disconnected. The advantage is also the capability of control in real time. Disadvanta‑ ges are limited computational and memory resources of the microcontroller. Following disadvantage is the need of repeating compiling and uploading the pro‑ gram into the microcontroller (mainly during develop‑ ment phase). The proposed solution is shown in Fig. 1. Computer

Microcontroller Petri net logic

Controlled system

Fig. 1. Simple scheme of proposed solution ‐ Petri net’s logic in microcontroller When the Petri net’s control logic is stored in spe‑ cialized SW application on PC, this solution gives an opportunity to control the system directly from it. In the microcontroller, only the program with communi‑ cation protocol is stored. This communication proto‑ col (in our case it is Firmata [16]) is used for commu‑ nication between PC and microcontroller. This solu‑

tion eliminates the necessity of recompiling and reu‑ ploading the program during development. The next advantage is the elimination of restrictions on compu‑ ting and storage resources because PC has (in compa‑ rison with microcontroller) almost unlimited resour‑ ces. One of the disadvantages is that the control sy‑ stem cannot react in real time. The proposed solution is shown in Fig. 2. Computer Petri net logic

Microcontroller

Controlled system

Fig. 2. Simple scheme of proposed solution ‐ Petri net’s logic in PC In Table 1, these differences are speci�ied. New software module PN2ARDUINO was based on the second approach. The Petri net runs on the perso‑ nal computer. For communication between SW appli‑ cation and microcontroller, the protocol Firmata [16] was used. Firmata is a protocol that is designed for communication between microcontroller and compu‑ ter (or mobile device like a smartphone, tablet, etc.). This protocol can be implemented in �irmware of va‑ rious microcontrollers. Mostly Arduino‑family micro‑ controllers are used. On PC the client library is needed. These libraries are available for many languages like Java, Python, .NET, PHP, etc. Firmata protocol is based on MIDI messages [22]. On the Arduino side, Standard Firmata 2.3.2 ver‑ sion is used. The client application on PC is based on Firmata4j 2.3.3 library which is programmed in Java. The advantage of using Firmata is that another micro‑ controller compatible with Firmata can be used. PN2ARDUINO extends PNEditor with many featu‑ res. For Petri nets modelling, there is a capability of adding time delay to transitions and capacity for pla‑ ces. Also, automatic mode of �iring transition was ad‑ ded for automatic system control purposes as only ma‑ nual mode was present in PNEditor. PN2ARDUINO brings a new communication mo‑ dule to PNEditor. This module communicates with the compatible microcontroller. This module consists of two parts. The �irst one provides the creation of connection with the microcontroller, so it sets COM port where the microcontroller is connected. The se‑ cond part provides the implementation of a capabi‑ lity of adding Arduino components to Petri net’s places and transitions. These types of Arduino components are supported: digital input and output, analog input, servo control, PWM output, message sending, custom SYSEX message [16] sending. In Fig. 3, the use‑case diagram of developed SW tool can be seen. Class diagram is shown in Fig. 4. As it was stated, transitions and places can be as‑ sociated with Arduino components. Digital and ana‑ log inputs serve as enabling conditions for transitions in Petri net. Digital and PWM outputs and messages serve as the executors of the respective actions. The interesting functionality is a capability of sen‑ ding custom SYSEX messages. The user must enter SY‑ Articles

21 21

Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N°33 2020

Tab. 1. Comparison of two concepts of system control using Petri nets Petri net logic in PC limited capability of real‑time control much more computation and memory resources available code in microcontroller does not need recompiling PC must be still online

Petri net logic in microcontroller real‑time control limited computation and memory resources during development repeated compiling is needed independence of control unit

User «extend»

Adding of a transition to the net

«extend»

Start automatic mode

Adding of time delay to the transition

Adding of PWM output

Adding of servo

Adding of message sending capability

Adding of custom SYSEX message sending capability

Fig. 3. PN2ARDUINO ‐ Use‐case diagram SEX command (0x00 - 0x0F) and optionally also the content of the message. The message is sent when the token comes to the place or when the transition is �i‑ red. For example, SYSEX messages are used in the pro‑ posed example of hybrid control in the last section of the paper. �ere, the SYSEX message noti�ies the mi‑ crocontroller that a different PID algorithm should be used for system control. Then PID algorithm is swit‑ ched, and the controlled system remains stable.

A main window of PN2ARDUINO consists of a quick menu, main menu, canvas for Petri net model‑ ling and log console. PN2ARDUINO supports two mo‑ des ‑ design mode and control mode. Control mode is manual and automatic.

22 22

Firstly, it is necessary to initialize communication with Arduino (Setup board in the menu). Then it is possible to add Arduino component to the place or the transition (Fig. 5). The example of analog input can be Articles

seen in Fig. 6. Time politics are also supported. To the transiti‑ ons, it is possible to add time delay which can be de‑ terministic or stochastic.

3. Case Study: Control of Laboratory Discrete‐ Event System For veri�ication of proposed software tool and met‑ hod of discrete‑event systems control it was necessary to design an education laboratory model of such sy‑ stem. A �ire alarm model was built. The scheme can be seen in Fig. 7. This model consists of an active buzzer, photo‑ resistor, three resistors and NPN transistor. NPN tran‑ sistor is mandatory for active buzzer connection. The LED of Arduino in pin 13 is also used. Photo‑resistor was used instead of the smoke sensor because of the less complicated feasibility of experiment.

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

«interface»

Subject

ArduinoListener

+ registerArduinoListeners (arduinoListener : ArduinoListener) + removeArduinoListener (arduinoListener : ArduinoListener) + notifyArduinoListeners (sourcePlaces : List<Node>, transition : Node, destinationPlacse : List<Node>) + notifyArduinoListenersPhase1 (sourcePlaces : List<Node>, transition : Node) + notifyArduinoListenersPhase2 (transition : Node, destinationPlaces : List<Node>)

+ update (sourcePlaces : List<Node>, transition : Node, destinationPlaces : List<Node>) + updatePhase1 (sourcePlaces : List<Node>, transition : Node) + updatePhase2 (transition : Node, destinationPlaces : List<Node>)

RootPflow # arduinoListeners: ArrayList<ArduinoListener> + getArduinoListeners (): ArrayList<ArduinoListener>

«use»

Marking + registerArduinoListeners (arduinoListener : ArduinoListener) + removeArduinoListener (arduinoListener : ArduinoListener) + notifyArduinoListeners (sourcePlaces : List<Node>, transition : Node, destinationPlacse : List<Node>) + notifyArduinoListenersPhase1 (sourcePlaces : List<Node>, transition : Node) + notifyArduinoListenersPhase2 (transition : Node, destinationPlaces : List<Node>)

ArduinoController + update (sourcePlaces : List<Node>, transition : Node, destinationPlaces : List<Node>) + updatePhase1 (sourcePlaces : List<Node>, transition : Node) + updatePhase2 (transition : Node, destinationPlaces : List<Node>)

«call»

ArduinoComponent # type: ArduinoComponentType # settings: ArduinoComponentSettings # arduinoManager: ArduinoManager + + + +

activate () deactivate () fire () isEnabled (): boolean

Fig. 4. PN2ARDUINO ‐ Class diagram

Fig. 6. PN2ARDUINO ‐ Analog input

Fig. 5. PN2ARDUINO ‐ Adding of Arduino component

Then the behaviour of the system must be de�ined. When the photoresistor detects an excessive lighting (it was experimentally determined as input value gre‑ ater than 799 on the analog pin of Arduino Uno which

resolution is from 0 to 1023) the intermittent tone of the buzzer is turned on. This tone alternates with LED lighting. When the value on the analog pin lowers be‑ low 800, these sound and light effects stop. This is re‑ peated cyclically. Initial marking of modelled timed Petri Net inter‑ preted for control (or sometimes called as interpreted timed Petri net) in PN2ARDUINO is shown in Fig. 8. Articles

23 23

Journal Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME N°33 2020 2020 VOLUME 14,14, N°

p3 - fire alarm is active t4 - alarm is turned off

t2 - alarm makes a noise t5 - alarm is turned off

p1 - alarm does not detect fire

t1 - alarm is turned on

t3 - signal light blinks

p2 - fire alarm is active

Fig. 10. PN for fire alarm (t2 is fired) Tab. 2. Specification of DC motor

Fig. 7. The scheme of laboratory model of fire alarm

p3 - fire alarm is active t4 - alarm is turned off

t2 - alarm makes a noise t5 - alarm is turned off

p1 - alarm does not detect fire

t1 - alarm is turned on

t3 - signal light blinks

p2 - fire alarm is active

Fig. 8. PN for fire alarm (initial marking) Places of Petri net (Fig. 8 ‑ Fig.10) corresponds with these states: ‑ p1 ‑ alarm does not detect �ire

‑ p2 and p3 ‑ alarm is active (�ire was detected) Transitions of Petri net (Fig. 8 ‑ Fig.10) corre‑ sponds with these actions/events: ‑ t1 ‑ alarm is turned on

‑ t2 ‑ alarm makes a noise ‑ t3 ‑ signal light blinks

‑ t4 and t5 ‑ alarm is turned off The token is in place p1 which corresponds with the state when the �ire alarm is not activated be‑ cause the photo‑resistor does not detect light intensity threshold. p3 - fire alarm is active t4 - alarm is turned off

t2 - alarm makes a noise t5 - alarm is turned off

p1 - alarm does not detect fire

t1 - alarm is turned on

t3 - signal light blinks

p2 - fire alarm is active

Fig. 9. PN for fire alarm (t1 is fired)

24 24

At the time when the value greater than 799 is de‑ tected on the analog pin of Arduino ‑ the transition t1 is �ired. This transition is associated with Arduino com‑ ponent Analog Input where a range of input values is set. This range determines when the transition is ena‑ bled. Now the token is in the place p2 (Fig. 9). Transi‑ tion t2 is associated with Arduino component Digital Articles

Actuators conditions Rated voltage Temperature range Humidity range No‑load characteristics No‑load current No‑load speed Load characteristics Rated load Rated current Rated speed Starting torque Locked‑rotor current

6.0V (DC) −20◦ C ~+60◦ C 0% − 90% ≤ 200mA 185 ± 10%rpm 0.0883N.m ≤ 550mA 135 ± 10%rpm 0.4413N.m ≥ 2.0A

Output (in this case pin 8) where the buzzer is con‑ nected. This transition has also associated the function of time delay ‑ 2 seconds. That means that transition �iring (and sound effect of buzzer) lasts for 2 seconds. Now the token is in the place p3 (Fig. 10). Transi‑ tion t3 is associated with Arduino component Digital Output (in this case pin 13) where the build‑in LED is connected. Time delay is set to 1 second. LED diode turns on for 1 second. This process is repeated cyclically, and it is stop‑ ped when the value on the analog pin is lowered under the value 800. Then the transition t4 or t5 is �ired and token moves to the place p1 when �ire alarm does not detect the �ire. We can conclude that the ability of discrete‑event control with PN2ARDU�N� was successfully veri�ied a generalized for other applications.

4. Case Study: Control of Laboratory Hybrid Sy‐ stem For veri�ication of proposed software tool for hy‑ brid systems control, it was necessary to design a la‑ boratory model of such system. A DC motor with en‑ coder was chosen. The encoder is used for feedback in the system because it is used for speed measurement. The actual speed of the DC is in is measured process value. See Table 2 for parameters of described DC mo‑ tor. DC motor was connected to Arduino Uno using the motor shield module. Arduino motor shield is based on dual full bridge driver L298. Using the motor shield, it is possible to independently control speed and mo‑ tion direction of DC motor. The encoder in this motor is of incremental type. For speed measurement, it is

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

necessary to use hardware interruptions functionality of Arduino Uno. The speed of the motor is set by pin described as ”PWM A”. When the input is set to ”PWM = 255” the Ar‑ duino program shows 186 rpm which approximately corresponds with parameters stated by the manufac‑ turer.

From solution analysis, it is obvious that for each working point it is necessary to use a different con‑ troller. One of the solutions is an option to switch be‑ tween multiple controllers according to the working point ‑ speed (rpm) of DC motor. It is possible to use developed software module PN2ARDUINO. It is possi‑ ble to switch between controllers and setpoints using SYSEX messages. Arduino and other microcontrollers that support Firmata protocol can be used. Develop‑ ment and veri�ication of this software module are one of the most interesting results of presented research.

Measurement of steady state I/O characteristics

300

uPWM y [rpm]

u PWM , y [rpm]

250

if analog_input<513

200

150

switch2 setpoint2

setpoint1 100

send SYSEX message 0x00: setpoint = speed_1; set_pid_mode = pid_1;

switch1

send SYSEX message 0x01: setpoint = speed_2; set_pid_mode = pid_2;

0 0

if analog_input>512

t [s]

Fig. 13. Control scheme for hybrid system using PN2ARDUINO

Fig. 11. Measurement of steady state I/O characteristics of DC motor The next step was a measurement of steady state I/O characteristics. The input is a voltage supplied to the motor. These inputs are of size from 0V to 5V which corresponds with PWM signal from 0 to 255 (8‑ bit resolution). Sampling is 0.05 seconds. In Fig. 11 the process of measurement of steady state I/O charac‑ teristics is shown. The signal was �iltered by 1‑D me‑ dian �ilter of 2nd order. Red line is input to the system (voltage or PWM). Output (rpm) is shown by magenta line. Steady state I/O characteristics is in Fig 12. Steady state I/O characterisrics

200 190 180 170 160 150

y [rpm]

140 130 120 110 100 90 80 70 60 50 40 30 20 10 0 0

100

110

120

130

140

uPWM

150

160

170

180

190

200

210

220

230

240

250

260

Fig. 12. Steady state I/O characteristics of DC motor In the process of working points choosing it was necessary to choose points which meet the certain condition. This condition is that behaviour of the sy‑ stem must be close to the linear behaviour around these points. From I/O characteristics two values of in‑ put (uP1 and uP2 ) and output (yP1 and yP2 ) were cho‑ sen (these values will be our working points): uP1 = 80 → yP1 = 140rpm

uP2 = 170 → yP2 = 174rpm

(1)

(2)

For illustration see the scheme in Fig. 13. It is an example for a demonstration of proposed control met‑ hod. Assume the mentioned DC motor. We require to operate it in 2 modes (working points or rpm). For ef‑ fective settlement of speed value to the setpoint, con‑ trollers with different parameters are needed (diffe‑ rent controller for each mode). We switch between rpm using potentiometer connected to the analog in‑ put of microcontroller Arduino Uno. The switching be‑ tween controllers is provided by transitions of Petri net named switch1 and switch2 according to the input value from potentiometer. Input from the analog pin in Arduino is represented by value between 0 and 1023. As the threshold, a half value was used (512). In the moment when the token in Petri net is moved to the place named setpoint1 or setpoint2, a SYSEX message is sent. This message ensures the execution of user de‑ �ined program code on the Arduino side. In this case, the control algorithm is executed. An algorithm (PID controller) for continuous control is independent of Firmata messaging, so it provides real‑time control. The case study of hybrid systems control proposed a basic example. Researchers in the �ield of hybrid con‑ trol design can use it for different and more complica‑ ted scenarios.

5. Conclusion

The article presents the new SW named PN2ARDUINO which extends PNEditor with the capability of communication with microcontrollers that supports protocol named Firmata. Then it is possible to control discrete‑event and hybrid systems using timed interpreted Petri nets with developed SW tool. This tool uses the control paradigm when the microcontroller has implemented only the commu‑ nication protocol. Petri net’s control logic is stored in the personal computer which communicates with Articles

25 25

Journal Journal of ofAutomation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

the microcontroller and sends control orders. The next research will focus on the concept of control with Petri nets where control logic will be directly implemented on the microcontroller.

AUTHORS

Erik Kučera∗ – Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava, Ilkovicova 3, Bratis‑ lava, Slovakia, e‑mail: erik.kucera@stuba.sk, www: www.uamt.fei.stuba.sk. Oto Haffner – Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava, Ilkovicova 3, Bratis‑ lava, Slovakia, e‑mail: oto.haffner@stuba.sk, www: www.uamt.fei.stuba.sk. Roman Leskovský – Faculty of Electrical Engineer‑ ing and Information Technology, Slovak University of Technology in Bratislava, Ilkovicova 3, Bratislava, Slovakia, e‑mail: roman.leskovsky@stuba.sk, www: www.uamt.fei.stuba.sk. ∗

Corresponding author

ACKNOWLEDGEMENTS This work has been supported by the Cultural and Educational Grant Agency of the Ministry of Educa‑ tion, Science, Research and Sport of the Slovak Re‑ public, KEGA 038STU‑4/2018, by the Scienti�ic Grant Agency of the Ministry of Education, Science, Research and Sport of the Slovak Republic under the grant VEGA 1/0819/17, and by the Tatra banka Foundation within the grant program Quality of Education, project No. 2019vs056 (Virtual Training of Production Operators in Industry 4.0).

REFERENCES

[1] M. Dotoli, M. Fanti, and G. Iacobellis, “A freeway traf�ic control model by �irst order hybrid petri nets”. In: 2011 IEEE Conference on Automation Science and Engineering (CASE), 2011, 425–431, 10.1109/CASE.2011.6042526.

[2] M. Fanti, G. Iacobellis, A. Mangini, and W. Uko‑ vich, “Freeway Traf�ic Modeling and Control in a First‑Order Hybrid Petri Net Framework”. In: IEEE Transactions on Automation Science and Engineering, vol. 11, no. 1, 2014, 90–102, 10.1109/TASE.2013.2253606.

[3] M. Dotoli, M. Fanti, and A. Mangini, “Fault moni‑ toring of automated manufacturing systems by �irst order hybrid Petri nets”. In: 2008 IEEE Inter‑ national Conference on Automation Science and Engineering, CASE, 2008, 181–186, 10.1109/CO‑ ASE.2008.4626493.

26 26

[4] N. Costantino, M. Dotoli, M. Falagario, M. P. Fanti, and A. M. Mangini, “A model for sup‑ ply management of agile manufacturing sup‑ ply chains”, International Journal of Production Articles

VOLUME 14, N° N°33 2020 VOLUME 14,

Economics, vol. 135, no. 1, 2012, 451 – 457, 10.1016/j.ijpe.2011.08.021.

[5] H. Matsuno, A. Doi, R. Drath, and S. Miyano, “Ge‑ nomic object net: object oriented representation of biological systems”, Genome Informatics, vol. 11, 2000, 229–230, 10.11234/gi1990.11.229.

[6] M. A. Drighiciu and G. Manolea, “Application des reseaux de petri hybrides a l’etude des systemes de production a haute cadence”. 2010. [7] M.‑A. Drighiciu and D. C. Cismaru, “Modeling a Water Bottling Line Using Petri Nets”, Annals of the University of Craiova, Electrical Engineering series, 2013.

[8] C. Rohr, W. Marwan, and M. Heiner, “Snoopy ‑ a unifying Petri net framework to investi‑ gate biomolecular networks”, Bioinformatics, vol. 26, no. 7, 2010, 974–975, 10.1093/bioinforma‑ tics/btq050. [9] E. Kucera, M. Niznanska, and S. Kozak, “Advanced techniques for modelling of AS/RS systems in au‑ tomotive industry using High‑level Petri nets”. In: 16th International Carpathian Control Confe‑ rence (ICCC), 2015, 261–266, 10.1109/Carpathi‑ anCC.2015.7145085.

[10] E. Kucera, O. Haffner, and S. Kozak, “Modelling and control of AS/RS using Coloured Petri nets”. In: 2016 Cybernetics & Informatics (K&I), 2016, 1–6, 10.1109/CYBERI.2016.7438532. [11] P. J. Mosterman, M. Otter, and H. Elmqvist. “Modeling Petri Nets as Local Constraint Equations for Hybrid Systems Using Mo‑ delica”, 1998. https://www.modelica.org/ publications/papers/scsc98fp.pdf, Acces‑ sed on: 2020.11.30.

[12] S. Fabricius and E. Badreddin, “Modelica library for hybrid simulation of mass �low in process plants”. In: Proceedings of the 2nd International Modelica Conference, Oberpfaffenhofen, Germany, 2002, 225–234. [13] S. Proß and B. Bachmann, “A Petri Net Li‑ brary for Modeling Hybrid Systems in Open‑ Modelica”. In: Proceedings of the 7th Inter‑ national Modelica Conference, 2009, 454–462, 10.3384/ecp09430014.

[14] S. Proß and B. Bachmann, “PNlib‑An Advanced Petri Net Library for Hybrid Process Modeling”. In: Proceedings of the 9th International MO‑ DELICA Conference, 2012, 10.3384/ecp1207647. [15] S. Wolfram, Mathematica: ein System für Mat‑ hematik auf dem Computer, volume 2, Addison‑ Wesley, 1994.

[16] H.‑C. Steiner, “Firmata: Towards Making Micro‑ controllers Act Like Extensions of the Computer”. In: The International Conference on New Inter‑ faces for Musical Expression (NIME), 2009, 125– 130.

Journal of Automation, Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

[17] M. �tter, K.‑E. A� rzé n, and I. Dressler, “StateGraph–A Modelica Library for Hierar‑ chical State Machines”. In: Modelica 2005 Proceedings, 2005, 569–578.

[18] P. Cesek. “DEDS control system based on Pe‑ tri nets and microcontrollers (in slovak)”, 2016. M.Sc. Thesis, Slovak University of Technology in Bratislava, 2016 (in Slovak). [19] M. Riesz, M. Secká r, and G. Juhá s, “PetriFlow: A Petri Net Based Framework for Modelling and Control of �ork�low Processes”. In: ACSD/Petri Nets Workshops, 2010, 191–205.

[20] A. Cesekova. “Control of laboratory discrete event systems”, 2016. M.Sc. Thesis, Slovak Uni‑ versity of Technology in Bratislava, 2016 (in Slo‑ vak). [21] E. Kucera. “Modelling and control of hybrid sys‑ tems using High‑level Petri nets”, 2016. PhD The‑ sis, Slovak University of Technology in Bratislava, 2016, (in Slovak). [22] MIDI Association. “Summary of MIDI 1.0 Messa‑ ges”, 2016. Accessed on: 2020.11.30.

Articles

27 27

Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME N°33 2020 2020 VOLUME 14,14, N°

TWO CASCADED AND EXTENDED KALMAN FILTERS COMBINED WITH SLIDING MODE CONTROL FOR SUSTAINABLE MANAGEMENT OF MARINE FISH STOCKS Submitted: 17th March 2020; accepted: 10th August 2020

Katharina Benz, Claus Rech, Paolo Mercorelli, Oleg Sergiyenko DOI: 10.14313/JAMRIS/3‐2020/30 Abstract: This paper deals with a possible approach to controlling marine fish stocks using the prey‐predator model descri‐ bed by the Lotka‐Volterra equations. The control strategy is conceived using the sliding mode control (SMC) appro‐ ach which, based on the Lyapunov theorem, offers the possibility to track desired functions, thus guaranteeing the stability of the controlled system. One of the most important aspects of this model is the identification of some parameters which characterizes the model. In this work two cascaded and Extended Kalman Filters (EKFs) are proposed to estimate them in order to be utilized in SMC. This approach can be used for sustainable manage‐ ment of marine fish stocks: through the developed algo‐ rithm, the appropriate number of active fishermen and the suitable period for fishing can be determined. Com‐ puter simulations validate the proposed approach. Keywords: Lotka‐Volterra Model, Sliding Mode Control, Extended Kalman Filter

1. Introduction

28 28

Marine ecosystems provide humanity with a mul‑ titude of goods and services, including water quality, �lood control and food supply, all of which are critical for human welfare. Since the human population is gro‑ wing continuously, the demand for these goods and services is also increasing and progressively exerting more pressure on aquatic ecosystems. As many �ish species migrate frequently and the oceans are mostly de�ined as public areas, the de�inition of clear bounda‑ ries and property rights regarding marine resources is rather complicated. As a result, most natural resour‑ ces exploited by the �ishing industry are de�ined as common‑pool resources. This has resulted in many pe‑ lagic ecosystems experiencing high levels of depletion and overexploitation [1], with 46 % of European com‑ munity �ish stocks currently below their minimum bi‑ ological level (European Environment Agency, [2]). The increasing intensity of human �ishing activities in turn diminishes the biodiversity within the affected systems, which is positively correlated with the provi‑ sion of the goods and services of the ecosystem that are of bene�it to the human population, see [3]. Le‑ vels of biodiversity have been shown to determine the stability of marine ecosystems and their ability to re‑ cover. Consequently, Worm et al. suggest that busi‑ ness as usual in the �ishing industry could potenti‑ ally threaten global food security and water quality, as well as ecosystem resilience, and thus jeopardise Articles

present and future generations, see [3]. The obser‑ ved trend is thus of increasing concern, so the topic of the conservation and restoration of aquatic biodiver‑ sity through sustainable �ishery management is incre‑ asingly visible in scienti�ic and political agendas. The United Nations has included this issue in its sustaina‑ ble development goals, dedicating goal number 14 to the conservation and sustainable usage of the planet’s oceans, seas and marine resources, [4]. The success‑ ful implementation of this goal includes the adapta‑ tion of sustainable methods to manage marine and coastal ecosystems in order to avoid signi�icant ad‑ verse effects, which is indicated by the proportion of national economic zones following ecosystem‑based approaches. By 2020, the United Nations aims to re‑ gulate destructive �ishing activities and end over�is‑ hing, alongside implementing a science‑based mana‑ gement approach to restore natural �ish stocks (Uni‑ ted Nations, 2019). In addition, the European Union has conducted several reforms of the Common Fisher‑ ies Policy (CFP), establishing different approaches to attempt to bring the situation under control, with the goal of reaching and maintaining a sustainable level of �ish in the oceans and in �ishermen’s nets. As com‑ mon practice in this �ield, scientists estimate the ex‑ isting level of �ish stocks within an area and suggest a number of total allowable catches (TACs) to politi‑ cal �ishery ministers. In turn, those ministers try to bargain and receive the highest shares for their regi‑ ons, which often leads to the amount of TACs excee‑ ding the maximum level recommended by scientists, rather than levels being allocated for mutual bene�it and optimal conservation purposes. As a result, the methods of the EU are rather unsuccessful for maintai‑ ning a sustainable yield of �ish and achieving the tar‑ gets adopted by all member states of the United Nati‑ ons: as [5] claims, the decision‑making process within the catch allocation should be managed by scientists rather than by politicians. One possible approach to enhancing this decision‑making process and expan‑ ding it based on an independent and objective compo‑ nent, driven by scienti�ic data, is to translate the obser‑ ved ecosystem into a mathematical model using MAT‑ LAB and simulate them with the integrated tool MAT‑ LAB/Simulink. MATLAB is a software package used to describe dynamic systems in a mathematical mo‑ del and can be used to identify the interdependences, mutual interactions, information feedback loops and circular causalities existing in the observed system. This article is an extension of the research presented in [6]. In this work the estimate of some parameters

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

which characterize the model taken into considera‑ tion using Extended Kalman Filters (EKF). Thus, this paper aims to offer a �irst attempt at exploring how MATLAB and Simulink can be utilised to facilitate the implementation of sustainable management approa‑ ches in the �ishing industry through strategic policy testing. The software will be used to formulate a sim‑ ple mathematical description of a marine ecosystem based upon the prey‑predator system represented in the Lotka‑Volterra equations. A number of papers de‑ aling with simulated prey‑predator systems have been published previously; however, adaptation of the mo‑ del to a marine ecosystem including �ish stocks and human �ishers has not yet been covered. In order to si‑ mulate the consequences of various possible policies through different controllers, these have been incor‑ porated into the code to eventually reach and main‑ tain a certain setpoint equal to the maximum sustai‑ nable yield of �ish. In terms of the proposed control technique, sliding mode control (SMC) is taken as one of the �irst possible approaches. In fact, the controllers obtained by an SMC approach show robust properties with respect to parameter uncertainties, as well with respect to more general dynamic uncertainties and to unknown signals. Another application for which SMC has suitable qualities is the �ield of fault‑tolerant cont‑ rol (FTC). In this area, due to intrinsic robustness, SMC models are able to overcome faults and uncertainties. Nevertheless, large uncertainties in the model imply strong chattering effects. Therefore, one of the most important aspects of this approach is the identi�ica‑ tion of some parameters which characterize the mo‑ del. In this work, two cascaded and EKFs are propo‑ sed to estimate them in order to be utilized in SMC. KF is one of the most important and used algorithms in the �ield of identi�ication of states and parameters of a system of any nature. During the last years many dif‑ ferent contributions appeared in many �ields of appli‑ cations and in different technical estimation and iden‑ ti�ication contexts, [7], [8]. Very often, to reduce pro‑ blems of curse of dimensionality KFs are split and or‑ ganized in cascaded forms as for instance in [9]. Just to recall very brie�ly, KF is one of the algorithms using series of the observed measurements over time and it also contains inaccuracies such as statistical noise. Estimates of unknown variables are produced by KF and they are more accurate than the estimates based on the only measurements by estimating a joint pro‑ bability distribution over the variables for each time frame, see [10], [11] and [12]. In fact, the controllers obtained by an SMC approach show robust properties with respect to parameter uncertainties, as well with respect to more general dynamic uncertainties and to unknown signals. Another application for which SMC has suitable qualities is the �ield of fault‑tolerant cont‑ rol (FTC). In this area, due to intrinsic robustness, SMC models are able to overcome faults and uncertainties. Concerning the measurements of the prey, recent re‑ search held at the University Laval and Quebec’s Mi‑ nistry of Forests, see [13], Wildlife and Parks treated the topic of DNA found in lake water which can be

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

used for estimation of the fertility of �ish which live there. This revolutionary approach presented in the Journal of Applied Ecology can contribute to under‑ standing how �ish stocks are managed in lakes. 10 one‑ liter samples of water from different areas of each lake under investigation were analysed by the researchers to be able to estimate the concentration of DNA of the lake trout. The water was �iltered and particles for ge‑ nomic analysis helped to measure the trout DNA in the water samples. A strong correlation between popula‑ tion estimates obtained by means of the traditional ap‑ proach and the one based on the DNA concentration is presented in the results. The paper is organised in the following way. In Section 2 the Lotka‑Volterra mo‑ del is presented. Section 3 is devoted to the control design performed using SMC without and with using EKF. Section 5.1 presents the obtained results and the paper ends with the conclusions drawn.

2. Model Design

The designed model is inspired by the ecological concept of the prey and predator relationship. This concept was formulated by Lotka and Volterra, and is based upon different mathematical theorems.

2.1. Lotka‐Volterra Equations The assumptions of Lotka and Volterra are taken as a basis to describe the relationship between na‑ tural �ish stocks and the �ishing activities of humans. Lotka and Volterra �irst describe the population dyna‑ mics of two species in a prey and predator relationship through two �irst‑order nonlinear differential equati‑ ons, as follows: dx(t) = αx(t) − βx(t)y(t), dt

(1)

dy(t) = δx(t)y(t) − γy(t), (2) dt where x(t) represents the number of prey and y(t) re‑ dy(t) presents the number of predators. dx(t) dt and dt re‑ present the growth rates of the populations based on the respective changes within their population sizes over time, which is denoted by the term t. α, β, δ, γ are positive real parameters and describe the inte‑ raction between the two populations. The expression (1) represents the dynamics of the prey population, which are calculated by subtracting the rate of preda‑ tion from the population’s intrinsic growth rate. Since it is assumed that the prey has an unlimited food sup‑ ply, its population grows exponentially if the popula‑ tion of predators and the rate of predation equal zero, which is expressed by the term αx(t). In turn, the rate of predation upon the prey is assumed to be propor‑ tional to βx(t)y(t). Thus, if either x(t) or y(t) equals zero, there is no predation. Equation (2) describes the dynamics of the predator population, which are determined by the rate at which it consumes the prey population, minus its intrinsic death rate. Since the growth rate of the predator po‑ pulation does not necessarily equal the rate of preda‑ tion of the prey, it is expressed by δx(t)y(t), which is Articles

29 29

Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N°33 2020

similar but not equal to the term representing the rate of predation in Eq. (1). In this equation, γy(t) denotes the loss rate of the predator population due to natural death or emigration. This results in an exponential de‑ cay if there is no prey available to be consumed. Since the main objective of designing this new approach is to achieve and maintain sustainable levels of �ish stocks and harvests alike, an equilibrium point between the two populations is intended. This point is reached if: dx(t) = 0, dt dy(t) = 0. dt

(3)

(4)

As a result, putting the corresponding equations also equal zero, wherefore one has: 0 = αx(t) − βx(t)y(t),

(5)

0 = δx(t)y(t) − γy(t).

(6)

x(t) = 0,

(7)

These equations yield two different solutions. One so‑ lution states that both populations become extinct: y(t) = 0.

�iven the second solution, a �ixed point can be achie‑ ved at which both populations sustain their current non‑zero numbers, depending on the settings of the four parameters α, β, δ, γ. This yields: α , β γ x(t) = . δ

(8)

y(t) =

(9)

Considering the Linearization Lyapunov Theorem it is possible to determine the nature of these two equili‑ brium points. The Jacobian matrix is as follows: [ ] α − βy(t) −βx(t) J= . (10) δy(t) δx(t) − γ At the extinction point (0, 0) the Jacobian matrix beco‑ mes: [ ] α 0 J= , (11) 0 −γ

with the following two eigenvalues λ1 = α and λ2 = −γ. This implies instable equilibrium points. Conside‑ ring the second equilibrium point stated by (9), then J=

[

αδ β

− βγ δ 0

]

(12)

with the following two complex eigenvalues λ1 = √ √ j αγ and λ2 = −j αγ. This implies oscillating point and no conclusion about the nature of this equilibrium point.

3. Sliding Mode Control

30 30

As the goal of the simulation is to realise and esta‑ blish sustainable �ishing activities in order to ensure Articles

the continuity of both marine ecosystems and the hu‑ man species, the current situation of over�ishing and ocean depletion has to be stopped and managed in a way that enables �ish stocks to recover. Therefore, the error between the desired setpoint, being the equili‑ brium point of the �ishery system, and the actual value, represented by the current level of �ish, has to be as‑ certained, harmonised and stabilised. This is explored through application of the Lyapunov Theorem. With zero being the intended value for ẋ(t) = f (x, u, t), the theorem de�ines that if: V (x(t)) > 0, ∀x(t), V (0) = 0, the function is positive and if:

V̇ (x(t)) < 0, ∀x(t)

and one has:

ẋ(t) = f (x, u, t),

(13)

(14) (15) (16)

then x(t) = 0 is an asymptomatic stabile point for function ẋ(t) = f (x, u, t). In order to reduce the error and harmonise the actual value of �ish with the desired value of �ish associated with a sustainable population size, an SMC is used as follows: ∫ t S(t) = (xd (t) − x(t)) + ks (xd (z) − x(z))dz , (17) 0

where ks is a parameter to be designed. Since the V‑ function is a positive‑de�ine function of x(t), it can be employed in the function above. Therefore, one gets: V (S(t)) =

1 2 S (t). 2

(18)

Thereupon, the function is differentiated, which yields: 1 (19) V̇ (S(t)) = 2S(t)Ṡ(t), 2 = S(t) [(ẋd (t) − ẋ(t)) + ks (xd (t) − x(t))] , [

= S(t)

(20)

( ) ] ẋd (t) − αx(t) − βx(t)y(t) + ks (xd (t) − x(t)) , (21)

if: y(t) = yeq (t) =

−ẋd (t) + αx(t) − ks (xd (t) − x(t)) , βx(t)

then V̇ (S(t)) = 0 and if:

y(t) = yeq (t) −

with

ηsgn(S(t)) , βx(t)

  1 if S(t) > 0 0 if S(t) = 0 sgn(S(t)) =  −1 if S(t) < 0,

(22) (23) (24)

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

then, if η > 0:

V̇ (S(t)) = S(t)[−ηsgn(S(t))] = −ηS(t)sgn(S(t)) = −η|S(t)| < 0.

(25)

In order to accelerate the process and reach the desi‑ red value more quickly, term λS(t), with λ > 0, can be included in the equation. The resulting control law is as follows: y(t) = yeq (t) −

ηsgn(S(t)) λS(t) − . βx(t) βx(t)

(26)

Remark 1 It is known that, if ∆ represents the upper bound of the uncertainties of the cancellation through the equivalent part of the control, see (22), then to gua‑ rantee the convergence it is suf�icient to i�pose η > ∆. 3.1. Euler Method Since the system in question has a relatively slow dynamics, it is not intended to measure its state second‑by‑second, but rather on a monthly basis. Therefore, the equation is discretised according to the Forward Euler method, where k represents the known counting integer variable and Ts represents the known sampling time, which yields: x(k) − x(k − 1) ẋ(t) = Ts

= αx(k − 1) − βx(k − 1)y(k − 1) (

→ x(k) = x(k − 1) + Ts αx(k − 1)

) − βx(k − 1)y(k − 1) .

(27) (28)

At this point, the respective equations are integrated into Matlab. With the number of predators and re‑ spectively the number of �ishermen represented by y(t), being the leverage point to control the level of �ish stocks in the regarded aquatic ecosystem, Eq. (26) re‑ presents one of the main equations in the SMC. Since the goal of the applied controller is to harmonise the desired and actual amounts of �ish, measured in kilo‑ gram biomass, the desired amount of �ish (denoted by xd (t)) and the actual amount of �ish (represented by x(t)) are the two main data inputs for the equation. Eq. (26) represents the main equation within the SMC strategy.

4. An Extended Kalman Filter in the Control Loop

The two KFs represeted in Fig. 1 consider the me‑ asured prey x(k) as output measured signal and para‑ meter α(k) and β(k) are the two augmented state to be estimated and y(k − 1) is the number of predators which represents the measured input. The a priori es‑ timation of the augmented state of EKF I is as follows: −

α̂ (k) = α̂(k − 1),

(29)

Fig. 1. Control Scheme and the augmented state of EKF II is as follows: β̂ − (k) = β̂(k − 1).

The a priori predicted covariance matrix is P − (k − 1) = Jd P (k − 1)JdT + Qw ,

(30) (31)

where Qw is the process noise covariance matrix and matrix Jd represents the discrete state Jacobian matrix which is an identity matrix and in our case is represen‑ ted by the scalar for both EKFs: Jd = 1.

Considering that for EKF I which estimates parameter β̂ − (k) = β̂(k − 1) as a stochastic augmented state, the equation of the output prey state is as follows: hI (k) = (1 + Ts α̂(k − 1))x(k − 1)

− β̂(k − 1)Ts x(k − 1)y(k − 1),

(32)

where Ts represents the sampling time and y(k − 1) represents the input predator variable. The output Ja‑ cobian of hI (k) is as follows: HI (k) = −Ts x(k − 1)y(k − 1).

Considering that for EKF II which estimates para‑ meter α̂− (k) = α̂(k − 1) as a stochastic augmented state, the equation of the output is exactly the same as for EKF I: hII (k) = (1 + Ts α̂(k − 1))x(k − 1)

− β̂(k − 1)Ts x(k − 1)y(k − 1)

(33)

Articles

31 31

Journal Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME N°33 2020 2020 VOLUME 14,14, N°

and its output Jacobian is as follows: HII (k) = −Ts x(k − 1).

The following equations state the correction (a poste‑ riori prediction) of for both EKF, EKF I and EKF II: ( )−1 T T H(.) P − (k − 1)H(.) +ζ , K(.) (k) = P − (k − 1)H(.) ( ) α(k) = α̂(k − 1) + K(.) (k) x(k) − h(k) , ( ) β(k) = β̂(k − 1) + K(.) (k) x(k) − h(k) , P(.) (k) = P(.) (k − 1) − K(.) (k)H. P(.) (k − 1),

Number of predators

1200

It is known, that in the presence of uncertainties SMC should provide to switch with a suf�icient large amplitude of η to guarantee the convergence, see (22). In this paragraph the use of EKF is proposed to relax the task of SMC. Tests without considering cancella‑ tion errors are shown at the beginning and after simu‑ lations using EKF in the presence of cancellation er‑ rors are shown.

5.1. Simulation Results Without Using EKF and Without Errors in α and β In order to test the designed model it is assumed that a sustainable level of �ish stocks is reached at a mi‑ nimum of 10.000 kg of �ish. The goal is then to test how the attendance of �ishermen affects the dynamics of the prey population and how a meaningful policy de‑ signed to regulate the activities of the �ishermen could be framed. Figure 2 shows the number of �ishermen in a system that is not restricted by political regula‑ tions. The line graph shows the development of the number of �ishermen over a period of 60 months. In the absence of political regulations, the number of �is‑ hermen immediately increases to 1.000 and remains stable over the entire period of time. The line graph depicted in Fig. 3 shows the corresponding dynamics of the �ish population over a period of 60 months, gi‑ ven the same situation that no political regulation of �ishing activities exists. In this scenario the amount of �ish peaks at 11.000 kg after approximately three months and stabilises at the desired amount of 10,000 kg after 60 months. In order to test how a political re‑ gulation regarding the number of active �ishermen af‑ fects the system, a hypothetical regulation has been assumed demanding that all �ishing activities are pro‑ hibited between the 5th and the 8th month of the pe‑ riod in question. This regulation is realised through an if‑clause in the m‑�ile of Matlab, as follows: if ((T < 5)|(T > 8)) y(t) = yeq (t) −

32 32

ηsgn(S(t)) . βx(t)

(35)

As a result, the number of �ishermen depicted in Fig. 4 rises to 1.000 and remains at that level until it drops Articles

600 400

0 0

30 40 Time [months]

Fig. 2. Number of predators without regulation Biomass of prey [kg]

5. Simulation Results

800

200

(34)

where ζ is the measurement noise covariance variable and K(k) is the Kalman gain and x(k) represents the measured biomass.

1000

12000 11000 10000 9000 8000 7000 6000 5000 0

Time [months] Fig. 3. Biomass of prey without regulation to 0 at the �ive‑month mark. It then remains at 0 until the 8th month and temporarily increases to 1.100 af‑ ter this point. Subsequently, the number slowly decre‑ ases again until it returns to a level of 1.000 after 60 months. The consequences of the regulation regarding the level of �ish stocks in kilogram biomass is depicted in Fig. 5. At the beginning of the time period in que‑ stion, when the number of �ishermen is high, the �ish biomass level is at 10.000 kg. As soon as the regulation takes effect, the �ish biomass increases exponentially, peaking at 1�.500 kg at eight months. Since the �isher‑ men resume their activities from the 8th month on‑ wards, the biomass level decreases again, stabilising at the desired level of 10.000 kg after 60 months. The results show that the designed model is indeed sensi‑ tive to regulatory changes, and that it is able to depict the dynamics of the interdependent populations.

5.2. Simulation Without Using EKF and Including Errors in α and β

If an error is considered in the parameter of α, then it is possible that SMC needs to work with large gain η

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

1.5 Biomass of prey [Kg]

Number of predators

1200 1000 800 600 400 200

105

1 0.5 0 -0.5 -1 -1.5 -2

0 0

Time [months]

-2.5 0

30 40 Time [months]

Fig. 6. Biomass of prey without Kalman estimator and with 20% of the error in parameter α

104

14 12 10 8 6 4 2 0 0

105

2.5 Sliding surface

Biomass of prey [kg]

Fig. 4. Number of predators with regulation

2 1.5 1 0.5 0

Time [months]

-0.5 -1 0

Fig. 5. Biomass of prey with regulation and λ to obtain the same performances. Nevertheless, using large η the chattering phenomenon results to be increased. In Fig. 6, a possible result in term of cont‑ rol is shown in which an error of 20% is considered in the parameter α without increasing the tuning pa‑ rameters of SMC. In Fig. 7, it is visible how the sliding surface does not reach zero.

5.3. Simulation Results Using EKF and Including Errors in α and β

Using the control scheme of Fig. 1 in which an EKF is utilized in the control loop the following results are obtained. Figures 8 and 9 show how the EKF can es‑ timate parameters α and β even with an initial condi‑ tion error on parameters α and β of 20%. Figure 10 shows the result of the controlled biomass inside the

Time [months] Fig. 7. Sliding surface without Kalman estimator and with 20% of the error in parameter α described regulation in which the proposed EKF is uti‑ lised in the control loop. A biased error on parameters α and β of 20% is simulated with error in the initial va‑ lue of biomass. Figure 11 shows the number of preda‑ tors with regulation and using EKF in the control loop. Figure 12 indicates the sliding surface in the presence of the EKF estimation action.

6. Conclusion

Since the implementation of a regulating if‑clause in the m‑�ile yields a reasonable result, the model seems to work and to be appropriate for policy testing Articles

33 33

Journal Journal of ofAutomation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

10-3 Biomass of prey [Kg]

1.5

Estimated parameter Real parameter

Parameter

1.4

1.2

105

0.5

1.1

Time [months]

1.5

Time [months]

Fig. 10. Biomass of prey using EKF in the control loop with an initial error of 20% in parameter α

1500 Number of predators

Estimated parameter Real parameter

1.45

0 0

Fig. 8. Estimation of parameter α using EKF in the control loop with an initial error of 20% in parameter α

1.4 1.35

Parameter

1.5

1.3

1 0

VOLUME 14, N° N°33 2020 VOLUME 14,

1.3

1.25

1000

500

1.2 1.15 0

30 40 Time [months]

Fig. 9. Estimation of parameter β using EKF in the control loop with an initial error of 20% in parameter α

34 34

in the �ishing industry. An algorithm is built in which the biomass of prey is controlled using an SMC stra‑ tegy. To estimate the necessary parameter of the mo‑ del, in this work two cascaded and Extended Kalman Filters (EKFs) are proposed to estimate them in order to be utilized in SMC. However, further research will be necessary in order to construct more complex mo‑ dels, and thus more realistic ones, by including addi‑ tional variables that may in�luence the system. In ad‑ dition, appropriate measurements must be taken and the values within the models must be adapted accor‑ dingly in order to obtain realistic and meaningful re‑ sults. Articles

30 40 Time [months]

Fig. 11. Number of predators using EKF in the control loop with an initial error of 20% in parameter α

AUTHORS

Katharina Benz – Institute of Product and Process In‑ novation, Leuphana University of Lueneburg, Univer‑ sitaetsallee 1, D‑21335 Lueneburg, Germany, e‑mail: katharina.benz@stud.leuphana.de. Claus Rech – Institute of Product and Process Innova‑ tion, Leuphana University of Lueneburg, Universitaet‑ sallee 1, D‑21335 Lueneburg, Germany, e‑mail: claus‑ rech@gmail.com. Paolo Mercorelli∗ – Institute of Product and Pro‑ cess Innovation, Leuphana University of Lueneburg,

Journal of Automation, Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

Sliding surface

[5] B. Leary, J. Smart, F. Neale, J. Hawkins, S. Newman, A. Milman, and C. Roberts, “Fis‑ heries mismanagement”, Marine Pollution Bulletin, vol. 62, no. 12, 2011, 2642–2648, 10.1016/j.marpolbul.2011.09.032.

104

[6] K. Benz, C. Rech, and P. Mercorelli, “Sustaina‑ ble Management of Marine Fish Stocks by Me‑ ans of Sliding Mode Control”. In: 2019 Federa‑ ted Conference on Computer Science and Informa‑ tion Systems (FedCSIS), vol. 18, 2019, 907–910, 10.15439/2019F221.

-5 -10 -15 -20 0

30 40 Time [months]

Fig. 12. Sliding surface with Kalman estimator and with 20% of the error in parameter α Universitaetsallee 1, D‑21335 Lueneburg, Germany, e‑mail: mercorelli@uni.leuphana.de. Oleg Sergiyenko – Applied Physics Department of En‑ gineering Institute of Baja California Autonomous Uni‑ versity, Blvd. Benito Juarez y Calle de La Normal, s/n, Col. Insurgentes Este, C.P.21280, Mexicali, BC, Mexico, e‑mail: srgnk@uabc.edu.mx. ∗

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

Corresponding author

ACKNOWLEDGEMENTS This work was realised within the lectures for the Complementary Studies course at Leuphana Univer‑ sity of Lueneburg during the winter semester 2018‑ 2019.

REFERENCES

[1] H. Gordon, “The economic theory of a common‑ property resource: The �ishery”, The Journal of Political Economy, vol. 62, no. 2, 1954, 124–142, www.jstor.org/stable/1825571.

[7] P. Mercorelli, “A hysteresis hybrid extended Kalman �ilter as an observer for sensorless valve control in camless internal combustion engines”, IEEE Transactions on Industry Ap‑ plications, vol. 48, no. 6, 2012, 1940–1949, 10.1109/TIA.2012.2226193. [8] P. Mercorelli, “A two‑stage augmented exten‑ ded Kalman �ilter as an observer for sensor‑ less valve control in camless internal combus‑ tion engines”, IEEE Transactions on Industrial Electronics, vol. 59, no. 11, 2012, 4236–4247, 10.1109/TIE.2012.2192892. [9] B. Haus, H. Aschemann, and P. Mercorelli, “Tracking control of a piezo‑hydraulic actuator using input‑output linearization and a Cascaded Extended Kalman Filter structure”, Journal of the Franklin Institute, vol. 355, no. 18, 2018, 9298 – 9320, 10.1016/j.jfranklin.2017.07.042.

[10] R. Kalman, “A New Approach to Linear Filte‑ ring and Prediction Problems”, Transactions of the ASME‑Journal of Basic Engineering, vol. 82, 1960, 35–45.

[11] P. S. Maybeck, Stochastic Models, Estimation, and Control, volume 1, Academic Press, Inc., 1979.

[12] F. L. Lewis, Optimal Estimation with an In‑ troduction to Stochastic Control Theory, Wiley‑ Interscience, 1986.

[13] A. Lacoursiè re‑Roussel, G. Cô té , V. Leclerc, and L. Bernatchez, “�uantifying relative �ish abun‑ dance with edna: a promising tool for �isheries management”, Journal of Applied Ecology, 2016.

[2] “EU 2010 biodiversity baseline”. European En‑ vironment Agency (EEA), 2010. Accessed on: 2020.12.02.

[3] B. Worm, E. B. Barbier, N. Beaumont, J. E. Duffy, C. Folke, B. S. Halpern, J. B. C. Jackson, H. K. Lotze, F. Micheli, S. R. Palumbi, E. Sala, K. A. Selkoe, J. J. Stachowicz, and R. Watson, “Impacts of Biodiversity Loss on Ocean Ecosy‑ stem Services”, Science, vol. 314, 2006, 787–790, 10.1126/science.1132294. [4] United Nations. “Sustainable development goal 14”. https://sustainabledevelopment.un. org/sdg14. Accessed on 2020.12.18.

Articles

35 35

Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME N°33 2020 2020 VOLUME 14,14, N°

PROPOSAL OF IOT DEVICES CONTROL USING MIXED REALITY AND QR CODES Submitted: 17th March 2020; accepted: 10th August 2020

Erich Stark, Erik Kučera, Oto Haffner DOI: 10.14313/JAMRIS/3‐2020/31 Abstract: The Internet of Things (IoT) and mixed reality are now among the most important areas in research or in practice. The aim of this paper is to propose an appropri‐ ate way of connection of these two areas, where is pos‐ sible to control and monitor mechatronic devices using a mobile device with augmented/mixed reality support. The main task will be to explore these options in the area and implement this solution as prototype. The proposed methodology for control and diagnostics of mechatronic devices is modern as it combines hardware management, Unity engine for mixed reality development, and commu‐ nication within the IoT network. Keywords: Mechatronic system, Mixed reality, Augmen‐ ted reality, QR code, System control, Cloud

1. Introduction

36 36

Actual trend in the industry is Industry 4.0 ‑ also called as fourth industrial revolution. It is a collective name for current automation, exchanging of data and manufacturing technologies. It can be de�ined as a col‑ lective name for concept and technologies for organi‑ sing value chain, which bonds Cyber‑Physical Systems, Industrial Internet of Things and Internet of Services. Industry 4.0 recognises manufacturing units as com‑ plex distributed systems made by ”smart” partial in‑ tegration of individual autonomous subsystems. The integration is provided by appropriate communica‑ tion of each other based on actual demand, activity coordination and coordination between autonomous subsystems. Communication standard OPC Uni�ied Ar‑ chitecture (OPC UA), virtual and mixed reality belong to the new emerging technologies that are used in ap‑ plications for Industry 4.0. Automation, communications, informatics and ar‑ ti�icial intelligence (A.I.) remain the key disciplines of the 21st century solving many inconveniences and bringing forth all the comforts in life. Industrial manu‑ facturing processes become more and more complex. The marriage of innovative manufacturing processes and advanced automation techniques connected with information and communication technologies, data, and analytics is driving force of the present indus‑ trial revolution. Nowadays manufacturing industry has been facing several important challenges, inclu‑ ding performance of production and sustainability. These challenges are sourced from many real needs and factors such as an aging workforce, changes in the landscape of global manufacturing and adaption Articles

of smart manufacturing by implementing advanced in‑ formation and communication technologies and smart control methods in manufacturing and industrial pro‑ cesses. The Industry 4.0 methodology is a present dri‑ ving force at the core of the industry development, and represents the realization and naming of great chan‑ ges in the present industry. These changes include digitization, mechatronization, automation and infor‑ mation and communication technologies integration at all levels of services and process control. Industry 4.0 represents the 4th industrial revolution in manu‑ facturing industry. The term Industry 4.0 is mainly used in Europe. In China you can �ind the strategy called ”�ade in China 2025”. In USA it is also called as ”Industrial Internet of Things (IIoT)” which highlights one of the main as‑ pects of this revolution. The 4th Industrial Revolution (Industry 4.0) is ba‑ sed on the following paradigms: ‑ Interoperability ‑ is the capability of integration and cooperation of smart machines, intelligent methods and human beings to interact through Internet of Things (IoT), Internet of Services (IoS) and Indus‑ trial Internet of Things (IIoT). ‑ Virtualization ‑ is creation of a virtual model (or copy) of an smart factory. Virtualization uses real data obtained from real plant applied to the smart factory model for decisions and control.

‑ Decentralization ‑ is the capability of each machine to carry out operations and decentralized (autono‑ mous) control, and to make quali�ied smart decisi‑ ons on each subprocess for enhancing process pro‑ duction. ‑ Real time (RT) ‑ data collection and analysis. Smart production control requires data to be collected and examined in real time. Based on the information col‑ lected, real‑time smart control and decision‑making approaches can be used for optimization and re‑ con�iguration, take into account failures and �ind ideal solutions such as component and device er‑ rors, transfer of production, etc. ‑ Service oriented communication and information exchange ‑ over the IoT / IIoT, providing data / in‑ formation / knowledge to other parties of the com‑ pany’s services.

‑ �odularity and recon�igurability ‑ The capability of an smart business to �lexibly adapt to the production situation by changing software and hardware mo‑ dules, module sharing, and recon�iguring processes

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

(multi criterial and multi variant optimal smart de‑ cisions).

Currently, computer networks are no longer just for connecting conventional computers like they once were. Their purpose gained a new dimension when mobile devices and embedded systems began to con‑ nect to these networks. At present, these boundaries are shifted to the level of connection of individual sen‑ sors, various household appliances, and even autono‑ mous cars to the network [4]. This expansion of con‑ nected devices has also happened because of the rise of microcomputers like Raspberry Pi, DragonBoard, and similar prototyping solutions. At the same time, people begin to realize the value of data that these sensors generate. They can help us streamline proces‑ ses in industry and services or make life easier with smart home solutions. As a result, the emergence of new types of networks such as Internet of Things (IoT) are needed. The concept of IoT can be found at almost every conference in the �ield of information and com‑ munication technologies or in scienti�ic articles [6], [3], [8]. The Gartner company makes regular analy‑ zes and research into the use of various technologies. Earlier in 2017, an analysis was made that states that IoT will have up to 20.5 billion connected devices in 2020 [11]. These paradigms would not take place wit‑ hout the development of new networks, data trans‑ mission protocols and the necessary software tools. At present, IoT devices are controlled by console, web, or mobile applications. Using these conventional met‑ hods of controlling IoT devices in a small room can be quite simple. Because the list of devices is on one screen, we can see and set properties almost instantly. But if there are multiple rooms or buildings, the seg‑ mentation of these devices may be totally unclear and cumbersome. Here is the opportunity to use current trends and modern technologies in the �ield of virtual, augmented and mixed reality. These technologies are able to put digital objects into the real world. Their convenience lies in the fact that objects from the real world are enriched with information relevant to the given object that one is looking at. This camera stream processing is real‑time. Mixed reality can now be developed and tracked with compatible headset ‑ such as Microsoft HoloLens or compatible mobile devices (smartphones and ta‑ blets) from both leaders in the segment ‑ Google An‑ droid and Apple iOS. The implementation of the pro‑ posed project involves the use of mobile devices for their wide availability ‑ whether for household or in‑ dustry. Compatible headsets are currently not suita‑ ble for this purpose, as businesses (especially small and medium‑sized ones) are often unwilling to in‑ vest in these headsets. The proposed methodology for controlling and diagnosing IoT devices is modern as it combines hardware management, a 3D engine for mixed reality development, and communication within the Internet of Things network ‑ all areas of me‑ chatronics. The proposed solution is unique and will contribute to the scienti�ic �ield of mechatronics.

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

2. Computer Generated Reality 2.1. Virtual Reality Virtual reality (VR) is a term that is mentioned in various areas, not only in information and commu‑ nication technologies. Films like Matrix have brought virtual reality from the sci‑�i world to the human mind. Examples of virtual and extended reality are becoming more and more real‑life, from military air simulators to simple smartphone applications. Everyone can have their own idea of virtual reality, so it is necessary to in‑ troduce a suitable de�inition. Virtual reality consists of an interactive computer simulation that senses the state of the user and repla‑ ces or extends sensory feedback information to one or more senses in such a way that the user gets a feeling of being immersed in the simulation (virtual environ‑ ment).

2.2. Augmented Reality Augmented reality (AR) is an overlapping of con‑ tent in the real world, but this content is not embedded or part of it. The content of the real world is not capa‑ ble of responding to computer‑generated content [1]. Augmented reality is therefore a live, direct or indirect view of a real world that is complemented by com‑ puter generated (CG) elements such as audio, video, graphics, or GPS data. Augmented reality is a layer of content above the real world, and this content is not anchored to this world or its part. As has been said, elements of the real world and CG content can not re‑ act with each other. The purpose of Augmented reality is to improve user perception and improve its effectiveness through additional information. The user retains awareness of the real world, but in an ideal extended reality it would not be able to recognize the difference between infor‑ mation from the real world and the virtual world.

Fig. 1. Augmented reality example [5] 2.3. Mixed Reality Mixed reality (MR) is an overlap of the real world with synthetic content that is embedded in it and in‑ teracts with the real world. The key feature of MR is that real‑world synthetic content and content can re‑ spond in real time to one another. Mixed reality is thus a combination of the real world and the virtual world, creating a new environment and visualization where physical and digital objects coexist and interact with Articles

37 37

Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

each other in real time. Mixed reality is the layer of ar‑ ti�icial (digital) content in a real world that is anchored and interacts with the real world. An important fact is that, in the case of mixed reality, advanced mapping of the environment is required for the placement of ad‑ ditional CG elements. If information is to be successfully combined, vir‑ tual objects must act physically in a suitable way. If a real and virtual object collision occurs, both must re‑ spond appropriately. In addition, virtual objects must overlay the view of real objects and also shadow on them. All this can only be achieved by a precise model of real and virtual environments.

Fig. 2. Mixed reality example [5] The �irst hardware for mixed reality, currently the most advanced device of the segment, is Microsoft Ho‑ loLens. The problem is still a relatively high price, but there is also an emulator for development.

VOLUME 2020 VOLUME 14,14, N°N°33 2020

with a unique IP address and capable of communica‑ ting with existing network infrastructures.

3.1. Iot Requirements The main IoT requirements are based on the above objectives. An IoT architecture must enable: ‑ data / information / knowledge collection, ‑ data / information / knowledge storage,

‑ data / information / knowledge analysis, ‑ results sharing,

‑ safety. Important requirements include interoperable and ef�icient data transfer and sharing, and therefore the choice of a suitable communication standard and data model. Another requirement for IoT systems is the processing of large volumes of data generated by IoT devices. From the point of view of user, it is also important that the solution is simple to deploy, easy to integrate with other applications and systems, and to achieve a clear organization and presentation of the processed data.

4. Related Work

In the paper [10], the authors presented the Aug‑ mented Things concept, where computer objects con‑ tain all the information needed to track and expand the information required by AR applications. This al‑ lows the user to connect to them, retrieve information using their mobile device, and get expanding informa‑ tion like, for example, maintenance, device, or usage information. The authors have created also a simple 3D framework that allows you to track objects using high quality 3D high resolution scans.

Fig. 3. Mixed reality spectrum [5] Based on the information we have mentioned, mixed reality seems to be the most exciting. However, it is possible to imagine the future in which synthe‑ tic content will be able to react in some way and even communicate with the real world [1].

3. Internet of Things

38 38

Internet of Things is currently a very widespread term in the �ield of modern information and commu‑ nication technologies. This issue is the subject of vari‑ ous debates as its deployment in industry and services brings about more effective action, but it also raises va‑ rious issues, such as safety. Thus, IoT concerns almost all �ields of human activity [9]. In general, IoT can be de�ined as a set of physical objects (or things) embedded in electronics, software, sensors and connected devices that are connected to‑ gether in the network to allow data exchange with other interconnected devices to achieve higher value and more services for users. These IoT devices create a linked network in which each is uniquely identi�iable Articles

Fig. 4. Augmented Things objects contain and share their AR information [10] Phillipe Lewicki has attempted to create a demon‑ stration program to help control the Philips Hue light bulb using the Microsoft HoloLens device as seen in Fig. 5. He realized that today’s solutions allow you to control the bulbs using the mobile application they need to open, to �ind a particular room, and then a par‑ ticular bulb. Often, such applications are limited be‑ cause smart bulbs contain more features than just turn on / off.

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

This system can be decoupled into several compo‑ nents described in subsections below.

Fig. 5. Application for HoloLens, enabling color adjustment of light [10] Thanks to the HoloLens on the head, it was only possible to look at the light and turn it into a simple gesture or change the color of the light. It was faster than a wall switch [5]. Fig. 6 shows Proof of Concept (PoC) by designer Ian Sterling and software engineer Swaroop Pala. Their concept shows how smart devices could be controlled by gestures. The task of this project was to provide a 3D user interface with Android Music Player and Ar‑ duino light fan [2].

Fig. 6. Light control using HoloLens and Arduino devices [5]

5. System Proposal The diagram of the proposed system can be seen in the Fig. 7. System description: 1) At the beginning of the system is mixed reality de‑ vice which is able to analyze data stream from ca‑ mera and detects QR code. 2) Application can connect to identical object in the cloud.

3) Data from the device sensors are sent to the cloud. 4) Mixed reality application gets information about device and shows tailored user interface.

5) The user can interact with that device using mixed reality experience.

6) It is possible to send some control commands to the cloud.

Fig. 7. Diagram of system proposal 5.1. Camera Device With Mixed Reality Support and Computer Vision Tools The best option for mixed reality experiences no‑ wadays is smartphone when comes to price or avai‑ lability for masses. Basically, there are two options: Android system with ARCore SDK or iOS with ARKit SDK support. ARKit SDK was chosen because of gre‑ ater support of functions needed for this project. For example, like persistent content or 3D object recogni‑ tion in recently announced ARKit 2. At �irst it has to be created software which can ana‑ lyze video stream from smartphone camera and de‑ tects physical objects. ARKit can be used with support of framework Vuforia inside Unity 3D editor, which helps a lot with software development. The main fea‑ tures of Vuforia SDK is Multi Target detection, User De‑ �ined Targets or Cloud Recognition. 5.2. Software Platform for IoT OPC is currently the most advanced standardized data exchange process for automation technology. It allows the collection and transmission of data in a uni‑ �ied form from various devices, control systems and applications throughout the organization. The design of this standard allows mapping almost all industrial data into the OPC data structure. OPC UA is an enhan‑ ced version of the OPC standard that has a uni�ied ar‑ chitecture that makes it a platform‑independent pro‑ tocol. In addition, it has built‑in security mechanisms and applications are fully scalable from microcontrol‑ lers to corporate servers. OPC UA organization offers industrial standard OPC UA (Open Platform Communication Uni�ied Architecture) for interoperability and horizon‑ tal and vertical integration of information from sensors/actuators/machines to ERP (Enterprise Resource Planning). Firstly, OPC UA was focused only on the needs of industrial automation, but OPC UA is independent of vendors and operating systems, and it quickly became an appropriate solution for interoperability for other ”markets and domains”. Ma‑ nufacturers’ machines and equipment are ”described” in the OPC UA by data structures and interfaces, and Articles

39 39

Journal Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

then security credentials are con�igured. The various network transport mechanisms are integrated for data and information transfer for the level of operati‑ onal management for both information technologies and cloud solutions, so that the best mechanism is chosen for the different scenarios. OPC UA technology has been recommended for communications techno‑ logy in German Industry 4.0. The German Federal Intelligence Agency (BSI) has thoroughly analysed data security in the OPC UA, resulting in very positive reviews. Many other organizations have used the possibility of modelling and OPC UA con�irmed the continuous integration of information between previ‑ ously incompatible systems. This leads to the removal of barriers to small and medium‑sized enterprises in setting up and expanding industrial communications in the context of Industry 4.0. An example of addressing the issue of uni�ied com‑ munication is also the cooperation of the VDMA Ma‑ chine Vision (Mechanical Engineering Industry Asso‑ ciation) and OPC Foundation. Machine vision systems have become an indispensable part of industrial pro‑ duction.

No other current component in the production process collects such data as machine vision. This re‑ sulted in the socalled ”OPC UA Machine Vision Com‑ panion Speci�ication” to create a standard for commu‑ nication and networking of camera systems within In‑ dustry 4.0. The vision of the OPC Foundation is to offer a multi‑domain platform for interoperability and data exchange from sensors to enterprise systems mana‑ gement. OPC UA is more than just a protocol. OPC UA is rather a framework for the representation and ex‑ change of object‑oriented data and information. 5.3. IoT Prototyping Hardware Kit There were many options for prototyping devices from single hardware to complex IoT kits. It was not easy to �ind IoT kit which meet system re�uirements the most. BigClown is a modular hardware and software sy‑ stem that allows to prototype and build real‑world te‑ lemetry, automation and other applications including IoT. BigClown can be imagined as a set of components with a single interface that can be connected together depends on application needs.

40 40

The core of each device is the so‑called Core Mo‑ dule. It is powered by a single core CPU with the Cortex‑M0�, speci�ically STM32�083C�. This chip was chosen for a number of reasons: it is proven and used ARM CPU from STM32 series, has a very low con‑ sumption (which is important for powering the nodes from battery), has integrated USB with ROM bootloa‑ der, enough number of interfaces, (Flash, RAM and EE‑ PROM), and above all, it has two cryptographic compo‑ nents: TRNG (True Random Number Generator) and AES‑128 computing accelerator [7]. Articles

VOLUME N°33 2020 2020 VOLUME 14,14, N°

Fig. 8. Whole BigClown ecosystem [7]

6. Conclusion The upcoming trend Internet of Things has an im‑ pact not only on applications for various services, hou‑ seholds and intelligent buildings, but also signi�icant impact on industry and industrial production. The ap‑ plication of IoT principles in industry is called In‑ dustrial Internet of Things (IIoT), where in this case instead of interconnected devices is used individual machine parts or their sensors and actuators, as well as sensors and actuators for HVAC (Heating, ventila‑ tion, and air conditioning) security. Device intercon‑ nection should be wireless in particular and should bring new interaction capabilities not only between systems, but also bring new capabilities to control, track and secure advanced services. This proposal of interconnection IoT and mixed re‑ ality can bring new form of Human Machine Interface which can save time for users or companies. The future work will focus on developing such a so‑ lution with technologies mentioned before.

AUTHORS

Erich Stark∗ – Рan‑European Univer‑ sity, Tematinska 10, Bratislava, Slovakia, e‑mail: erich.stark@paneurouni.com, www: www.paneurouni.com/en/kat‑kontakty/peu/. Erik Kučera∗ – Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava, Ilkovicova 3, Bratis‑ lava, Slovakia, e‑mail: erik.kucera@stuba.sk, www: www.uamt.fei.stuba.sk. Oto Haffner – Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava, Ilkovicova 3, Bratis‑ lava, Slovakia, e‑mail: oto.haffner@stuba.sk, www: www.uamt.fei.stuba.sk. ∗

Corresponding author

ACKNOWLEDGEMENTS This work has been supported by the Cultural and Educational Grant Agency of the Ministry of Educa‑ tion, Science, Research and Sport of the Slovak Re‑ public, �EGA 038STU‑4/2018, by the Scienti�ic Grant Agency of the Ministry of Education, Science, Research and Sport of the Slovak Republic under the grant VEGA 1/0819/17, and by the Tatra banka Foundation within the grant program Quality of Education, project No. 2019vs056 (Virtual Training of Production Operators in Industry 4.0).

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

REFERENCES [1] “VR/AR/MR, what’s the difference? | Virtual re‑ ality | Foundry”. www.foundry.com/insights/vr‑ ar‑mr/vr‑mr‑ar‑confused. Accessed on: 2020.12.15. [2] “Control with your smart devices by staring and gesturing”, 2016. blog.arduino.cc/2016/07/26/control‑with‑ your‑smart‑devices‑by‑staring‑and‑gesturing/. Accessed on: 2020.12.15. [3] R. Falkenberg, M. Masoudinejad, M. Buschhoff, A. K. R. Venkatapathy, D. Friesel, M. t. Hompel, O. Spinczyk, and C. Wietfeld, “PhyNetLab: An IoT‑ Based Warehouse Testbed”. In: Annals of Com‑ puter Science and Information Systems, vol. 11, 2017, 1051–1055, ISSN: 2300‑5963.

[4] S. Hammar. “Connected cars: Driving the Internet of Things revolution”, 2017. www.iot‑ now.com/2017/04/03/60270‑connected‑cars‑ driving‑internet‑things‑revolution/. Accessed on: 2020.12.15. [5] P. Lewicki. “Controlling lights with the Hololens and Internet of Things”, 2016. www.afternow.io/controlling‑lights‑with‑ the‑hololens‑and‑internet‑of‑things. Accessed on: 2020.12.15. [6] M. P. Loria, M. Toja, V. Carchiolo, and M. Malgeri, “An ef�icient real‑time architecture for collecting IoT data”. In: 2017 Federated Conference on Com‑ puter Science and Information Systems (FedCSIS), 2017, 1157–1166, 10.15439/2017F381. [7] M. Malý. “BigClown: IoT jako modulá rnı́ sta‑ vebnice”, 2016. www.root.cz/clanky/bigclown‑ iot‑jako‑modularni‑stavebnice. Accessed on: 2020.12.15.

[8] J. Mocnej, T. Lojka, and I. Zolotová , “Using infor‑ mation entropy in smart sensors for decentrali‑ zed data acquisition architecture”. In: 2016 IEEE 14th International Symposium on Applied Ma‑ chine Intelligence and Informatics (SAMI), 2016, 47–50, 10.1109/SAMI.2016.7422980. [9] P. Pohanka. “Internet of Things”. https://pavelpohanka.cz/en/internet‑of‑ things‑2/. Accessed on: 2020.12.15.

[10] J. Rambach, A. Pagani, and D. Stricker, “Aug‑ mented Things: Enhancing AR Applications le‑ veraging the Internet of Things and Univer‑ sal 3D Object Tracking”. In: 2017 IEEE In‑ ternational Symposium on Mixed and Augmen‑ ted Reality (ISMAR‑Adjunct), 2017, 103–108, 10.1109/ISMAR‑Adjunct.2017.42.

[11] R. van der Meulen. “Gartner Says 8.4 Bil‑ lion Connected ”Things” Will Be in Use in 2017, Up 31 Percent From 2016”, 2017. www.gartner.com/en/newsroom/press‑ releases/2017‑02‑07‑gartner‑says‑8‑billion‑ connected‑things‑will‑be‑in‑use‑in‑2017‑ up‑31‑percent‑from‑2016. Accessed on: 2020.12.15.

Articles

41 41

Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME N°33 2020 2020 VOLUME 14,14, N°

ONLINE CONTROL EDUCATION USING 3D HOLOGRAPHIC VISUALISATION Submitted: 17th March 2020; accepted: 10th August 2020

Jakub Matišák, Matej Rábek, Katarína Žáková DOI: 10.14313/JAMRIS/3‐2020/32 Abstract: Interactive 3D visualization technology has brought many benefits into education. Since it is possible to visu‐ alize behavior of wide range of devices, it is much easier to imagine processes where these devices are included. The paper demonstrates the application for interactive teaching of control theory. It allows to simulate a holo‐ graphic model of a selected mechatronic system that is a digital visualization of the real device. The behaviour of the device is controlled by Scilab/Xcos, which is open‐ source, cross‐platform numerical computational environ‐ ment. The main purpose of the application is to help stu‐ dents with better understanding of physical meaning of abstract mathematical models, that describes dynamical systems. Keywords: Control theory, Hologram, Furuta pendulum, Education, 3D model, Scilab

1. Introduction

42 42

Last years, fast improvement of technology, many innovations and digitization of the world have chan‑ ged the way how companies and institutions work. They also adapt and create new methods and proces‑ ses for more effective and innovative educating. They try to make the process for students easier ( [10], [7]) and more ef�icient. The progress in education also cre‑ ates a contribution to science, technology and deve‑ lopment. The ideal outcome is to bring education and research activities together, creating innovations that support the industry [1]. Better education can be ulti‑ mately bene�it for everyone. More effective education can be supported by digi‑ tizing the subject matter. Simplifying device designs, understanding technical speci�ications, facilitating de‑ vice prototyping, or even making manufacturing pro‑ cess cheaper are just a few of the many different uses of 3D hardware digitization ( [19], [3]). There are many three‑dimensional environments around the world that try to incorporate, work, and simulate kno‑ wledge from different areas [12]. Study says that 58% students agreed that methodologies like simulations, demonstrations and virtual laboratories make them more comfortable in lab sessions [14]. Nowadays, we can observe the trend of digitizing in almost every working segment [4]. It allows us to face real situations before they happen, to learn from them, see issues from another perspective, respond to them much faster and, �inally, to save costs. There are many ways how to digitize 3D object. Articles

Lately many institutions have specialized in virtual and augmented reality (AR), like in ( [13], [11]) and brought attention to it. Study [8] says that more than 70% of the students thought that AR made the clas‑ ses more interesting. However, the presented project wants to focus on an area that is not used so wide‑ spread. The aim is to show students another angle of innovative learning using holographic technology. Op‑ tical holography for recording three‑dimensional sce‑ nes can be traced back to the early sixties. Since then, the art of holography has been applied in many areas, primarily as a tool for 3D imaging, processing, and dis‑ play [17]. Study in [6] says that 45.5% of teachers be‑ lieve that hologram technology would have affect in the �ield of teaching. The holographic technology could be used in vari‑ ous areas of life. First example related to its use in cars was published in [18]. The authors attempt to pre‑ sent a holographic display, that should reduce the time when drivers were guided to the dashboard. Hologram should be projected onto the front glass, so time of in‑ attention would be reduced. Another example is me‑ dicine area. The paper published in [16] presented a possibility of displaying real heart beating on a model of heart in four‑sided hologram pyramid. The aim of presented paper is to connect the ho‑ lographic technology and control education. Created application should help students dealing with basics of automatic control. Our system can simply help to vi‑ sualize the behavior of mechatronic system using a 3D digital model in holographic device.

2. State of Technology

The “Hologram” word refers to “a three‑ dimensional picture made by laser light re�lected onto a photographic substance without the use of a camera [2]. Hologram device could be used to play video, represent some system behavior, show object models, etc. We know many varieties of holograms, and there are various ways for classifying them. For our purpose, we can divide them into three types: re�lection holograms, transmission holograms and hybrid holograms.

Reflection holograms The re�lection hologram is the most common used type of the hologram. It can be seen in galleries and in presentation places. Such holo‑ gram is formed when the reference beam and the ob‑ ject beam are incident on opposite sides of the holo‑ graphic surface. They interfere and record an image.

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

To reconstruct the image, a point source of white light illuminates the hologram from the proper angle, and the viewer looks at it from the same side as the light source emits. The setup of re�lection holograms is very simple and holograms are visible without a laser light [5].

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

security hologram stickers, in passports or on cre‑ dit cards. Computer‑generated holograms are used to make optical elements, for scanning, splitting, in ge‑ neral for controlling laser light ( CD player can be an example) [9]. Thanks to its technology this type of ho‑ lograms is not relevant for our work, so we will not pay more attention to it.

3. Holographic Hardware

Fig. 1. Reflection hologram

Transmission holograms Transmission holograms are also known as Laser‑Transmission Holograms. They are created when the reference beam and the object beam are incident on the same side of the holographic surface. They are viewed by shining a spread‑out laser light through the emulsion side of the hologram at the same angle the hologram was recorded at with the viewer looking on from the opposite side. The light is transmitted from behind the hologram device to the side of the observer [5]. Image which is displayed can be very precise. For instance, through a small hologram, a full‑size room with people in it can be seen as if the hologram were a window [9]. The materials, methods, and processes used to make transmission hologram are the same as at re�lection holograms.

Each hologram has to be generated by a holo‑ graphic equipment. In our case we used product of Re‑ al�iction company. The company offers a wide range of devices that differ in sizes and in the number of dis‑ play areas. The price of devices is between (2,000$ ‑ 10,000$). Their main purpose is to present the good in markets. In the paper we used the Real�iction Dreamoc HD3.2 (Fig. 3). The device has three‑sided view, 23” screen and built‑in loudspeakers. Its connectivity is realized via HDMI port and RJ45 port. More details about this equipment can be found in [15].

Fig. 3. Front side of Realfiction Dreamoc HD3.2 The Dreamoc HD3.2 device is big enough to help teacher with 3D model presentation in the lesson. On the other hand its biggest disadvantage is the already mentioned higher price. Fig. 4 shows how the Dreamoc HD3.2 projection works.

Fig. 2. Transmission hologram

Hybrid holograms Hybrid hologram could be consi‑ dered as a combination of transmission and re�lection hologram. Hybrid hologram can be speci�ied as Multi‑ channel holograms, Holographic interferometry, Inte‑ gral holograms, Embossed holograms, and Computer‑ generated holograms. For example, embossed holo‑ grams are used for authenticity applications such as

Fig. 4. Projection in Realfiction Dreamoc HD3.2 The device consists of two main parts. The �irst part is the image‑emitting screen, which is placed on top and emits the image. The Dreamoc HD3.2 uses Full HD screen. The quality of projected image is conditio‑ ned by used resolution. The second part is a projection Articles

43 43

Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

glass with a semi‑permeable layer, which is placed at a 45‑degree angle below the screen. The screen provi‑ des an image that is re�lected on the glass. By placing the glass at the speci�ic angle, the image is presented as if it is behind the glass. It acts as a delusion on the eyes, which creates virtual image of the presented ob‑ ject.

4. Application

Building the application we wanted to ful�il several goals:

‑ to make realistic view of 3D mechatronic experi‑ ment,

VOLUME 2020 VOLUME 14,14, N°N°33 2020

parameters is generated. This form is generic to ens‑ ure the possibility of adding new experiments in the future. The realistic movement in visualization is achie‑ ved by the use of numerical values computed in Sci‑ lab/Xcos simulation environment that is available as a web service. Each experiment is represented by one block schema and computed outputs correspond to changing position of dynamical parts of the simulated mechatronic system. The system architecture is shown in Fig. 6.

‑ to visualize the 3D experiment via holographic de‑ vice, ‑ to control the behaviour of the experiment using pa‑ rameters entered by the user.

Front‐end side of application The application was de‑ veloped with minimal requirements and possibility of future adoption. Using web technologies was proba‑ bly the simplest solution. The main structure of web page is made by HTML, CSS and JavaScript with the use of its libraries. The front‑end side is divided into control and view part. The view uses the Three.js li‑ brary, which allows us to render the 3D model on the screen. It is necessary to realize that the display area is in the hologram, so it is not possible to change the control parameters from there. To do so, we needed to use two browser windows. The �irst one is opened in hologram as the view and the second one in computer screen as the control window. It is allowed via multi‑ screen model (Fig. 5).

Fig. 6. Architecture of the created system Fig. 5. Windows multi‐screen mode

44 44

Back‐end side of application The back‑end of appli‑ cation is driven by Laravel PHP framework, which is mostly used because of its well‑designed object relation mapping implementation. It is connected to a MySQL database, which contains information about the used control algorithm de�ined in the block schema and default input values for experiment con‑ troller. Back‑end is managed by Model‑View‑Controller architecture. When the control interface is opened, data for speci�ic experiment are automatically down‑ loaded from MySQL database and an HTML form with Articles

Web service To facilitate calculation of numerical va‑ lues needed for visualisation, the suitable simulation or computation environment such as Matlab, Scilab or Octave can be used. We decided for Scilab/Xcos. It is an open source distribution modeling and simulation software for numerical computation. There were se‑ veral reasons why it was chosen. Firstly, there is commercial and noncommercial software. Matlab is the most used and professional software, but the license is needed and not everyone can have access to it. Other programs can be included in open source category. However, Octave does not provide a a possibility to build block diagrams suitable for easier creation of controllers. The remaining Sci‑

Journal of Automation, Journal Automation, Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

lab/�cos environment ful�il this requirement and also has appropriate numerical methods to solve differen‑ tial equations. In addition, it is the closest open source option to Matlab/Simulink. Secondly, for the future development it is more suitable to create an application as a modular system. For these purposes it is very useful to have an Appli‑ cation Programming Interface (API) that allows us to use the simulation software as a web service. The ad‑ vantage of the simulation environment that is loca‑ ted on the server is that everyone who requires to re‑ trieve the data needs to know only the URL and how to access the software through the appropriate API. We had API for Scilab at the time of implementation, so therefore we decided to use it. Since the entire data processing module is running independently, it can be easily changed or replaced. Also, application can be ex‑ panded to a different simulation environment in the future. Module with Scilab has its own interface, it works as standalone application. It requires uploading block schema to the server before the �irst simulation. Then, it can be accessed via URL by an authorized program. 3D visualization process The main problem of sho‑ wing results using the 3D model is to ensure how to communicate between the control and the view inter‑ face. During visualization the change of parameters should be possible. Since each client window works separately, it is necessary to inform the view interface about parameter changes in control interface. The application uses simple solution. When the control view is opened, a random SHA code is genera‑ ted and saved as cookie value. Then, the cookie is also used as a �ile name for data sent from simulation. The user view interface gets these data and starts to render the movement of experiment. The animation is gene‑ rated in a loop, so experiment never stops.

5. Furuta Pendulum Experiment

The whole solution is demonstrated on Furuta Pendulum example. In similar way another 3D models can also be realised. Once the view for holographic device is created, it should look like in Fig. 7.

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

The used device needs to have model displayed three times and rotated by ninety degrees, because of its three‑side projection unit. Then, the 3D model on the bottom is shown on the front side of the device and the other two ones on its left and right side. The dark background behind the displayed objects enables to create more realistic movement of the hologram. To start the experiment it is necessary to open the view interface (Fig. 7) in the holographic device. The computer during this process is connected to the screen via the HDMI port. To create a movement of the pendulum, data from simulation are required. In Fig. 8 GUI (Graphical User Interface) of created web service is shown.

Fig. 8. Graphical user interface of the Scilab module One can see there the uploaded block schema con‑ taining a state space controller with prede�ined para‑ meters: ‑ Filename is a �ile name used in Scilab/�cos block schema. The name on the server after uploading is automatically encoded into random hash code. It uses zcos format, the newest standard used in Sci‑ lab 6.0.2. ‑ Input variables are names of variables used in uplo‑ aded zcos �ile. K1, K2, K3, K4 are gains of the used controller. In addition, it is necessary to de�ine the animation time (i.e. length of the simulation), sam‑ pling period and the required angle of pendulum.

‑ Output variables are time, height and speed. Sy‑ stem uses these variables to map output from Scilab console to JSON format. Fig. 7. Furuta pendulum 3D model visualization for Dreamoc HD3.2

Fig. 9 shows the control interface that enables user to enter own parameters for simulation. After sending the request to the server (by clicking on the button Start simulation), the data will arrive Articles

45 45

Journal Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

Fig. 9. Control view interface form for Furuta pendulum using state space controller within seconds and the model simulation will auto‑ matically start. The initial parameters can be altered by sending a new request. The process is designed to change the movement automatically. Fig. 10 shows the holographic device Dreamoc HD3.2 connected to com‑ puter via HDMI.

VOLUME N°33 2020 2020 VOLUME 14,14, N°

led by the prede�ined control algorithm. The applica‑ tion uses Scilab/Xcos as simulation environment for computing data needed for output animation of visu‑ alised model. The presented 3D model was displayed as holo‑ gram. Unfortunately, the holographic device that was used to show the model and behaviour of Furuta pen‑ dulum cannot be considered as a solution that could be used massively due to its higher price. However, the tool can be used as a display device for teachers on lessons. On the other hand, the application can be simply modi�ied to a device that is not subject of a high price and this version could be deployed for an every‑ day use. As a future work authors would like to change the used device for the cheaper one. Also, they would like to extend the application to use it through augmented reality and deploying it on smartphones, using Google ARCore platform.

AUTHORS

Jakub Matišák∗ – Institute of Automotive Mechatro‑ nics, Faculty of Electrical Engineering and Informa‑ tion Technology, Slovak University of Technology in Bratislava, Ilkovič ova 3, Bratislava, Slovakia, e‑mail: ja‑ kub.matisak@stuba.sk, www: uamt.fei.stuba.sk. Matej Rábek – Institute of Automotive Mechatro‑ nics, Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bra‑ tislava, Ilkovič ova 3, Bratislava, Slovakia, e‑mail: ma‑ tej.rabek@stuba.sk, www: uamt.fei.stuba.sk. Katarína Žáková – Institute of Automotive Mechatro‑ nics, Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bra‑ tislava, Ilkovič ova 3, Bratislava, Slovakia, e‑mail: kata‑ rina.zakova@stuba.sk, www: uamt.fei.stuba.sk. ∗

Corresponding author

ACKNOWLEDGEMENTS The work presented has been supported by grants KEGA 030STU‑4/2017 and by the Tatra Banka Foun‑ dation within the grant program E‑talent, project No. 2018et016 (Holographic technology and augmented reality in online experimentation). Authors would like to thank to all colleagues for help with implementation. Fig. 10. Dreamoc HD3.2 displaying Furuta pendulum controlled via control interface

6. Conclusions

46 46

Every educational institution should help students with understanding the subject matter. The presented paper deals with the use of holographic technology in control education. The created application enables to simulate 3D ob‑ jects presenting mechatronic devices that are control‑ Articles

REFERENCES

[1] J. Autiosalo, “Platform for industrial internet and digital twin focused education, research, and in‑ novation: Ilmatar the overhead crane”. In: 2018 IEEE 4th World Forum on Internet of Things (WF‑ IoT), Singapore, 2018, 241–244, 10.1109/WF‑ IoT.2018.8355217. [2] A. H. Awad and F. F. Kharbat, “The �irst design of a smart hologram for teaching”. In: 2018 Advances in Science and Engineering Technology Internati‑ onal Conferences (ASET), Abu Dhabi, 2018, 1–4, 10.1109/ICASET.2018.8376931.

Journal of Automation, Automation, Mobile Journal MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

[3] P. Bistak, “3D Three‑Tank Remote Laboratory Ba‑ sed on Matlab and Websockets”. In: 2019 5th Experiment International Conference (exp.at’19), Funchal (Madeira Island), Portugal, 2019, 85–89, 10.1109/EXPAT.2019.8876585.

[4] B.‑M. Block, “Digitalization in engineering education research and practice”. In: 2018 IEEE Global Engineering Education Confe‑ rence (EDUCON), Tenerife, 2018, 1024–1028, 10.1109/EDUCON.2018.8363342. [5] K. Fey. “Creation of Simple Holograms with a Single Beam Setup ‑ PDF Free Download”, 2000. https://docplayer.net/41350960‑Creation‑ of‑simple‑holograms‑with‑a‑single‑beam‑ setup.html, Accessed on: 2020.12.15.

[6] H. Ghuloum, “3D Hologram Technology in Lear‑ ning Environment”. In: In SITE 2010: Informing Science + IT Education Conference, vol. 10, 2010, 693–704, 10.28945/1283.

[7] T. Haramaki and H. Nishino, “An Engineering Education Support System Using Projection‑ Based AR”. In: 2016 19th International Con‑ ference on Network‑Based Information Systems (NBiS), Ostrava, Czech Republic, 2016, 267–272, 10.1109/NBiS.2016.67.

[8] C.‑Y. Hsu, M.‑W. Chen, and C.‑C. Wu, “Teaching High School Computer Science with Videos of Historical Figures – An Augmented Reality Ap‑ proach”. In: 2015 International Conference on Le‑ arning and Teaching in Computing and Engineer‑ ing, Taipei, Taiwan, 2015, 22–25, 10.1109/La‑ TiCE.2015.30.

[9] A. Jeong and T. H. Jeong. “INTEGRAF | What Are the Main Types of Holograms”, 2019. https://www.integraf.com/resources/articles/a‑ main‑types‑of‑holograms. Accessed on: 2020.12.15.

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

[14] D. Moud, S. Tuli, S. Bijawat, M. Bhardwaj, and S. Sharma, “An analysis to �ind effective teaching methodology in engineering education”. 2013, 185–189, 10.1109/MITE.2013.6756331.

[15] Real�iction. “Dreamoc HD3 Fact Sheet v1”. https://www.realfiction.com/hubfs/ Brochures%20and%20flyers/Dreamoc% 20HD3%20Fact%20Sheet%20v1.pdf. Acces‑ sed on: 2020.12.15.

[16] T. Thap, Y. Nam, H.‑W. Chung, and J. Lee, “Sim‑ pli�ied 3D Hologram Heart Activity Monitoring Using a Smartphone”. In: 2015 9th International Conference on Innovative Mobile and Internet Ser‑ vices in Ubiquitous Computing, Santa Cantarina, Brazil, 2015, 447–451, 10.1109/IMIS.2015.87.

[17] P. W. M. Tsang and T. Poon, “Review on the State‑ of‑the‑Art Technologies for Acquisition and Dis‑ play of Digital Holograms”, IEEE Transactions on Industrial Informatics, vol. 12, no. 3, 2016, 886– 901, 10.1109/TII.2016.2550535. [18] W. Wang, X. Zhu, K. Chan, and P. Tsang, “Digital Holographic System for Automotive Augmented Reality Head‑Up‑Display”. In: 2018 IEEE 27th International Symposium on Industrial Electro‑ nics (ISIE), Cairns, Australia, 2018, 1327–1330, 10.1109/ISIE.2018.8433601. [19] K. Zakova and M. Hok, “Interactive three dimen‑ sional presentation of Segway laboratory model”. In: 2016 International Conference on Emerging eLearning Technologies and Applications (ICETA), Starý Smokovec, High Tatras, Slovakia, 2016, 377–380, 10.1109/ICETA.2016.7802064.

[10] B. Kraut and J. Jeknic, “Improving education ex‑ perience with augmented reality (AR)”. In: 2015 38th International Convention on Informa‑ tion and Communication Technology, Electronics and Microelectronics (MIPRO), Opatija, Croatia, 2015, 755–760, 10.1109/MIPRO.2015.7160372. [11] E. Kuč era, O. Haffner, and R. Leskovský, “Inte‑ ractive and virtual/mixed reality applications for mechatronics education developed in unity en‑ gine”. In: 2018 Cybernetics & Informatics (K&I), Lazy pod Makytou, 2018, 1–5, 10.1109/CY‑ BERI.2018.8337533.

[12] J. Lebiedź and M. Szwoch, “Virtual Sightseeing in Immersive 3D Visualization Lab”. In: Proceedings of the 2016 Federated Conference on Computer Science and Information Systems, 2016, 1641– 1645, 10.15439/2016F227. [13] M. Ma, P. Fallavollita, T. Blum, U. Eck, C. Sandor, S. Weidert, J. Waschke, and N. Navab, “Kinect for Interactive AR Anatomy Learning”. 2013, 277– 278, 10.1109/ISMAR.2013.6671803.

Articles

47 47

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

NÂ° 3

2020

Preface to Special Issue on Modern Intelligent Systems Concepts

DOI: 10.14313/JAMRIS/3-2020/28 The whole world is experiencing an unprecedented technological revolution lately. The latest developments in Artificial Intelligence and Big Data through the advances recorded in the Internet of Things and cognitive computing sector are that we are only in the beginnings of this revolution. All these disciplines find their applications in different fields including health, environment, transport, industry, finance, security, and in general in any field that one can think of. Our role as well as our responsibility in this sector may lie in guiding and directing all these methods to contribute to the well-being of all humanity and not to its destruction. This JAMRIS special issue includes several topics in Intelligent Systems varying from problem modeling to machine learning and deep learning approaches. The included papers are selected from the International Conference on Modern Intelligent Systems Concepts, which was held in Morocco in December 2018 (MISCâ&#x20AC;&#x2122;2018). In this special issue four papers are presented. In the first paper, S. Yousfi, M. Rhanoui and M. Mikram propose to guide researchers to choose between one of the most commonly used models CNN and LSTM. To do so, they compared and applied both models for opinion mining from long text documents using real datasets. They compared the performances of both models using real-world datasets collected from electronic newspapers. They found that combining Doc2vec and CNN models slightly surpasses LSTM performances.

In the second paper, Z. Bakraouy, W. Abbass, A. Baina and M. Bellafkih deal with a problem of designing multi-agent systems with fuzzy characteristics to meet the needs of complex modern systems that must deal with imperfect information. To solve this problem, they proposed a new approach for the design of fuzzy multi-agent systems, the model FMASACNQOS (Fuzzy Multi Agents System for Automatic Classification and Negotiation of QOS). The proposed model consists in integrating logic fuzzy in a multi-agent level, by the use of fuzzy agents independent of other system agents. This model can help build complex applications that can benefit from the advantages of the multi-agent approach and the capabilities of fuzzy logic, such as its ability to represent and manipulate imperfect knowledge. The authors applied their Framework in modeling and implementing a system of classification and negotiation of services in a virtual market that based on the Cloud Computing. As for the third paper, F. Zegrari and A. Idrissi tackle the problem of optimizing and processing large amounts of data in a high heterogeneity system such as the cloud results in a variability of the workload. To ensure the viability of cloud computing, IT resources must be managed effectively by a dynamic monitoring of the current workload of virtual machines (VMs). In their paper, the authors proposed the design of a cloud services simulation tool at the infrastructure level based on cloud computing simulation platform called CloudSim. It allows real-time monitoring of the load of each VM in terms of CPU utilization, memory utilization and bandwidth utilization ratio. The result of this case study can be useful for carry out dynamic environment simulations for VMs monitoring and fast decision making that can be used in load balancing mechanisms. The authors developed a simulator called CloudSimulator based on CloudSim. It enables dynamic and intelligent simulation in the cloud environment and provides monitoring of the current workload of VM resources such as CPU, memory and bandwidth. In this system, the information collected on the current load of resources determines the state of load of each VM that can be used to solve load balancing problems. When an overload is detected on a node, the overloads are transferred to the less loaded nodes. The goal of this study is to manage effectively the Cloud Computing resources to improve the performance of the system.

The last and not the least effort, given by M. Mikram, M. Rhanoui, S. Yousfi and H. Briwa, deals with enhancing the management of unit load devices (ULD) flow. Specifically, they built prediction models: ARIMA following the BOX-JENKINS approach and exponential smoothing methods, in order to obtain more forecasts that are accurate. The authors tested their approach using the operational data of flight processing and the results are compared with four benchmark method (SES, DES, Holt-Winters and Naive prediction) using different performance indicators: MAE, MSE, MAPE, WAPE, RMSE, SMPE. The authors stated that the results obtained with the exponential smoothing methods surpassed the benchmarks by providing forecasts that are more accurate. Following the authors, the sim-

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

NÂ° 3

2020

ple exponential smoothing model is a model that minimizes KPIs and therefore considered as the best performing model for their forecast.

We consider that this special session presents some real advances in the field of Intelligent Systems particularly Artificial Intelligence. It contributes to its evolution, its development, its emergence, and particularly in its orientation and good practice in the service of the humanity. We would like to thank all the authors for their interactions and interesting contributions as well as all the reviewers for their time, advices and suggestions.

In addition, we will not close this preface without warmly acknowledging the great efforts of the Editors, especially Professor Janusz Kacprzyk and the Managing Editor Katarzyna Rzeplinska-Rykala for their great help and support and to any person whom contribute to promote the International Journal JAMRIS. Editor: Abdellah Idrissi MISCâ&#x20AC;&#x2122;2018 General Chair Artificial Intelligence Group, Intelligent Processing Systems Team (IPSS), Computer Science Laboratory (LRI), Computer Science Department, Faculty of Sciences Mohammed V University in Rabat, Morocco email: idrissi@fsr.ac.ma, idrissi@ieee.org

Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

NÂ° 3

2020

Comparative Study of CNN and LSTM for Opinion Mining in Long Text Submitted: 26th June 2019; accepted: 25th March 2020

Siham Yousfi, Maryem Rhanoui, Mounia Mikram

DOI: 10.14313/JAMRIS/3-2020/34 Abstract: The digital revolution has encouraged many companies to set up new strategic and operational mechanisms to supervise the flow of information published about them on the Web. Press coverage analysis is a part of sentiment analysis that allows companies to discover the opinion of the media concerning their activities, products and services. It is an important research area, since it involves the opinion of informed public such as journalists, who may influence the opinion of their readers. However, from an implementation perspective, the analysis of the opinion from media coverage encounters many challenges. In fact, unlike social networks, the Media coverage is a set of large textual documents written in natural language. The training base being huge, it is necessary to adopt large-scale processing techniques like Deep Learning to analyze their content. To guide researchers to choose between one of the most commonly used models CNN and LSTM, we compare and apply both models for opinion mining from long text documents using real datasets. Keywords: Deep learning, Long text opinion mining, CNN, LSTM

1. Introduction

The Web 2.0 has become an official communication space for the press, companies and many governmental or non-governmental organizations. It is also an unofficial communication space, as it allows Internet users to express their ideas, opinions and critics regarding products, services, individuals and special events such as economic or cultural ones. Many organizations are becoming aware of this digital revolution and are implementing new innovative tools to monitor the opinion that the public has built about them and implement, if necessary, preventive or corrective actions. The Press coverage is an essential element to analyze quantitatively and qualitatively the opinion expressed in the traditional and Web Media. They are realized after a media monitoring and consist of the set of documents related to a brand or a product, following a public-relations operation, a press release, a publicity stunt or an event operation.

The Analysis of the press coverage has many advantages. Indeed, it allows measuring the gain or the lack of reputation of an organization and/or its competitors, regarding a particular action, and identifying good and bad actions in order to take preventive or corrective measures. Although sentiment analysis has been widely discussed in the literature, most of the published papers focus on social networks. However, compared to the posts and the comments of the social networks, the press articles are a set of large textual documents. Machine learning techniques that have proven their effectiveness for short texts, lead to poor performances for documents, since the knowledge base becomes wider. Moreover, the accuracy of these techniques is going down because it is more likely that a word appears in long text than in a short text and techniques such as BOW (Bag of Words) show low performances [1], [2]. From another point of view, the Deep learning techniques revolutionized the world of data science during the last years. We wondered the contribution the famous learning models to analyze opinions for long text. Thus, through this paper, we propose to apply and compare both CNN (Convolutional Neural Network), and LSTM (long short-time memory) models for opinion mining from press coverage using real world datasets. The present paper is organized as follows. First, in the section II we provide a general overview about document level sentiment analysis and the deep learning models CNN and LSTM. Then, the section III presents the related works. Sections IV and V present the results and analysis of our benchmark.

2. Background This section briefly describes the general concepts we are using in this paper, namely Document-Level Sentiment Analysis, Word Embedding techniques and Deep Leaning models CNN and LSTM.

2.1. Document-Level Sentiment Analysis

Sentiment mining or opinion mining of textual data is a field of research that attracted the interest of academia and industry during the last decade, especially with the explosion of data through the mas-

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

sive use of social media[2]. Several studies target the building of powerful models to analyze sentiments within different fields such as financial forecasting [3], [4] healthcare[5], [6]and others [7], [8]. Although technically challenging, this field is very useful. Indeed, from one hand, it allows companies to discover the opinion of the public regarding their products, and from the other hand, it helps the users to take advantage from the experience of other customers. There are different levels on which sentiment analysis can be performed according to the level of granularity required [9]: • Word level sentiment analysis that determines the subjectivity, polarity, and strength of orientation of a word. • Sentence level sentiment analysis, which determines the subjectivity or the polarity of a sentence. • Document level sentiment analysis. In fact, at the document level, the objective of sentiment analysis is to assign a global opinion expressed in a document and determine if this opinion is positive or negative. Generally, the whole document is supposed to express the same opinion.

man behavior. This method is based on artificial neural networks. It had a great success in the field of image recognition, natural language processing and speech recognition. The artificial neuronal network represents a set of neurons, each receiving an input value with a certain weight. Then, a combination function allows the comparison of the inputs sum of the neuron. And, finally, an activation function captures the difference and compares it to a certain threshold to choose the output and ensure transmission to other artificial neurons[13]. The activation function is an increasing and differentiable function that takes as a parameter the weighted sum of the “x” entries multiplied by the corresponding weights “Wt”. The most common functions are the sigmoid function, the hyperbolic tangent function (Tanh), the Softmax function and the rectified linear function (ReLU). The sigmoid function allows having an output range between zero and one. It is expressed as follows:

Word embedding is a method that aims to learn the representation of words, by using a vector of real numbers, which facilitates the semantic representation. Two main techniques are used: • Word2Vec is an unsupervised neural network model that produces word embedding according to the words meaning [10]. Similar words are grouped in a vector space, which preserves the semantic relationship between words. • Doc2Vec is an extension of Word2Vec developed by Le and Mikolov [11] that deals with the whole document instead of single words. The model creates a numerical representation of the document in order to determine the meaning of a word and to find similarities between documents.

The hyperbolic tangent function provides an output range of -1 and 1. It is expressed as follow:

2.2. Word Embedding Techniques

2.3. Bag-of-Words (BoW) Model

The bag-of-words (BoW) model considers documents as a bag of words. It is mainly used to generate textual representations in the NLP (Natural Language Processing) and text mining. However, this model ignores the order of the words. Thus, two documents containing the same words are considered as similar. Several techniques based on neural networks have been proposed to generate dense vectors representing both semantic and syntactic properties of words [12].

Sigmoïde (Wt.x) =

Tanh (Wt.x) =

1 (1) 1 + e −Wt . x

eWt . x − e −Wt . x (2) eWt . x + e −Wt . x

The rectified linear function allows having an output with a threshold of 0 when the input is less than 0. It is expressed as follows:

ReLU (Wt.x) = Max (0, Wt.x)

(3)

An artificial neural network is a set of neurons assembled and connected between the layers that constitute it. It is composed of three main layers; an input layer, an output layer, and an intermediate layer, which can be hidden. The following Fig. 1 summarizes the architecture of an artificial neural network:

2.4. Deep Learning Models

Deep learning is a subset of artificial intelligence that uses algorithms to build models that mimic hu-

Fig. 1. A simple example of an artificial neural network[14]

Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

2.5. CNN Model

2.6. LSTM (Long Short-Time Memory) Model

CNN is one of the most successful algorithms used in computer vision. It can detect, segment and recognize objects with excellent noise resistance and variations in position, scale, angle and shape [15]. CNN has also been largely used for NLP tasks such as sentiment analysis, summarization, machine translation, and question answering [16]. The CNN architecture helps to automatically learn the representative characteristics of a given category it receives during the training phase. Subsequently, the CNN seeks these characteristics at the level of a new input data in order to classify it. An example of the architecture of CNN for natural text processing is presented in Fig.2. In fact, it is composed of three different layers namely: input layer, convolution layer and pooling layer.

It is a network of artificial neurons, where the direction of information diffusion is bidirectional, using an internal memory. The LSTM model is based on back propagation over time and allows prediction with sequential data. It is used whenever there is a sequence of data, and the temporal dynamic that connects the data is more important than the spatial content. LSTM is a deep learning model intended for longterm dependencies. It uses an internal memory that allows reading, writing and deleting data, according to their importance. The weights learned by the algorithm determine this level of importance. As represented in Fig. 3, LSTM contains three gates: an Input Gate that allows receiving or not a new entrance, a Forget Gate that ensures the suppression of unnecessary information and an exit door (Output Spoiled). This architecture helps to solve the problem of the disappearance of the gradient the various gates are analog and allow back propagation.

Fig. 2. Example of CNN architecture[17] Input layer. Each word of the input sentences is represented thanks to the techniques of word embedding in vector wi ∈ Rd where d is the dimension of the word embedding. Therefore, the input sentence that contains n words is represented as a matrix with dimensionality d × n. Convolution layer. Convolution is now the most used concept in deep learning. It defines a mathematical operation that takes two input signals (U1 and U2) and returns a new signal (S), where: +∞

S ( t ) = ∫ U 1 (Ʈ).U2(t–Ʈ)dƮ

−∞

(4)

It can therefore be seen as an operation where two sources of information are mixed to produce a result. Convolution is performed over the results of the input layer. It combines input data with a filter K k ∈ Rhd which is applied to a window of h words to produce a new feature, called the convolution kernel. Pooling layer. The pooling layer reduces the inputs by taking only a sample, which helps to minimize the number of parameters and calculations. The most common pooling is the Max-pooling that applies a “max” operation to each region of the filter. Fully connected layer. This layer attributes to each of the extracted “Features” during convolutions a weight representing the connection strength between the same feature and the corresponding category.

Articles

Fig. 3. Detailed description of LSTM architecture [18]

3. Related Works Different deep-learning-based techniques including CNN, and LSTM are compared in papers [12], [19] and [20]. Yin et al. [1] provide a comparative between different Deep learning models including CNN , GRU (Gated Recurrent Unit) and LSTM among a large number of NLP tasks including sentiment classification. The authors conclude that the performances of the different studied models are very close with a slight overshoot for LSTM. The benchmark targeted social network sentences that contain between 5 and 65 words. As discussed above, these benchmarks used for short text analysis are not necessarily adapted to long texts. In our case, we focus our study on large documents.

Journal of Automation, Mobile Robotics and Intelligent Systems

Different researches have applied the Deep Learning models for sentiment analysis in documents; Tang et al. [21] propose a new neural network model called User Product Neural Network (UPNN) to capture user-level and product-level information for opinion analysis on documents. Xu et al. [22] present a Cached Short-Term Neural networks (CLSTM) to capture the overall semantic information in long texts. The proposed method divides memory into memory with a low forgetting rate, that captures the global semantic features, and memory with high forgetting rate captures the local semantic features. Yang et al. [23] proposed a hierarchical attention network for document classification. In order to build the representation of the document, the presented model includes two levels of attention mechanisms; word level and sentence level. In fact, it builds a document vector by aggregating important words into sentence vectors and then aggregating important sentences vectors to document vectors that improves performances. These works are applications of Deep Learning models for short text; they did not provide a comparative study to measure the performance and define the optimal model in the context of the sentiment analysis in documents.

4. Experimental Environment This section aims to present the technical details about the implementation of our benchmark as well as an analysis of results.

4.1. Dataset Description

We have performed our experiments on real datasets collected from web media including electronic newspapers and magazines. The documents were collected from January 2018 to June 2018 in a CSV file with 2275 rows written in French. The CSV file contains the following columns: • Sector: that represents the context of data, such as agriculture, automobile, healthcare, etc. • Brand: the brand concerned with the review. • Media: the name of the journal and magazine that published the data. • Title: the title of the published content. • Text: the published content. • Polarity: it indicates whether the text is positive or negative.

4.2. Technical Environment and Architecture

We have conducted the test on Inter® Xeon® CPU E3-1240 with 3.50GHW, 8,00 GO of memory. The environment is based on the following tools and libraries: • MongoDB: this is the NoSQL database where the input data and the results of the analysis are stored. • Conda: this is the Python-based analysis environment that allows the management of used packages.

VOLUME 14,

N° 3

2020

• PyMongo: this is the python distribution containing tools that connect to MongoDB. • Pandas, NumPy, Sklearn: these are Python libraries that provide tools for operations performed on data during processing. • TensorFlow: this is a library of software dedicated to Deep Learning that provide complex calculations and analyzes. Keras is one of them. • Keras: the API of neural networks.

4.3. Data Preprocessing

Before applying the CNN and LSTM models, we had to preprocess our initial data in order to make them exploitable. First, we have extracted the two attributes useful for the analysis step namely “text” and “polarity”. Then, using beautiful soup library we performed a cleaning step in order to eliminate spaces, html codes, etc. Finally, we divided the dataset into three parts, with 70% for the learning base, 22% for the validation base and 8% for the test database.

4.4. CNN Architecture

We adopted the following CNN architecture: 1) Input layer: As explained before, our dataset contains 2275 rows. The maximum number of words contained in a document is 4500. Therefore, the size of the input matrix is 10237500. Since our documents are very long, we choose the Doc2vec model in order to build the embedding vector of the input data since it offers better performances for documents use cases and allows building a general overview about the document. 2) Convolution: The input matrix is very large we choose to perform many convolution steps in parallel in order to reduce the number of parameters. In order to build our convolution layer we were inspired by [15]. The first convolution step applies 100 bigram filters with kernel = 2. The second convolution applies 100 trigram filters with kernel = 3. The third convolution applies 100 forgrams filters with kernel = 4. The fourth convolution applies 100 fivegrams filters with kernel = 5. 3) Max-pooling: We build each convolution layer followed by a Max-pooling layer. Then the different Max-pooling layers are merged in one output layer. 4) Fully connected Layer 5) The function of activation: Sigmoid This function is used to transform the results obtained and to assign the features to the category designating their polarity, whether they are positive or negative. 6) The loss function “binary cross entropy”. 7) The Adam optimization algorithm. Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

4.5. LSTM Architecture For the LSTM model, that we designed the following architecture: – The Word Embedding (WE) For this step, we define a set of K vectors representative of each polarity by associating with each word belonging to a polarity a vector Xj in a space of dimension d equal to 10237500. – The LSTM layers: They have the particularity of memorizing the chronological order of words, which is beneficial for long sentences. – Fully-connected layer: This layer provides the connection to the output layer by determining the connection weight between the vectors and their category. – The Softmax activation function: This layer allows converting the vector into probabilities on the polarity that we want to detect. The Softmax function provides an output range from 0 to 1, while the sum of all the outputs is equal to one. We choose this function because our model tries to define the category of each input. – The loss function “binary cross entropy”. – The Adam optimization algorithm.

4.6. Results and Analysis

Table 1 shows the experimental results for both CNN and LSTM deep learning models for opinion mining from long textual documents. In fact, Fig.4 and Fig.6 show that the training and validation loss are getting closer between CNN and LSTM with the increase of the number of epochs. Also, as shown in Fig. 5 and Fig.7 both models provide good results with a slight outperformance for CNN with 97% of accuracy during the testing step. Finally, the accuracy of the CNN model improves considerably from the first draft. This is not the case for the LSTM model, since the results improve slowly with the increase of the number of epochs. This is because the CNN implementation uses Doc2Vec model that helped to build the polarity of the whole document. In return, LSTM provides good results, although it’s not combined with the doc2vec model. Indeed, unlike the CNN, LSTM model keeps a memory that allows locating a word in context, which can be similar to doc2vec. Tab. 1. Performance comparison results Model

Training Loss

Doc2vec+CNN 5% LSTM 21%

Validation

Test

Accuracy

Loss

Accuracy

Loss

Accuracy

97% 92%

16% 28%

95% 93%

10% 16%

97% 94%

Fig. 4. CNN loss curve according to the number of epochs 54

Articles

Fig. 5. CNN accuracy curve according to the number of epochs

Fig. 6. LSTM loss curve according to the number of epochs

Fig. 7. LSTM accuracy curve according to the number of epochs

5. Conclusion In this paper, we applied the two famous deep learning models CNN and LSTM for opinion mining from long textual documents. We compared the performances of both models using real-world datasets collected from electronic newspapers. We found that combining Doc2vec and CNN models slightly surpasses LSTM performances. As a perspective, we are currently looking for other models combinations involving Doc2vec, CNN and LSTM in order to improve performances. Acknowledgements. The authors gratefully thank Miss Naqachi Hajar, graduate students from the School of Information Sciences in Rabat for her cooperation and invaluable contribution to the validation of this work.

AUTHORS Siham Yousfi* – SIP Research Team, Rabat IT Center, Mohammed V University in Rabat, Rabat, Morocco, Meridian Team, LYRICA Laboratory, School of Information Sciences, Rabat, Morocco, email: sihamyousfi@research.emi.ac.ma. Maryem Rhanoui – IMS Team, ADMIR Laboratory, Rabat IT Center, ENSIAS, Mohammed V University in Rabat, Morocco, Meridian Team, LYRICA Laboratory, School of Information Sciences, Rabat, Morocco, email: mrhanoui@gmail.com.

Journal of Automation, Mobile Robotics and Intelligent Systems

Mounia Mikram – LRIT, Faculty of Science, Mohammed V University in Rabat, Rabat, Morocco, Meridian Team, LYRICA Laboratory, School of Information Sciences, Rabat, Morocco, email: mikram.mounia@gmail.com. *Corresponding author

REFERENCES [1]

W. Yin, K. Kann, M. Yu and H. Schütze, “Comparative Study of CNN and RNN for Natural Language Processing”, arXiv preprint, 2017, arXiv:1702.01923. [2] N. C. Dang, M. N. Moreno-García and F. De la Prieta, “Sentiment Analysis Based on Deep Learning: A Comparative Study”, Electronics, vol. 9, no. 3, 2020, 10.3390/electronics9030483. [3] S. Sohangir, D. Wang, A. Pomeranets and T M. Khoshgoftaar, “Big Data: Deep Learning for financial sentiment analysis”, Journal of Big Data, vol. 5, no. 1, 2018, 10.1186/s40537-017-0111-6. [4] H. Jangid, S. Singhal, R. R. Shah and R. Zimmermann, “Aspect-Based Financial Sentiment Analysis using Deep Learning”. In: Companion of the The Web Conference 2018, 2018, 1961–1966, 10.1145/3184558.3191827. [5] R. Satapathy, E. Cambria and A. Hussain, Sentiment Analysis in the Bio-Medical Domain, Springer, 2017. [6] A. Rajput, “Chapter 3 – Natural Language Processing, Sentiment Analysis, and Clinical Analytics”. In: M. D. Lytras and A. Sarirete (eds.), Innovation in Health Informatics, 2020, 79–97. [7] M. Kraus and S. Feuerriegel, “Sentiment analysis based on rhetorical structure theory: Learning deep neural networks from discourse trees”, Expert Systems with Applications, vol. 118, 2019, 65–79, 10.1016/j.eswa.2018.10.002. [8] L. Li, T.-T. Goh and D. Jin, “How textual quality of online reviews affect classification performance: a case of deep learning sentiment analysis”, Neural Computing and Applications, vol. 32, no. 9, 2020, 4387–4415, 10.1007/s00521-018-3865-7. [9] M. M. S. Missen, M. Boughanem and G. Cabanac, “Opinion mining: reviewed from word to document level”, Social Network Analysis and Mining, vol. 3, no. 1, 2013, 107–125, 10.1007/s13278-012-0057-9. [10] T. Mikolov, I. Sutskever, K. Chen, G. Corrado and J. Dean, “Distributed Representations of Words and Phrases and their Compositionality”. In: Advances in Neural Information Processing Systems, vol. 26, 2013, 3136–3144, arXiv:1310.4546. [11] Q. Le and T. Mikolov, “Distributed Representations of Sentences and Documents”. In: International Conference on Machine Learning, 2014, 1188–1196.

VOLUME 14,

N° 3

2020

[12] L. Zhang, S. Wang and B. Liu, “Deep learning for sentiment analysis: A survey”, Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 8, no. 4, 2018, 1188–1196, 10.1002/widm.1253. [13] Y. LeCun, Y. Bengio and G. Hinton, “Deep learning”, Nature, vol. 521, no. 7553, 2015, 436–444, 10.1038/nature14539. [14] A. Krenker, J. Bešter and A. Kos, “Introduction to the Artificial Neural Networks”, Artificial Neural Networks – Methodological Advances and Biomedical Applications, IntechOpen, 2011, 10.5772/15751. [15] A. Krizhevsky, I. Sutskever and G. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks”. In: Advances in Neural Information Processing Systems, vol. 25, 2012, 1097–1105. [16] T. Young, D. Hazarika, S. Poria and E. Cambria, “Recent Trends in Deep Learning Based Natural Language Processing”, IEEE Computational Intelligence Magazine, vol. 13, no. 3, 2018, 55–75, 10.1109/MCI.2018.2840738. [17] A. Ibrahim and I. Yasseen, “Using Neural Networks to Predict Secondary Structure for Protein Folding”, Journal of Computer and Communications, vol. 5, 2017, 10.4236/jcc.2017.51001. [18] K. Greff, R. K. Srivastava, J. Koutník, B. R. Steunebrink and J. Schmidhuber, “LSTM: A Search Space Odyssey”, IEEE Transactions on Neural Networks and Learning Systems, vol. 28, no. 10, 2017, 2222–2232, 10.1109/TNNLS.2016.2582924. [19] Q. T. Ain, M. Ali, A. Riaz, A. Noureen, M. Kamran, B. Hayat and A. Rehman, “Sentiment Analysis Using Deep Learning Techniques: A Review”, International Journal of Advanced Computer Science and Applications (IJACSA), vol. 8, no. 6, 2017, 10.14569/IJACSA.2017.080657. [20] P. Singhal and P. Bhattacharyya, “Sentiment Analysis and Deep Learning : A Survey”, Center for Indian Language Technology, Indian Institute of Technology, Bombay, 2016. [21] D. Tang, B. Qin and T. Liu, “Learning Semantic Representations of Users and Products for Document Level Sentiment Classification”. In: Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, 2015, 1014–1023, 10.3115/v1/P15-1098. [22] J. Xu, D. Chen, X. Qiu and X. Huang, “Cached Long Short-Term Memory Neural Networks for Document-Level Sentiment Classification”, arXiv preprint, 2016, arXiv:1610.04989. [23] Z. Yang, D. Yang, C. Dyer, X. He, A. Smola and E. Hovy, “Hierarchical Attention Networks for Document Classification”. In: Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2016, 1480–1489, 10.18653/v1/N16-1174. Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

NÂ° 3

2020

Fuzzy Multi Agent System for Automatic Classification and Negotiation of QOS in Cloud Computing Submitted: 26th June 2019; accepted: 25th March 2020

Zineb Bakraouy, Wissam Abbass, Amine Baina, Mostafa Bellafkih

DOI: 10.14313/JAMRIS/3-2020/35 Abstract: The use of Multi Agents Systems (MAS), Cloud Computing (CC) and Fuzzy Inference System (FIS) in e-commerce has increased in recent years. The purpose of these systems is to enable users of electronic markets to make transactions in the best conditions and to help them in their decisions. The design and implementation is often characterized by the constant manipulation of information, many of which are imperfect. The use of the multi-agent paradigm for the realization of these systems implies the need to integrate mechanisms that take into account the processing of fuzzy information. This makes it necessary to design multi-agent systems (MAS) with fuzzy characteristics. For the modeling and realization of this system, we chose to use the FMAS model. This paper deals with the presentation of the use of the Fuzzy MAS model for the development of a management and decision support application in a virtual market with high availability. After the presentation of the system to be realized in the first section, we describe in the second section the application of the model FMAS for the design and the realization of this system. We then specify the JADE implementation platform and how the fuzzy agents of our model (Expert, Choice and Query) can be implemented using this platform Keywords: MAS, SLA, Negotiation, QOS, Availability, Web Services, Service Broker, Classification, Fuzzy Logic, Inference System, Fuzzy Inference System

1. Introduction

Modern systems such as industry, economics, marketing, and ecology are increasingly based on computer systems, which are responsible for helping users to deal with problems and make decisions [1]. The complexity of the computer systems that need to be put in place leads many researchers to use modeling based on the multi-agent paradigm for the realization of these systems. This approach makes it possible to take into account the complexity of these systems by extracting the entities of actions that must be presented in the modeling of the system. The implementation of modern systems [2] (with the accompanying human-machine communication problems)

is often characterized by the constant manipulation of information, many of which are imprecise, vague, uncertain and incomplete. However, the use of the multi-agent paradigm implies the need to integrate mechanisms that allow it to take into account uncertainty, inaccuracy and incompleteness of information. It is known that fuzzy logic [3] is the only framework in which inaccuracies and uncertainties can be dealt with, which also allow the treatment of certain incompleteness. It is also the only framework in which digital knowledge and knowledge symbolically expressed by natural language skills can be processed. This makes it necessary to design multi-agent systems (MAS) with fuzzy characteristics. Many applications of fuzzy logic that have been performed around the world have been proven effective in solving various types of problems in which the available knowledge is imperfect. Multi-agent technology with its features is the best way to model and build complex distributed systems [4, 5]. The combination of these two technologies should therefore open up a new avenue of research for the design and realization of modern systems that are often complex systems, composed with several entities interacting in distributed mode, and furthermore they are characterized by imprecision and uncertainty. As part of our work, our goal is to take into account the fuzziness in agent-based systems in order to be able to manage applications in which some of the knowledge leading to the decision is imperfect, which is often the case in many applications. Indeed, the purpose of our work is to enrich the MAS enough for their behavior to acquire fuzzy characteristics. This means that the behavior of the system must be able, in particular, to deal with problems with imperfect knowledge, to make decisions (or help decision-makers in their decisions) under uncertainty and with unclear criteria and objectives, and respond to user queries expressed through vague and imprecise language skills. In this paper, we provide an overview of the MAS and Fuzzy Inference System (FIS) approaches adopted in the QoS Classification and Negotiation platform to expose negotiation functionalities to Web services. Service Level Management (SLA)-based negotiation is a crucial support to handle the widely-ranging requirements that characterize Web services. The paper is organized as follows: Section II is dedicated to related work, the Cloud Computing, FIS and MAS

Journal of Automation, Mobile Robotics and Intelligent Systems

concept, in Section III, we present the Fuzzy System Multi Agents for Automatic Classification and Negotiation Framework used in our contribution and we discussed it, eventually in Section VI we draw our conclusions.

2. Materials and Methods 2.1. Motivation In the field of artificial intelligence, attempts are often made to equip artificial agents with techniques of representation and use of knowledge that allow them to solve problems and wait for their objectives. This problem of representation and use of knowledge is at the center of a relatively new and in any case controversial scientific discipline, called artificial intelligence. This discipline has had a limited impact, until recently and exclusively, on the symbolic processing of knowledge, as opposed to the numerical modeling traditionally used in the engineering sciences. More recently, there has been a return to digital in these artificial intelligence problems, with neural networks, genetic algorithms and fuzzy logic. While neural networks offer an implicit “black box” approach and genetic algorithms are iterative optimization algorithms, fuzzy logic is more in keeping with symbolic artificial intelligence, which puts forward the notion of reasoning, and where the knowledge is explicitly coded. The main objective of this section is to highlight the originality of our work and to motivate our choice to use fuzzy logic with MAS in Cloud Computing network.

2.2. Related Work

Cloud Computing. There are many definitions of the term Cloud Computing (CC) and there is little consensus on a single and universal definition. This multitude of definitions reflects the diversity and technological richness of Cloud Computing. In what follows, we cite some of the most relevant. According to [6], based on a close-up view of the Grid computing grids [7], Cloud Computing [8] is mainly based on the paradigm of distributed computing [9] on a large scale to ensure an on-demand service accessible through the Internet. A second definition, proposed in [10], [11] and which is more abstract, defines cloud computing by using the computing resources (hardware and software) that are offered as a service through a network (typically the Internet). A third definition, developed by a working group of the European Commission [12], considers Cloud Computing as an elastic performance environment for resources involving multiple actors to offer a service with a certain level of quality of service. This definition has been extended in [13] taking into account the perspectives of the different players in the Cloud Computing ecosystem (supplier, develop-

VOLUME 14,

N° 3

2020

er, user). However, the definition proposed by the National Institute of Standards and Technology (NIST) in [14], defines Cloud Computing as a model that allows access via a network in a simple and on-demand way to a set of shared and configurable computing resources. These IT resources can be allocated and released quickly with minimal management effort or interaction with service providers. In addition, NIST states that Cloud Computing is composed of five essential features, three service models (Infrastructure As A Service, Platform As A Service and Service As A Service). Multi Agents System. Several definitions of an agent have been given throughout the years when each definition addresses one or more aspects of this paradigm: • According to [15], an agent is a system capable of autonomous action and reflected in a real environment. • According to [16], an agent is a persistent software that has a specific purpose, the agent can be distinguished from a conventional software by its size because smaller and by those objectives and agendas on which it is based to accomplish its tasks. • According to [17], an agent is a computer system that is in a complex and dynamic environment, and that sees and reacts autonomously, in order to achieve the purposes for which it was created. • According to IBM [18], an intelligent agent is a software that performs a set of predetermined tasks, with a certain degree of independence and autonomy, based on a set of knowledge and a representation of predetermined objectives. Agent Features [19]: • Autonomous: the agent is able to act without the influence or intervention of a human or agent and controls his own actions as well as his internal state; • Proactive: the agent must exhibit opportunistic behavior; • Social: the agent must be able to interact with other agents, especially when the situation requires it; • Cooperation: able to coordinate with other agents to achieve the common objective; • Mobility: the agent can be mobile, able to move to another environment; • Rationality: the agent is able to act according to his internal objectives and his knowledge ; • Learning: the agent is able to evolve and learn; As a function of this learning, he is able to change his behavior. A multi-agent system (MAS) [15], [20] is a system composed of a set of autonomous agents, located in a certain environment and interacting according to certain relationships (ACL: Agent Communication Language) to arrive at a global objective. According to Ferber [15], as illustrate Figure1, a multi-agent system is a system composed of the following elements: • An environment E, that is to say a space generally having a metric. Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

• A set of objects O. These objects are located, that is to say that for any object it is possible, at a given moment, to associate a position in E. These objects are passive; That is, they can be perceived, created, destroyed and modified by agents. • A set A of agents, which are particular objects, which represent the active entities of the system. • A set of relations R that unite objects (and therefore agents) between them. • A set of operations Op allowing the agents of A to perceive, produce, consume, transform and manipulate O objects.

Fig. 1. Multi Agents System Fuzzy Inference System. Fuzzy logic is a concept first developed by Lotfi zadeh [21], at the University of California in 1964. A concept that is essentially based on the notion of degrees of belonging to a set according to functions and fuzzy operations. The objective of fuzzy logic is to solve the problems encountered in the use of conventional logic. Problems that are often encountered because of absolute values (true or false) for evaluation of a logical expression. For this, it integrates a new concept of partial truth by assigning different degrees of absolute truth according to a membership function dedicated to the domain studied. Fuzzy logic is used in various fields and research for the control and inference of conclusions based on vague and ambiguous inputs, which require special treatment. To this end, fuzzy logic establishes the link between numerical and symbolic modeling using linguistic variables. This makes it possible to manage numerical uncertainties, based on symbolic values resulting from fuzzy sets. It is on the notion of these that fuzzy logic is essentially based. They support degradation in the membership of elements to a class, such that an element can belong to at least one or more classes.

Fig. 2. Fuzzy Inference System 58

Articles

VOLUME 14,

N° 3

2020

Figure 2 schematizes the steps of the nominal execution of a fuzzy inference system. These steps make it possible to identify the output value needed to derive the state or decision sought. The system entry is generally vague and ambiguous data, so the first phase is to define the membership of the entries, to link with the symbolic variables allowing processing in several steps: • Fuzzification: The first step of the inference system is the conversion of the collections entries by the system to fuzzy sets. These are represented by linguistic variables, whose values are words or phrases of a natural language. The conversion uses the membership functions explained in the previous section. The shape and the end of the functions are found experimentally, since no technique exists for the automatic identification. • Inference: The processing phase that links the input and the output of the system, via rules defined in advance. These are simple conditions that take the form of IF (Condition) THEN. The conditions are a combination of the input variables and the usual operations. The membership of the outputs is identified on the combination of the results of the set of conditions. • Defuzzification: This phase aims to recover the numerical value of the output, via the symbolic value resulting from the treatment phase. An operation that is based on the membership function of the output fuzzy set. There are several techniques for defuzzification, the most used of which is the center of gravity Modern complex systems such as in industry, economics, finance, marketing and ecology are composed with multiple agents interacting in distributed mode and, almost always, they are characterized by inaccuracy, uncertainty and Incompleteness of information. It is clear that fuzzy logic-based multi-agent systems are the most appropriate approach for the analysis, design, and realization of systems with such properties [22]. In the rest of this section, we will present a quick overview on the use of fuzzy logic with MAS in the literature with examples of the work done in this area. Among the first works on fuzzy MAS are those of [23, 24], who introduced the concept of a distributed intelligent multi-agent system (Fuzzy Distributed Multi-Agent). Intelligent System (FDIS)). In particular, they considered the problem of coordination between autonomous agents, namely, fuzzy scheduling, fuzzy scheduling, dispatching, and online dynamic systems. Their model is used for the implementation of a control system in an industrial system. Another approach for designing fuzzy multi-agent systems is that presented in [25]. This model is a fuzzy and multi-criteria decision support system, its main idea is to break down the task between a number of parallel and competitive agents; each intelligent agent composed with a fuzzy knowledge-based system; each agent proposes a solution to the total problem (not only for a partial problem); the total solution of the problem is

Journal of Automation, Mobile Robotics and Intelligent Systems

determined as a proposal of one of the agents after a competition procedure (not with coordination and integration of partial solutions of agents as in the classical distributed system). In [26], the author addresses the problem of group decision making in MAS, a decision in which a collection of n agents must collaborate on the choice of an alternative between a set, X = x1, x2, ... xn, possible alternatives. Based on fuzzy set theory, the author described some methods for helping with this type of decision. In [27, 28], fuzzy linguistic models for multi-agent system design are proposed. These models are offered for cooperative and linguistic filtering of information on the Web. [29] Proposes a method of aggregation of preferences in collective decision problems in the presence of uncertainty. The interest of this article lies in the fact that the decision problem is considered in the form of a possible multi-agent decision (four optimists, and four pessimists) for which a representation theorem is demonstrated. [30] Presents a fuzzy-logic based multi-agent e-commerce system capable of negotiating computer/laptop between the seller and buyer.

3. FMASACNQOS Framework (Fuzzy Multi Agent System for Automatic Classification and Negotiation of Quality of Services) Disciplines have been produced, in order to make the operation of any natural and intelligent machine like the human being. Artificial Intelligence (AI) is one of those disciplines that aims to understand the nature of intelligence by building programs that mimic human intelligence. The classical approach of artificial intelligence, which is related to knowledge-based system design, was not suitable for progressive applications that require a large amount of knowledge and diversity of knowledge. In the same way, individuals usually work in groups. This led to the emergence of a new artificial intelligence approach called Distributed Artificial Intelligence (DAI). This distributed artificial intelligence is based on the principle of the distribution of intelligence between a set of entities that cooperate to achieve a global goal where each of them cannot achieve individually. An extension of the DAI systems is proposed [28], in which the components that have a certain autonomy must be endowed with the capacities of perception and action on their environment. We then talk about agent and therefore multi agent systems Moreover, the implementation of modern systems often poses problems of representation and manipulation of imperfect reversals, this imperfection is due mainly to the nature of the real environment of the world, which cannot always be represented, in a precise quantitative format. Observation instruments that produce errors, and man-machine communication problems that characterize these systems. However, the use of MAS implies the need to equip intelligent agents with reasoning and decision-making skills close to those of the human being, such as

VOLUME 14,

NÂ° 3

2020

the ability to solve problems and make decisions under uncertainty and with inaccurate and incomplete knowledge. This leads researchers to associate fuzzy logic with intelligent agents to establish new kinds of autonomous systems. Fuzzy logic is the most effective way to take into account the vagueness and uncertainty of information, and to formalize processes of reasoning and human decision-making. It is currently generating a general interest among all those who feel the need to formalize empirical methods, to generalize natural reasoning, to automate decision-making in their field, and to create intelligent agents performing the tasks. Usually taken care of by humans. It is for the possibilities it offers to manage uncertainty and inaccuracy, and for its ability to model mechanisms of reasoning and human decision-making, that we are interested in fuzzy logic. The subject of FMASACNQOS case study is the realization of an online decision support system in a virtual market. This system must allow users of the electronic market to carry out their online transactions in a flexible manner and must assist them in their decision. Our system must allow market customers to use linguistic vague terms and imprecise in the criteria for qualifying the data sought and expressing preferences between these criteria, which is often a legitimate demand of the users. For example, the system must allow a user who seeks, via the Internet, a Telecom service, to use fuzzy requests. Thus, the data returned by the system must be ordered and presented to the user according to the QOS of the service provided by the service providers: The provider, the cost and the URL link The objective of FMASACNQOS is to design a system of decision support (Classification and Negotiation of the Service Level Management (SLA)) allowing customers to obtain offers on the services requested, their price and URL. In order to realize this system, we chose to use our fuzzy MAS design model. This paper presents the application of the FMAS model for the desired decision support system modeling. As descripted on the Figure 3 the Fuzzy Multi Agent System for Automatic classification and negotiation of QOS is based on the inference system to make the decision:

Fig. 3. FMASACNQOS Generally, QoS [31], [32] is defined as degree of satisfaction of clients to use the service. The achievement of this satisfaction will be achieved when the QoS metrics for various network applications based on technology or user factor are respected. Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

Both of these elements play an important role to satisfy the desired requirement of QoS. The Service Level Agreements (SLA) is a compromise between user and provider of services by a document that defines the QoS, the prescribed service. In other words, these are clauses based on a contract defining the precise objectives expected and the level of service a client wishes to obtain from the provider and sets out the responsibilities. The mechanism of negotiation, monitoring and classification of SLAs provokes many issues like, congestion in providers side, increasing delay, low availability The way SLA between cloud service providers and cloud service consumers are established and managed is currently far from being ideal from the customer’s point of view because of the huge number of requests which cause saturation of the buffers in consequence congestion of networks and Unavailability of services. This topic naturally gains crucial importance for customers being companies whose success depends, even partially, on the advertised QoS. The overall aim is the development of an intermediary that generates SLA management tasks. The introduction of a Broker of Agreement facilitates the mechanisms of service discovery and automates negotiation and monitoring of SLAs, moreover it allows comparing the services according to their QoS requirements as mentioned on Table 1. Furthermore, clients can negotiate SLAs based on the QoS requirements outlined in the table below. Tab. 1. Requirements of general QoS [29] Service

Data Rate

Delay

Loss Rate

FTP

11,8 Kbps

64 Kbps

< 150 ms

< 0,1%

4-60 Mbps

< 150 ms

< 0,0001%

VoIP

Video Email

< 10 Kbps

~ 10 sec < 4 sec

Articles

N° 3

2020

plication concerned. Our choice was on the Java programming language and the MAS: JADE development platform for the following reasons: • It is easy to make a mapping of fuzzy agents from our model in a JADE code, • It is simple to create agents with JADE, • JADE manages communication between agents and offers agent management interfaces, • The JADE platform respond to several features and offer a wide range of libraries, • The agents developed in JADE are written totally in JAVA, which is an easy language and based on the notion of object. The Framework used in FMASACNQOS to implement fuzzy inference system is JFuzzyLogic It present many advantages by facilitate and speed up the development of fuzzy systems. The main advantages are: • Using standard programming language (FCL), • Providing a fully functional and complete implementation of FIS, • Creating API that can be extend by developers, • Implementing an Eclipse plugin to easily write and test FCL code, • Independency of the software platform. As is shown below in Figure 4 every provider must make registration by sending its offers. Every offers must contain: • Id of provider, • Name of Service, • Data Rate, • Loss Rate, • Link, • Cost.

The best way to build a multi-agent system (MAS) is to use a multi-agent platform. A multi-agent platform is a set of tools needed to build and commission agents in a specific environment. These tools can also be used for analysis and testing of the MAS created. These tools can be in the form of a programming environment (API) and applications to help the developer. There are currently several platforms for the development of multi-agent systems, such as Madkit [33], Zeus [34], Swarme [35] and JADE [36], etc. However, these platforms do not offer a solution to facilitate the use of interaction protocols, with the exception of the JADE platform. The realization of our application, on the case study, which concerns the realization of a management application and decision support in a market, requires us to choose implementation tools. The development stage requires technological tools adapted to the field of research and the architecture of the ap-

VOLUME 14,

Fig. 4. Service Provider Container

Fig. 5. Execution of deployment of Service Provider

Journal of Automation, Mobile Robotics and Intelligent Systems

Fig. 6. Membership functions for LossRate

VOLUME 14,

NÂ° 3

2020

Fig. 5 presents the execution of deployment of a service provider by its container. The membership functions of Qos parameters (Loss rate, Data Rate and Delay) are illustrated in the following figures. Fig. 6 defines the membership of Loss Rate. The value, called membership value or degree of membership, quantifies the grade of membership of the element to the fuzzy set. Fig. 7 defines the membership of Data Rate. The membership function characterizes the degree of certainty and truth in FIS. Fig. 8 defines the membership of Delay. Fig. 9 describes the negotiation between the broker and the providers of service using the sniffer of JADE framework. As it mentioned above, we have use the library FuzzyLogic to implement the fuzzy inference system (FIS) in our framework. Table II shows example of fuzzy control language code executed in our framework. Tab. 2. Example OF Fuzzy Control Language (FCL) Code

Fig. 7. Membership functions for dataRate

RULE 1 : IF dataRate IS excellent AND delay IS average AND lossRate IS average THEN qos IS poor; RULE 2 : IF dataRate IS average AND delay IS excellent AND lossRate IS average THEN qos IS excellent; RULE 3 : IF dataRate IS average AND delay IS average AND lossRate IS excellent THEN qos IS excellent; RULE 4 : IF dataRate IS average AND delay IS average AND lossRate IS average THEN qos IS average;

Fig. 8. Membership functions for Delay

The use of decision support systems in e-commerce has increased in recent years. The purpose of these systems is to enable users of electronic markets to make transactions in the best conditions and

Fig. 9. Exchanged messages in the network Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

to help them in their decisions. The design and implementation of decision support systems, with the accompanying man-machine communications problems, is often characterized by the constant manipulation of information, many of which are imperfect. The subject of our case study is the realization of an online decision support system in a virtual market (Cloud Computing). This system must allow users of the electronic market to carry out their online transactions in a flexible manner and must assist them in their decision. Our system must allow market customers to use vague and imprecise language terms in the criteria for qualifying the data sought and to express preferences between these criteria, which is often a legitimate request from users. Thus, the data returned by the system must be ordered and presented to the user according to preferences. On the other hand, our system must help service providers estimate their property at the right price. The use of the multi-agent paradigm for the realization of these systems implies the need to integrate mechanisms that take into account the processing of fuzzy information. This makes it necessary to design multi-agent systems (MAS) with fuzzy characteristics. For the modelling and realization of this system we chose to use the FMAS model

we used the JADE platform. In this case, we proposed rules that allow to easily implement the fuzzy agents from the FMASACNQOS framework, as well as the different modules. Portions of codes are presented to clarify certain operations

4. Conclusion

REFERENCES

In our work, we are interested in the problem of designing multi-agent systems with fuzzy characteristics to meet the needs of complex modern systems that must deal with imperfect information. To solve this problem, we have proposed in this paper a new approach for the design of fuzzy multi-agent systems, the model FMASACNQOS (Fuzzy Multi Agents System for Automatic Classification and Negotiation of QOS). In our contribution, the proposed model is initially generic since we found, in the literature, the absence of a generic model for the design of fuzzy MAS. And unlike the work currently being done on fuzzy multi-agent systems, which offer architectures for specific applications or use logic to deal with problems in MAS our design model is based on the idea of integrating logic fuzzy in a multi-agent level, by the use of fuzzy agents independent of other system agents, it is therefore independent of the application in which we will use it. Indeed, our model will help build complex applications that can benefit from the advantages of the multi-agent approach and the capabilities of fuzzy logic, such as its ability to represent and manipulate imperfect knowledge. In this paper, we presented the application of our Framework, the FMAS model, in a case study, which consists of modeling and implementing a system of classification and negotiation of services in a virtual market that is the Cloud. Computing. After the system modeling using our framework with the soft focus, we presented the execution scenarios with an example on the use. And for the system implementation,

[1]

Articles

AUTHORS Zineb Bakraouy* – STRS Lab., National Institute of Posts and Telecommunications, Rabat, Morocco, email: bakraouy@inpt.ac.ma.

Wissam Abbas – STRS Lab., National Institute of Posts and Telecommunications, Rabat, Morocco, email: abbass@inpt.ac.ma. Amine Baina – STRS Lab., National Institute of Posts and Telecommunications, Rabat, Morocco, email: baina@inpt.ac.ma.

Mustafa Bellafkih – STRS Lab., National Institute of Posts and Telecommunications, Rabat, Morocco, email: bellafkih@inpt.ac.ma. *Corresponding author

[2]

[3]

[4]

[5]

E. Marcon, S. Chaabane, Y. Sallez, T. Bonte and D. Trentesaux, “A multi-agent system based on reactive decision rules for solving the caregiver routing problem in home health care”, Simulation Modelling Practice and Theory, vol. 74, 2017, 134–151, 10.1016/j.simpat.2017.03.006. W. Abbass, Z. Bakraouy, A. Baina and M. Bellafkih, “Classifying IoT security risks using Deep Learning algorithms”. In: 2018 6th International Conference on Wireless Networks and Mobile Communications (WINCOM), 2018, 10.1109/WINCOM.2018.8629709. A. V. Lovato, C. H. Fontes, M. Embiruçu and R. Kalid, “A fuzzy modeling approach to optimize control and decision making in conflict management in air traffic control”, Computers & Industrial Engineering, vol. 115, 2018, 167– 189, 10.1016/j.cie.2017.11.008. S. Torres, O. M. Salazar and D. A. Ovalle, “A Fuzzy-Based Multi-agent Model for Group Formation in Collaborative Learning Environments”. In: P. Vittorini, R. Gennari, T. Di Mascio, S. Rodríguez, F. De la Prieta, C. Ramos and R. Azambuja Silveira (eds.), Methodologies and Intelligent Systems for Technology Enhanced Learning, vol. 617, 2017, 3-11, 10.1007/978-3-319-60819-8_1. Z. Bakraouy, A. Baina and M. Bellafkih, “System Multi Agents for Automatic Negotiation of SLA

Journal of Automation, Mobile Robotics and Intelligent Systems

[6]

[7]

[8] [9] [10] [11]

[12] [13] [14]

[15]

[16]

[17]

in Cloud Computing”. In: A. Abraham, A. Haqiq, A. K. Muda and N. Gandhi (eds.), Innovations in Bio-Inspired Computing and Applications, vol. 735, 2018, 234–244, 10.1007/978-3-319-76354-5_21. L. M. Vaquero, L. Rodero-Merino, J. Caceres and M. Lindner, “A break in the clouds: towards a cloud definition”, ACM SIGCOMM Computer Communication Review, vol. 39, no. 1, 2008, 50–55, 10.1145/1496091.1496100. R. Buyya, C. S. Yeo, S. Venugopal, J. Broberg and I. Brandic, “Cloud computing and emerging IT platforms: Vision, hype, and reality for delivering computing as the 5th utility”, Future Generation Computer Systems, vol. 25, no. 6, 2009, 599–616, 10.1016/j.future.2008.12.001. C. N. Höfer and G. Karagiannis, “Cloud computing services: taxonomy and comparison”, Journal of Internet Services and Applications, vol. 2, no. 2, 2011, 81–94, 10.1007/s13174-011-0027-x. I. Foster, Y. Zhao, I. Raicu and S. Lu, “Cloud Computing and Grid Computing 360-Degree Compared”. In: 2008 Grid Computing Environments Workshop, 2008, 1-10, 10.1109/GCE.2008.4738445. D. S. Kushwaha and A. Maurya, “Cloud Computing - A Tool For Future”, International Journal of Mathematics and Computer Research, vol. 1, no. 1, 2013, 9-14. A. Mohiuddin, M. R. C. Abu Sina, A. Mustaq and H. R. Md. Mahmudul, “An Advanced Survey on Cloud Computing and State-of-the-art Research Issues”, IJCSI International Journal of Computer Science Issues, vol. 9, no. 1, 2012, 201–207. L. Schubert, K. Jeffery and B. Neidecker-Lutz, “The future of cloud computing: Opportunities for European cloud computing beyond 2010”, Expert Group report, public version, 2010. “Advances in clouds: Research in future cloud computing”. L. Schubert and K. Jeffery, https:// ec.europa.eu/newsroom/dae/document.cfm? doc_id=1174. Accessed on: 2020.12.16. P. Mell and T. Grance, “The NIST Definition of Cloud Computing”, Technical Report, DOI: 10.6028/NIST.SP.800-145. https://csrc.nist. gov/publications/detail/sp/800-145/final. Accessed on: 2020.12.16. J. Ferber, Les systèmes multi-agents. Vers une intelligence collective, Dunod, 1997. J. C. Brustoloni, Autonomous agents: characterization and requirements, Carnegie Mellon Technical Report CMU-CS-91-204, School of Computer Science, Carnegie Mellon University, 1991. P. Maes, “Agents that Reduce work and information Overload”, Readings in Human–Computer Interaction, 1995, 811–821, 10.1016/B978-0-08-051574-8.50084-4

VOLUME 14,

N° 3

2020

[18] D. Gilbert, “Intelligent Agents: The Right Information at the Right Time”. In: IBM Corporation, IBM Intelligent Agent White Paper, 1998. [19] J.-P. Briot and Y. Demazeau, Principes et architecture des systèmes multi-agents, Hermes Science Publications, 2001. [20] D. El Bourakadi, A. Yahyaouy, J. Boumhidi, “Multi-agent system based on the fuzzy control and extreme learning machine for intelligent management in hybrid energy system”. In: 2017 Intelligent Systems and Computer Vision (ISCV), 2017, 10.1109/ISACV.2017.8054922. [21] L. A. Zadeh, “The concept of a linguistic variable and its application to approximate reasoning”, Information Sciences, vol. 8, no. 3, 1975, 199–249, 10.1016/0020-0255(75)90036-5. [22] R. A. Aliev, B. Fazlollahi and R. R. Aliev, Soft Computing and its Applications in Business and Economics, Springer Berlin Heidelberg, 2004, 10.1007/978-3-540-44429-9. [23] R. A. Aliev and R. R. Aliev, Fuzzy Distributed Intelligent System for Continuous Production in Applications of Fuzzy Logic: Towards High Machine Intelligence Quotient Systems, Prentice Hall, 1997. [24] R. A. Aliev and R. R. Aliev, “Fuzzy Distributed Intelligent Manufacturing System”. In: First European Congress on Fuzzy and Intelligent Technologie, Aachen, Germany, 1999, 229–235. [25] B. Fazlollahi, R. M. Vahidov and R. A. Aliev, “Multi-agent distributed intelligent system based on fuzzy decision making”, International Journal of Intelligent Systems, vol. 15, no. 9, 2000, 849–858, 10.1002/1098-111X(200009)15:9<849::AIDINT2>3.0.CO;2-I. [26] R. R. Yager, “Penalizing strategic preference manipulation in multi-agent decision making”, IEEE Transactions on Fuzzy Systems, vol. 9, no. 3, 2001, 393–403, 10.1109/91.928736. [27] E. Herrera-Viedma, F. Herrera, L. Martı́nez, J. C. Herrera and A. G. López, “Incorporating filtering techniques in a fuzzy linguistic multiagent model for information gathering on the web”, Fuzzy Sets and Systems, vol. 148, no. 1, 2004, 61–83, 10.1016/j.fss.2004.03.006. [28] E. Herrera-Viedma, E. Peis and J. M. Moralesdel-Castillo, “A Fuzzy Linguistic Multi-agent Model Based on Semantic Web Technologies and User Profiles”. In: E. Herrera-Viedma, G. Pasi and F. Crestani (eds.), Soft Computing in Web Information Retrieval, vol. 197, 2006, 105–120, 10.1007/3-540-31590-X_6. [29] N. Ben Amor, F. Essghaier and H. Fargier, “Décision collective sous incertitude possibiliste. Principes et axiomatisation”, Revue d’intelligence artificielle, vol. 29, 2015, 515–542, 10.3166/ria.29.515-542. Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

[30] B. M. Balachandran and M. Mohammadian, “Development of a Fuzzy-based Multi-agent System for E-commerce Settings”, Procedia Computer Science, vol. 60, 2015, 593–602, 10.1016/j.procs.2015.08.186. [31] Y. Chen, T. Farley and N. Ye, “QoS Requirements of Network Applications on the Internet”, Information-Knowledge-Systems Management, vol. 4, no. 1, 2004, 55–76, 10.1145/1030194.1015475. [32] “MaDKit”. The Multiagent Development Kit, http://www.madkit.org/. Accessed on: 2020.12.16. [33] Z. Bakraouy, W. Abbass, A. Baina and M. Bellafkih, “MAS for Services Availability in Cloud of Things Network: Monitoring and Reactivity”. In: Proceedings of the 12th International Conference on Intelligent Systems: Theories and Applications, 2018, 10.1145/3289402.3289510. [34] B. Azvine, D. D. Nauck and N. Azarmi (eds.), Intelligent Systems and Soft Computing: Prospects, Tools and Applications, Springer, 2000, 10.1007/10720181. [35] M. Askenazi, R. Burkhart, Ch. Langton, N. Minar, “The Swarm Simulation System: A Toolkit for Building Multi-Agent Simulations”. https:// www.santafe.edu/research/results/workingpapers/the-swarm-simulation-system-a-toolkit-for-building. Accessed on: 2020.12.16. [36] F. Bellifemine, G. Caire, T. Trucco and G. Rimassa, “JADE Programmer’s Guide”. https://jade. tilab.com/doc/programmersguide.pdf. Accessed on: 2020.12.16.

Articles

VOLUME 14,

N° 3

2020

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

NÂ° 3

2020

Modeling of a Dynamic and Intelligent Simulator at the Infrastructure Level of Cloud Services Submitted: 26th June 2019; accepted: 25th March 2020

Faouzia Zegrari, Abdellah Idrissi

DOI: 10.14313/JAMRIS/3-2020/36 Abstract: Cloud environments made up of a large number of compute and storage servers provide ondemand services in a usage-based consumption model (pay-as-you-go). Load balancing is one of the major problems in the cloud. Indeed, the dynamics of demand requirements and QoS, as well as the variability of cloud resources and its provisioning models make difficult the operation of performance evaluation of the system. To face this issue and to ensure the viability of cloud computing, IT resources must be managed effectively by a dynamic monitoring of the current workload of virtual machines (VMs). In this study, we propose the design of a cloud services simulation tool at the infrastructure level based on cloud computing simulation platform named CloudSim. It allows real-time monitoring of a load of each VM in terms of CPU utilization, memory utilization and bandwidth utilization ratio. The result of this case study can be useful for carry out dynamic environment simulations for VMs monitoring and fast decision making that can be used in load balancing mechanisms. Keywords: Load balancing, Cloud computing, Resource utilization, Dynamic environment simulation

1. Introduction The explosion of numerical data and the need for high availability of service are the critical factors in the emergence of the concept of cloud computing. The cloud model is a new paradigm in IT aiming at modernizing the Internet. It allows access to a pool of computing resources that can be allocated and released on demand with minimal interaction with the service provider [1]. The cloud provides hosted services in Datacenters of high performance, which are categorized according to the technical layer provided. There are three usage models at the disposal [2]: IaaS, PaaS and SaaS. [3] The IaaS layer corresponds to the architecture and IT infrastructure part where the provider hosts virtualized resources like servers, data storage, network and virtualization. The variability of cloud resources, the diversity of requirements for applications in terms of performance and workload are among the most important problems in the field of re-

search. The dynamic of demands can be managed by dynamically provisioning cloud resources capacities. Several researchers, who integrate state of load control techniques of resources in a cloud datacenter, have proposed various dynamic scheduling algorithms. The simulation of these algorithms in real time for the evaluation of the performances of various metrics is a very difficult job to realize. In our study, we propose a dynamic simulator based on CloudSim Framework [4, 5, 6]. Let us recall that CloudSim allows a modeling and a correct simulation of the infrastructure and application services of the cloud computing. It is an open source framework developed in Java. The communication between the various CloudSim entities, such as Datacenters, hosts, virtual machines VMs and cloudlets, occurs using events with static triggering. Indeed, after launching the simulation, CloudSim does not make it possible to interact with the system or to add tasks dynamically. Tasks are assigned to VMs through the Datacenter broker before the beginning of the execution. That is explained by the fact why the Cloudsim clock is based on events for its operation. Each execution, resumption or stop execution is defined as being an event. At the beginning of its execution, it predicts, based on the information of simulation, the duration of simulation as well as the successions of the events, which will take place. We are not thus able to obtain a load of VMs in real time. There exists in the literature some tools, that measure the use of resources, but these tools are not easy to adapt them in certain contexts. The rest of this paper is organized as follows: the next section presents the analysis of some frameworks used in previous work for the measurement of workload and the possibility to adapt them in dynamic environment simulations. In section 3, we propose the architecture and modeling of the simulation. The CloudSimulator simulation scenario is described in Section 4. Section 5 is devoted to evaluating and analyzing the results obtained by CloudSimulator. The conclusion of this paper is presented in Section 6 with perspectives for future related work.

2. Related Work In our study, several tracks have been explored to analysis some frameworks used in previous work in measuring the load of VMs and their possibility

Journal of Automation, Mobile Robotics and Intelligent Systems

to adapt them according to our context to perform simulations in real time. The first track was to analyze the CloudSim distributions that were developed in order to find one that would be able to meet our request. Two distributions have attracted our attention: Dynamic CloudSim and Real Time CloudSim [6], but none of these distributions overcomes our problem. As a second solution, the CloudSched [7]. This tool compares existing simulation systems at the application level for the cloud and defines a new lightweight simulation system for dynamic resource scheduling in cloud datacenters. The results are then analyzed and discussed. The goal of CloudSchedâ&#x20AC;&#x2122;s analysis is to understand the logic of real-time simulation in order to adapt it to our solution. Since the source code is not available, we were able to access the compiled code and rewrote the source code. After analysis, we noticed that it randomly generates the tasks and executes them without defining a real allocation policy. In addition, the defined classes do not allow us to reuse them in order to implement our solution since several elements are not defined, like policies of supply and allocation at the level of the hosts and VMs. The third possible solution was to install a virtual machine hypervisor, free virtualization software, which will be responsible for the management of VMs. The principle of the hypervisor is to run the operating system in the same kernel and not emulate them, which allows keeping performances close to the native ones. There are several hypervisors and we chose XEN [8] which is (para) virtualization software. This distribution integrates a XenMon monitoring application [9] to monitor a Xen-based environment. It allows running several operating systems on the same hardware resource (PC, Server...), but consumes many resources and today, it is considered hypervisor completely overstepped. Similarly, for the VMWare hypervisor, the physical server on which the VMs are hosted must have a high memory capacity to be able to share it between the VMs and it is a software, which requires the purchase of a license for its use. Since no explored solution was conclusive, we then opted for the development of a simulator called CloudSimulator, adapted to our need to determine in a more flexible way, performance metrics in terms of use of CPU, ram and bw. Taking into account its success, we decided to develop our simulator around CloudSim to take advantage of the plethora of algorithms and basic models, which it proposes. The design and implementation of this proposed load measurement tool first required an analysis of the architecture of CloudSim and its various modes of operation and then a stage of understanding its source code in order to be able to redefine some classes and methods of the CloudSim simulator. The use of this tool in load balancing mechanisms represents a definite advance in real-time monitoring of cloud resource status and load balancing decision-making. 66

Articles

VOLUME 14,

NÂ° 3

2020

3. Architecture and Simulation Modeling The operation of the simulator proposed is not based on events in order to make the dynamic simulation. Its architecture is described by the whole of the components as illustrated below, in Fig. 1. Datacenter: it manages and groups hundreds of physical machines connected to each other and characterized by physical resources such as mips, ram, bandwidth and storage, also logical specifications like architecture, hypervisor Vmm, operating system, time zone and pricing at the second of the various resources used making it possible to bill the cost of consumption to its customers. It implements resource allocation policies for hosts and VMs. Host: This class models a physical node assigned to one or more VMs by a VMs allocation policy named VmScheduller. A host is characterized by CPU speed (mips), storage capacity; one or more physical processing cores Pes, bandwidth and memory capacity. Datacenter Broker: the broker allows access to Datacenter and plays the role of mediator. It manages the execution of virtual machines and acts on behalf of the cloud provider. Virtual machine: This class models a VM, which is managed by a host and allows running cloudlets according to the scheduling policy that it uses. The elements characterizing a VM are CPU capacity, memory capacity, the number of CPUs, bandwidth and storage size. VmAllocationPolicy [4]: This abstract class represents provisioning policies for allocating hosts to virtual machines with the least Pes used.

Fig. 1. Class Diagram of the CloudSimulator Simulation System VmScheduler [10]: This abstract class models how to distribute the available processing capacity of each core of a host between the VMs that host them, according to the chosen allocation policy: time-shared or space shared. For proposed CloudSimulator, we used VmSchedulerTimeShared because it allows Pes sharing. CloudletScheduler [10]: This class establishes the virtual Pes allocation policy for a given VM to run cloudlets. It implements two strategies: spaceshared and time-shared. We have created a third

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

type to adapt it to the dynamic environment, which we called DynamicCloudletSpaceShared based on the principle of shared space type. Evaluating the effect of these two policies on cloudlet completion time was described in [11] and proved that space shared policy has been better than time-shared policy. Cloudlet: This class represents the task to run on a VM, defined by a number of instructions and a quantity of data to be transferred, expressed by a size of the input and output files. As for the allocation of resources, such a scheduling of VMs [10] uses policies at the host level and the VM level: the first policy relates the sharing of host cores between the VMs assigned to it. The second policy allows the VMs to allocate from the available capacity, a quantity of mips to the tasks for their executions. Two modes of execution are suggested: timeshared and space-shared. [4] In the space-shared policy, the cloudlet is executed once the resource is released, the other cloudlets are queued. The completion time of the cloudlet depends on the number of Pes necessary for its execution and the capacity of the assigned processing element. In time-shared, a scheduler is used to allocate resources to a cloudlet during a certain time, once the usage time has elapsed, the cloudlet is queued to execute the next entity from the queue waiting, and however, running the cloudlets happens almost at the same time. The resource utilization of cpu, ram, and bw by a Cloudlet can be determined according to the usage model chosen. The models proposed by CloudSim are UtilizationModelFull, UtilizationModelNull and UtilizationModelStochastic. For our simulator, we choose the stochastic model, which consists in assigning to each Cloudlet for its execution on a VM, a percentage of random use between 0 and 1. A load is specified as a rate of utilization resource, associated with a duration of use. In our case, since time is variable, we associated it with the length of the task (expressed in Millions of instructions MI) which is a fixed value. When the simulation is launched, we want to collect in real time information on the load rate of each VM. The utilization rate of the resource is the average value of utilization rates for each cloudlet in the execution list. The total utilization rate on a VM is expressed by the following equation:

  = UseVm %  Use cpu + Use ram + Use bw  × 100 . (1)   3 The execution time of cloudlet in real time is calculated by using the following formula:

Execution = finishTime − startTime time

(2)

where finishTime is the time where this Cloudlet completes and startTime is the start time of executing this Cloudlet.

N° 3

2020

The communication latency of the node can be easily deduced by:

= Wait time startTime − submissionTime (3)

where Waittime is the communication latency of the node and submissionTime is Cloudlet’s submission time to a Cloud Resource. The response time is calculated as hereafter:

Response = Executiontime + Delaytime time

Knowing that:

(4)

Delay = Wait time + Transfertime (5) time

where DelayTime is the transmission time of the cloudlet and ExchangeData is the size of data exchanged between cloudlets that communicate with this cloudlet. The workload of each VM is determined from the formula defined in (1) and can be categorized by three states: under loaded, normal and overloaded.

4. Simulation Scenario • • • • •

Initialize the simulator clock; Create a Datacenter and a Datacenter broker; Generate VMs assuming nodes are heterogeneous; Generate cloudlets dynamically; Submit the VMs and Cloudlets created to the broker; • Start the simulation; • Save and display the monitoring of the execution of each VM: – The measurement of the load of CPU, ram and bw load as described in (1); – The execution time as defined in (2); – Simulation start time, Start time of executing cloudlet and time where this Cloudlet completes; – The number of cloudlets processed; • Stop the simulation. Tab. 1. Host configuration N° of Host

Processing Power (MIPS)

6 000

RAM (MB)

10 000

Storage (MO)

1000000

Bw (Mb/s)

10 000

Tab. 2. Virtual Machines configuration Virtual machines

VM0

RAM (MB)

256

P.Power (MIPS)

250

VM1

1 500 1 024

VM2

VM3

2 048

1 024

1 000

500

Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

Tab. 3. Cloudlets configuration Number of Cloudlets

1 100

Length (MI)

40 000 Ă 140000

Output Size (KO)

300

File Size (KO)

300

5. Experiments and Evaluation of Results In order to evaluate the behavior and the efficiency of our proposed simulator, we present the experiments and the evaluations of the results of the simulation that we undertook along this study. The simulation was carried out on a PC with an Intel Core i5 CPU 2.40GHz, 32-bit Windows 8.2 Professional Operating System, 4GB Ram, Development Environment: NetBeans IDE 8.2, Cloudsim-3.0.3 Framework and JAVA development language. The experiment consists of measuring the workload on each VM in real time.

5.1. Experimentation

CloudSimulator allows us to perform multiple tasks on multiple Vms. The simulation is carried out with a single host created in a single Datacenter. The goal is to perform an experiment for various quantities of tasks, which we add dynamically, and then measure a load of CPU, memory and bandwidth on each VM. The scenario begins with a simulation configuration step: We proceed by setting the parameters of the various components of the simulator: a Datacenter made up of 4 virtual machines whose capacities are respectively 250, 1500, 1000 and 500 MIPS as shown on Tab. 2. These VMs are instantiated on a host whose configuration is set on Tab. 1. The tasks to be performed on the VMs are set to an interval [40 000, 140 000] Million instructions, as shown on Tab. 3. In our experiment, we suppose that the tasks communicate with each other by the exchange of a quantity of data, to send or to receive. After this initialization phase, we proceeded to the application of the simulation scenario as described above. The program that we have developed uses a multithreaded environment. Among the classes that have threads: EntitiesGenerator.java, DynamicVm.java DynamicDatacenterBroker.java. EntitiesGenerator.java is a class that allows creating tasks dynamically, to assign them to the appropriate VMs, and to send them to the Datacenter broker. When launching the simulation, the thread of the DynamicDatacenterBroker class performs four operations: the first operation starts the thread of each VM, which triggers the updateVmProcessing(vmInfo) method of the DynamicCloudletScheduler class, for updating the task state (in execution, finished) and updating the processing time. The second operation sends the cloudlets received on the broker to the waiting list of the corresponding VM. The third operation calls on with the method, which will collect on each VM, the 68

Articles

VOLUME 14,

NÂ° 3

2020

information of the load of the CPU, ram and Bw and also the processing time of the accomplished tasks. Lastly, the fourth operation executes the method that ends the simulation. The results of simulation are shown in the curves below.

5.2. Results Evaluation

The real-time monitoring performed throughout the simulation on each VM showed the resource utilization that varies according to the size of running cloudlets expressed in million instructions. The measured values vary in the interval [0, 1] from which we can determine the state of underload and overload. Fig. 2 illustrates that the CPU load varies depending on the amount of data processed, the sizes of the cloudlets vary from 40 000 to 140 000 Millions of Instructions as shown on Tab. 3, which leads to a variability of the workload of the processor that can range from 0.81% to 98.38%. Fig. 3 shows that the rate of memory usage varies rapidly depending on the number of tasks. The values of the measurements collected depend on the amount of data in execution and range from 0.89% to 95.23%. Fig. 4 shows that the different load measurements of the bandwidth collected during a period on each VM as a function of the number of tasks take values ranging from 1.75% to 93.69% representing the occupancy rate of the data transferred on the bandwidth. Fig. 5 shows the evolution of the execution time as a function of the number of tasks that uses a sharedspace policy in a dynamic environment. The required time increases gradually as the tasks are added dynamically. VMs with large capacity are faster and take less time to complete a task, as it is clearly visible on the curves.

5.3. Graphical Representation

Fig. 2. Real time CPU utilization based on the number of tasks

Fig. 3. Real-time memory utilization according to the number of tasks

Journal of Automation, Mobile Robotics and Intelligent Systems

VOLUME 14,

N° 3

2020

Abdellah Idrissi – Intelligent Processing Systems Team (IPSS), Computer Science Laboratory (LRI), Computer Science Department, Faculty of Sciences, Mohammed V University in Rabat, Rabat, Morocco, email: idrissi@fsr.ac.ma. Fig. 4. Real time bandwidth utilization based on the number of tasks

*Corresponding author

REFERENCES [1]

Fig. 5. Evolution of the execution time according to the number of tasks

6. Conclusion Processing large amounts of data in a high heterogeneity system such as the cloud results in a variability of the workload is not a simple task. In this paper, we developed a simulator that we named CloudSimulator based on CloudSim. It enables dynamic simulation in the cloud environment and provides monitoring of the current workload of VM resources such as CPU, memory and bandwidth. We proceeded by defining a class diagram that is adapted to the dynamic environment. Some classes of CloudSim have been modified by redefining methods and others have been newly created, and a new dynamic resource allocation policy has been implemented using the principle of space-shared policy. The information collected on the current load of resources determines the state of load of each VM that can be used to solve load balancing problems. When an overload is detected on a node, the overloads are transferred to the less loaded nodes. The goal of this study is effectively managed Cloud Computing resources to improve system performance. In future work, our proposed simulator may well be integrated into load balancing mechanisms and resource allocation algorithms. The information collected on the measurement of the execution time can be used as allocation metric in algorithms for assigning tasks to VMs based on a minimal processing time.

AUTHORS Faouzia Zegrari* – Intelligent Processing Systems Team (IPSS), Computer Science Laboratory (LRI), Computer Science Department, Faculty of Sciences, Mohammed V University in Rabat, Rabat, Morocco, email: z.faouzia@gmail.com.

P. Mell and T. Grance, “The NIST Definition of Cloud Computing”, Technical Report, DOI: 10.6028/NIST.SP.800-145. https://csrc.nist. gov/publications/detail/sp/800-145/final. Accessed on: 2020.12.16. [2] V. Sangeetha, V. Jaganraja and T. Gnanaprakasam, “A General Study of Homomorphic Encryption Algorithm with Cloud Computing”, Global Journal of Advanced Engineering Technologies and Sciences, vol. 3, no. 3, 2016. [3] S. Rajan and A. Jairath, “Cloud Computing: The Fifth Generation of Computing”. In: 2011 International Conference on Communication Systems and Network Technologies, 2011, 665–667, 10.1109/CSNT.2011.143. [4] R. N. Calheiros, R. Ranjan, A. Beloglazov, C. A. F. De Rose and R. Buyya, “CloudSim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms”, Software: Practice and Experience, vol. 41, no. 1, 2011, 23–50, 10.1002/spe.995. [5] T. Goyal, A. Singh and A. Agrawal, “Cloudsim: simulator for cloud computing infrastructure and modeling”, Procedia Engineering, vol. 38, 2012, 3566–3572, 10.1016/j.proeng.2012.06.412. [6] “The CLOUDS Lab: Flagship Projects – Gridbus and Cloudbus”. www.cloudbus.org/cloudsim/. Accessed on: 2020.12.16. [7] W. Tian, Y. Zhao, M. Xu, Y. Zhong and X. Sun, “A Toolkit for Modeling and Simulation of Real-Time Virtual Machine Allocation in a Cloud Data Center”, IEEE Transactions on Automation Science and Engineering, vol. 12, no. 1, 2015, 153–161, 10.1109/TASE.2013.2266338. [8] “xen [Wiki ubuntu-fr]”. https://doc.ubuntu-fr. org/xen. Accessed on: 2020.12.16. [9] D. Gupta, R. Gardner and L. Cherkasova, “XenMon: QoS Monitoring and Performance Profiling Tool”, Technical Report – HPL-2005-187, www.hpl.hp.com/techreports/2005/HPL2005-187.pdf. Accessed on: 2020.12.16. [10] R. Kumar and G. Sahoo, “Cloud Computing Simulation Using CloudSim”, International Journal of Engineering Trends and Technology, vol. 8, no. 2, 2014, 82–86, 10.14445/22315381/IJETT-V8P216. Articles

Journal of Automation, Mobile Robotics and Intelligent Systems

[11] S. Mehmi, H. K. Verma and A. L. Sangal, “Simulation modeling of cloud computing for smart grid using CloudSim”, Journal of Electrical Systems and Information Technology, vol. 4, no. 1, 2017, 159–172, 10.1016/j.jesit.2016.10.004.

Articles

VOLUME 14,

N° 3

2020

Journal of Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

UNIT LOAD DEVICES (ULD) DEMAND FORECASTING IN THE AIR CARGO FOR OPTIMAL COST MANAGEMENT Submitted: 26th June 2019; accepted: 25th March 2020

Mounia Mikram, Maryem Rhanoui, Siham Yousfi, Houda Briwa DOI: 10.14313/JAMRIS/3‐2020/37 Abstract: In recent decades, the airline industry has become very competitive. With the advent of large aircraft in service, unit load devices (ULD) have become an essential ele‐ ment for efficient air transport. They can load a large amount of baggage, cargo or mail using only one unit. Since this results in fewer units to load, saving time and efforts of ground crews and helping to avoid delayed flig‐ hts. However, a deficient loading of the units causes ope‐ rating irregularities, costing the company and contribu‐ ting to the dissatisfaction of the customers. In contrast, an excess load of containers is at the expense of cargo. In this paper we propose an approach to predict the de‐ mand for baggage in order to optimize the management of its ULD flow. Specifically, we build prediction models: ARIMA following the BOX‐JENKINS approach and expo‐ nential smoothing methods, in order to obtain more accu‐ rate forecasts. The approach is tested using the operatio‐ nal data of flight processing and the results are compared with four benchmark method (SES, DES, Holt‐Winters and Naive prediction) using different performance indicators: MAE, MSE, MAPE , WAPE, RMSE, SMPE. The results obtai‐ ned with the exponential smoothing methods surpass the benchmarks by providing more accurate forecasts. Keywords: Air Transport, ULD, Machine Learning, ARIMA, Exponential Smoothing

1. Introduction With the increasing importance of air cargo [1], many traditional airlines have shifted from simple passenger carriers to ”combined” carriers (cargo and passengers). Although passenger traf�ic remains the main source of revenue for mixed carriers, air cargo transport has become an increasingly important source of revenue for these companies. Usually, airlines use the bunkers of their passenger plane to transport goods. Thus, the delivery of freight for these carriers is strongly in�luenced by several fac‑ tors, as the number of the passengers, the �light sche‑ dule, the routing and the amount of baggage each pas‑ senger can bring. For these companies, it is very common to only load freight into the space remaining in the bunkers af‑ ter the total loading of all the passenger baggage. The‑ refore, there is no guarantee that a shipment will be sent in a speci�ic �light. For large aircraft, the transport of cargo and baggage is carried out by means of load units (ULD): pallets or containers, which allows rapid loading and unloading of freight and baggage and a

gain in terms of time and effort. Luggage demand forecasting is required to deter‑ mine the number of ULDs that are required to load baggage on planes and leave enough space to load cargo. Thereby, optimal use of ULDs for passenger baggage will improve passenger service and freight for maximum pro�it and service. Currently, ULD allocation to a �light is very empi‑ rical. It is therefore necessary to estimate the demand for �light baggage, to provide a scienti�ic basis for this allocation of ULDs for the passenger baggage to be em‑ barked and to improve the ef�iciency of the service. In this context, our research aims to use supervi‑ sed learning methods to predict short‑term (7 days) demand for baggage. The objective of this study is to build a prediction based on the ARIMA model and to compare its accuracy with the exponential smoothing models. The remainder of this paper is organized as fol‑ low: in Section II we present the background of for‑ castings models, section III presents relevant related works, section IV introduces our work and its motiva‑ tion, V presents the application of the forcasting mo‑ dels, and �inally, section VI gives a summary and re‑ commendations.

2. Motivation: ULD Management 2.1. ULD and Baggage Typology

Cargo units are pallets or containers used to load baggage, cargo and mail on containerized planes. They can load a large amount of baggage or freight or mail in one unit. Each ULD has its own packing list (or ma‑ nifest) so that its content can be tracked. In our case, there is 4 types of containers that can be loaded according to the types of machines:

‑ AKE: These types of containers are used on machi‑ nes B747‑400, B787‑800 and B767‑300;

‑ DQF: These types of containers are only used on B767‑300 machines; ‑ AAK: These types of containers are used on machi‑ nes B747‑400, B787‑800 and B767‑300; ‑ DPE: These types of containers are only used on B767‑300 machines, they are rarely used;

Below (Fig. 1) is an illustrating example of an AKE ULD. ULDs are identi�ied by these types of baggage, which facilitates their management. However, the lack of a suf�icient number of ULDs at the level of stopovers Articles

75 71

Journal Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME N°33 2020 2020 VOLUME 14,14, N°

This makes predictive planning to send the right number of ULDs even more dif�icult and delicate. The purpose of this paper is:

‑ Analyze baggage behavior during different periods for a transatlantic air route.

‑ Construct a predictive method for forecasting bag‑ gage demand. ‑ Develop a short‑term baggage forecast, delivering reliable and credible results for decision‑making re‑ garding the number of ULDs to be loaded for each �light. Fig. 1. Example of ULD specification [2] generates a mixed load of baggage making their treat‑ ment heavier, which can lead to delayed baggage and passenger complaints.

3. Context and Background: Econometrics and Forecasting Model

2.2. Provision of Stopovers and One‐Off Events

3.1. Time Series

A minimum number of ULDs is allocated at the le‑ vel of each stopover based on the types of machines used during the season for �lights to the stopover con‑ cerned, the �illing of �lights (number of passengers de‑ parting from the stopover concerned ), as well as the operational constraint of the stopover (example: Non‑ use of AA� type ULDs for �lights to Montreal). During periods of high traf�ic (such as weekends, holiday periods ..), the stopover time allocated to �light processing does not allow the repatriation of all ULDs unloaded at the stopover making stock rebuilding at the local HUB dif�icult. It is necessary to wait for the next �lights to these stopovers to repatriate the ULD, which is not always the case as it happens times that no jumbo jet is programmed at the level of this stopo‑ ver. To ensure that all baggage at a destination will be transported and delivered to their owners in the same �light, the planner checks each morning the movement of the ULDs, the stock status of each destination and the passenger forecasts of the �lights of the day. To then send instructions to the station managers on the number of ULDs containing the baggage, and empty (if necessary) to be loaded on board the aircraft for ope‑ rational �lights of the day.

A time series is a succession of observations over time. A time series usually consists of several ele‑ ments: ‑ Trend: represents the long‑term evolution of the se‑ ries.

2.3. Motivation

Baggage demand forecasting was required to de‑ termine the fair number of ULDs required to load air‑ craft baggage for each �light and leave enough space to load cargo.

The existing system allows for effective manage‑ ment and planning of ULDs on the company’s network, with real‑time visibility of the location and status of the ULD, and inventory control at the stations to ens‑ ure availability. However, the anomalies present at the level of sen‑ ding and processing of messages, the management of stopover stock as well as the irregularities related to the management of occasional events (with high traf‑ �ic), make the dynamic management of ULDs dif�icult, which leads sometimes to overstock or under‑stock at stopovers and generates delayed baggage. Articles

‑ Seasonality: evolution repeated regularly every year.

‑ Stationary (or residual) component: what remains when the other components are removed and des‑ cribes the short‑term evolution of the series. A time series comes from the realization of a family of random variables {Xt , t ∈ I}, where the set I is a time interval that can be discrete or continuous. For our study, we note the set I = {0, 1, ..., T }, where T is the total number of observations. 3.2. Forecast Models

In this work, we are interested in time series ana‑ lysis in order to understand the behavior of a variable and its dynamics, to discover the regularities and then to establish a short term forecast. Prediction methods are often subdivided into ca‑ tegories. We focus on forecasts based on the ARIMA model and the exponential smoothing models. 3.3. BOX JENKINS Method

The Box‑Jenkins method [6] refers to a set of pro‑ cedures for identifying and estimating time series mo‑ dels in the class of autoregressive integrated moving average (ARIMA) models. Box‑Jenkins’ approach to building the ARIMA mo‑ del includes the following steps: ‑ 1. Identify the parameters p, d and q of the model ‑ 2. Select the appropriate model ‑ 3. Diagnose the chosen model

‑ 4. Use the model for forecasting

Journal of Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

3.4. Smoothing Method The methods of exponential smoothing were intro‑ duced by Holt in 1957 [13] and by Winters in 1960 [22] and popularized by Brown in 1962 [7]. They constitute the set of empirical techniques that assign exponentially decreasing weights as the observation is more older. Thus, recent observations have more weight in the forecast than older observations. Simple exponential smoothing Simple exponential smoothing is used for short‑term forecasts. It assu‑ mes that the data �luctuates around a reasonably sta‑ ble average without seasonality and a locally consis‑ tent trend. The speci�ic formula for simple exponential smoothing is: X̃(h) = St

St = α × Xt + (1 − α) × St−1

With α is a real 0 < α < 1 The predicted value is the weighted average of pre‑ vious observations. If α = 0, then the current value is ignored, the new value depends entirely on the smoot‑ hed value that precedes it. The smaller the value of α, the greater the selection of the initial value of S. Thus the choice of the initial value affects the calculation of the values which follow it; it can be initialized by the average of � or 5 �irst observations. Double exponential smoothing Double exponential smoothing is used when the data show a trend. It is a generalization of simple exponential smoothing that assumes that the series approaches locally through an af�ine transformation of time. It is an exponential smoothing with a trend [15]. Its speci�ic formula for is: X̃(h) = St + hTt

St = α × Xt + (1 − α)(St−1 + Tt−1 ) Tt = γ(St − St−1 ) + (1 − γ)Tt−1

Where 0 < γ < 1 and 0 < α < 1 And h represents the horizon of the forecast made at time T.

Triple exponential smoothing (Holt‐Winters) This type of exponential smoothing makes it possible to add to the autoregressive component of the model, a trend and a seasonality. But that can be adapted to the series without seasonality by adjusting them by a line in the vicinity of T [15]. The formula of the additive seasonal HW model is: X̃(h) = St + hTt + It−p+h

St = α(Xt + It−p ) + (1 − α)(St−1 + Tt−1 ) Tt = γ(St − St−1 ) + (1 − γ)Tt−1 It = δ(Xt − St ) + (1 − δ)It−p

Where 0 < γ < 1 and 0 < α < 1 and 0 < δ < 1 And It is the seasonality index smoothed at the end of period t and p is the seasonality cycle.

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

3.5. Performance and Model Comparison To assess the credibility of a given model, valida‑ tion is an essential activity when faced with the need to make critical decisions based on modeling results. It allows us to decide whether the model responds cor‑ rectly and ef�iciently to our problem. Several methods are used in the benchmarking ap‑ proach for time series data to study the accuracy of a given model: Partitioning data for time series The partitioning of the data will divide the series to study in 2 periods:

‑ Train is the set of data used for the analysis and con‑ struction of the model. ‑ Test is the dataset used to verify and validate model performance. We assume that we do not have these data and we want to predict them.

‑ Future is the period of which we do not really know and we want to predict.

Generally, more data is allocated for training and less for testing. �ne can choose a �ixed data partition or by advancing the learning period (the partitioning is done several times). The latter has several advanta‑ ges; it allows us to compare the performance of roll‑ forward deployment scenarios. Performance Indicators (KPIs) To evaluate the accu‑ racy of the forecast, the validation period must be exa‑ mined by comparing the actual values Xt and the Ft va‑ lues generated by the model, by comparing their per‑ formance indicators. Several KPIs are possible for our study: Mean Absolute Error T 1 |Xt − Ft | T t=1

M AE =

The weaker it is, the smaller the gap between ob‑ servation and prediction. MAE is not used when the series is intermittent. Mean Squared Error M SE =

T 1 |Xt − Ft |2 T t=1

It is preferred to MAE because it is more sensitive to errors with small deviations. Mean absolute percentage error M AP E =

T 1 |Xt − Ft | T t=1 |Xt |

It can only apply to strictly positive values. Weighted Absolute Percentage Error W AP E =

|Xt − Ft | T t=1 Xt

t=1

Articles

Journal Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

Root mean squared error T 1 (Xt â&#x2C6;&#x2019; Ft )2 RM SE = T t=1 Symmetric Mean Percentage Error T Ft â&#x2C6;&#x2019; X t SM P E = t=1 T t=1 Ft + Xt

The best performing model is the one that minimiâ&#x20AC;&#x2018; zes the most these key performance indicators. Time series are widely used in various ďż˝ields, such as ďż˝inance [8] [21], energy consumption [20] [10], cloud performance [16]. We are interested in this paâ&#x20AC;&#x2018; per in predicting the demand for ULDs to optimize the cost of travel.

4. Related Works

The quality of an airlineâ&#x20AC;&#x2122;s service depends on its timeliness, accuracy, functionality, quality, and price. For these purposes, airlines need optimizationâ&#x20AC;&#x2018;based decision support systems Optimization [25]. There are several works in the literature that address different aspects of air transport optimization, as airline crew scheduling [4] [12], crew pairing [3] and ďż˝light planâ&#x20AC;&#x2018; ning among others [17]. However, the study and optiâ&#x20AC;&#x2018; misation of the management of ULDs is still rare, and for the most part does not exploit the techniques of arâ&#x20AC;&#x2018; tiďż˝icial intelligence and machine learning. Thus, Lu et al. [14] estimate safety stock levels of ULDs for international airline operations which is as the minimum quantity that can support the utilization during the entire trip. Limbourg et al. [19] deal with the problem of an optimal loading of ULDs in an airâ&#x20AC;&#x2018; craft. Wong et al. [23] has studied the issue of loading passengersâ&#x20AC;&#x2122; luggage in the cargo hold in an optimal way. Yan et al. [24] proposed a mixed integer non liâ&#x20AC;&#x2018; near model to address the problem of how to load the containers into an aircraft in a stochastic environment. Deploying machine learning in the business proâ&#x20AC;&#x2018; cess by using data from logistic information systems offers the company several advantages: anticipating the evolution of its stocks, optimizing ďż˝low manageâ&#x20AC;&#x2018; ment by reducing costs, thus enabling the steering committee to focus on decisionâ&#x20AC;&#x2018;making with better control and visibility. The positive impact of the machine Learning for the supply chain lies in the management of demand forecasting, and the anticipation of product needs, which eliminates operating irregularities (baggage not routed ..) and consequently allows a better service delivery for customers. ďż˝n the ďż˝ield of passenger air transport, several stuâ&#x20AC;&#x2018; dies have been conducted to study the performance of the ďż˝light based on historical data (number of passenâ&#x20AC;&#x2018; gers, number of baggage lost, number of delays ...) but there are very few works on the estimated demand in terms of baggage (especially ULD). Among them is Chengâ&#x20AC;&#x2122;s comparative study on bagâ&#x20AC;&#x2018; gage demand forecasting methods [9]. This work comâ&#x20AC;&#x2018; Articles

VOLUME NÂ°33 2020 2020 VOLUME 14,14, NÂ°

pares prediction models: neural networks and multiâ&#x20AC;&#x2018; ple regression to predict baggage demand. The models were built based on the historical data of the ďż˝light bagâ&#x20AC;&#x2018; gage claim. ďż˝n order to provide a scientiďż˝ic basis for the allocation of resources for checked baggage and to improve the efďż˝iciency of the passenger service, foreâ&#x20AC;&#x2018; casts were made by analyzing three types of data (data for all ďż˝lights, data of a single ďż˝light, and data of ďż˝ligâ&#x20AC;&#x2018; hts with the same destination). The authors suggest to optimize the neural network model or to choose a more adequate predictive model and address this isâ&#x20AC;&#x2018; sue more accurately. Also, Li conducted an analytical study on the deâ&#x20AC;&#x2018; parture baggage check at the airport based on passenâ&#x20AC;&#x2018; ger behavior [18]. This study was based on operatiâ&#x20AC;&#x2018; onal data from the airport to establish an analysis of the behavior of the baggage claim process and bagâ&#x20AC;&#x2018; gage claim characteristics such as weight and quanâ&#x20AC;&#x2018; tity, which can provide support for scientiďż˝ic decisionâ&#x20AC;&#x2018; making for the demand forecast of baggage. The reâ&#x20AC;&#x2018; sults of this analysis showed that baggage weight folâ&#x20AC;&#x2018; lows a widespread distribution of extreme value and demand varies according to the type of ďż˝light, which has led to improvements in the baggage registration process. Bokern [5], who was inspired by Dâ&#x20AC;&#x2122;Engelbronner [11], conducted research on creating a forecast based on two data sources: historical ďż˝light data and reservaâ&#x20AC;&#x2018; tion data. He showed in his thesis that a forecast can be made over a 10â&#x20AC;&#x2018;day horizon with an error of 2â&#x20AC;&#x2018;3%. We are interested in the prediction made on the basis of the ďż˝light data obtained from the ALTEA informaâ&#x20AC;&#x2018; tion system. To create this forecast, Bokern used two types of models: Autoregressive Moving Average Moâ&#x20AC;&#x2018; dels (ARMA) and Exponential Smoothing (ES) models. A comparison between the models was made on the basis of error measurements to determine which foâ&#x20AC;&#x2018; recasting model is the best in terms of forecasting.

5. Forecasting 5.1. Data Source

The source of the data is the AMADEUS ALTEA Deâ&#x20AC;&#x2018; parture Control System â&#x20AC;&#x2018; Costumer Management. Two data ďż˝iles will be extracted from the DCS and saved in Excel ďż˝iles: â&#x20AC;&#x2018; Statistics Baggage by period: this ďż˝ile contains staâ&#x20AC;&#x2018; tistics on checked baggage in each ďż˝light per peâ&#x20AC;&#x2018; riod (ďż˝light number, type of machine, date of deparâ&#x20AC;&#x2018; ture, departure, destination, number of pieces, total weight of coins ..).

â&#x20AC;&#x2018; Filling of ďż˝lights: provides information on the numâ&#x20AC;&#x2018; ber of passengers boarded for a ďż˝light (ďż˝light numâ&#x20AC;&#x2018; ber, type of machine, departure date, departure, desâ&#x20AC;&#x2018; tination, type of cabin, number of PAX recorded, number of PAX on board). After preprocessing these two ďż˝iles, a table of two columns will be created which will include the date of the ďż˝light and the corresponding baggage ďż˝ passenger ratio and which will represent our series to be analyâ&#x20AC;&#x2018; zed.

Journal of Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

Fig. 2. Structure of the final file to exploit

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

Fig. 5. Dickey‐Fuller test results

5.2. Data Analysis Before the implementation of the prediction mo‑ del, it is essential to visualize our series, analyze it and study its behavior over time to check if it contains null values (if yes, use the interpolation), and check if the observed values represent insigni�icant outliers to re‑ move them. A �irst analysis of the series makes it pos‑ sible to analyze the behavior of the data (Fig. 3):

Fig. 6. Correlogram of the original series (FAC, FACP) Note that the autocorrelation is positive and non‑ zero for a large number of lags. This con�irms that dif‑ ferentiation is necessary.

Fig. 3. The series 5.3. ARIMA Model Recall that the ARIMA model assumes that the se‑ ries is stationary. Thus, to identify the parameters of the model, a �irst step will be the study of stationarity.

Fig. 4. Mobile Average and Standard Deviation of the series The Dickey‑Fuller test allows us to check if the se‑ ries is stationary: From the results, we note that the statistical test is not less than the critical values of 10%, 5% and 1% (Fig. 5) . Thus, We can not reject the null hypothesis H0: ”the series is not stationary”. So, according to the Dickey‑Fuller test we conclude that the series is not stationary and needs to be differentiated. Let’s analyze the correlogram of the original series (Fig. 6):

Fig. 7. Correlogram of the series after differentiation (FAC, FACP) This graph represents FAC and FACP of the series after a �irst differentiation. Note that offset autocorre‑ lation 1 is negative and greater than ‑0.5. This indica‑ tes that our series does not need to be differentiated and therefore we estimate d = 1. Negative offsets jus‑ tify differentiation. On the FAC graph as well as the FACP, we notice that there is a peak at offset 1. This allows us to estimate the parameters p and q, with p = 1 and q = 1. Model Selection From the foregoing, the model of ARIMA (1, 1, 1) is a suitable model. To make sure of the validity of our choice, we compare the different possi‑ ble models with the information criteria AIC and BIC (Fig. 1): We note that the ARIMA model (1, 1, 1) minimizes the two criteria the most. Thus, the model to be imple‑ mented will be ARIMA (1, 1, 1). Diagnosis of residual According to the diagnosis of the residual (Fig. 8), it is found that the autocorrela‑ Articles

Journal of of Automation, Automation, Mobile Journal MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14, N°N°3 3 2020

Tab. 1. Comparison of the AIC and BIC of the different models ARIMA Model ARIMA (0, 0,1) ARIMA (0, 1, 1) ARIMA (1, 0,0) ARIMA (1, 0,1) ARIMA (1, 1,0) ARIMA (1, 1,1)

AIC 1830,63593742 ‑345,135984579 ‑177,874517485 ‑345,0467419 ‑247,992863849 ‑372,325641125

BIC 1839,78050251 ‑335,991419493 ‑168,729952399 ‑331,329894272 ‑238,848298764 ‑358,608793497

tion is signi�icantly zero for almost all offsets, with the exception of shift 7. Again, the distribution of the re‑ sidues follows the linear trend of the samples taken from a normal distribution and therefore can also con‑ sider that they follow a normal distribution.

Fig. 9. Simple exponential smoothing Tab. 3. Simple exponential smoothing Prediction Results Date 04‑01 04‑02 04‑03 04‑04 04‑05 04‑06 04‑07

Real Value 0,551838 0,559496 0,576744 0,610526 0,761290 0,750789 0,633952

SES 0,666563 0,602323 0,586976 0,601106 0,697217 0,729360 0,672115

It can be seen that the predicted values with simple exponential smoothing are higher but very close to the real values.

Fig. 8. Diagnosis of residues The short‑term forecast (7 days) gives the follo‑ wing results: Tab. 2. ARIMA(1, 1, 1) Prediction Results Datel 04‑01 04‑02 04‑03 04‑04 04‑05 04‑06 04‑07

Real Value 0,551838 0,559496 0,576744 0,610526 0,761290 0,750789 0,633952

ARIMA (1,1,1) 0,798146 0,804509 0,806367 0,806910 0,807069 0,807115 0,807128

According to these results, it is noted that the pre‑ dicted values are higher than the real values. To know to what extent the forecast of this model is accurate, we calculate in the next part the different indicators mentioned above. 5.4. Construction of Smoothing Models

Simple exponential smoothing (SES) Recall that the simple exponential smoothing makes it possible to cal‑ culate the prediction from the weighted average, by as‑ signing to each value a weight and where the weight decreases according to an exponential function. By partitioning the data into training data (Train) and test data (Test), we obtain the following graph : The table below summarizes the 7‑day results of our prediction: Articles

Double exponential smoothing (DES) Double expo‑ nential smoothing, as already seen, is an exponential smoothing suitable for series with a tendency Tt and a level St . The graph below shows the behavior of the series over the three periods: the Learning Train series in blue, the validation test series in red and modeling with the DES in purple:

Fig. 10. Double exponential smoothing And the results of the forecast are presented in the following table: Tab. 4. Simple exponential smoothing Prediction Results Datel 04‑01 04‑02 04‑03 04‑04 04‑05 04‑06 04‑07

Real Value 0,551838 0,559496 0,576744 0,610526 0,761290 0,750789 0,633952

DES 0,674789 0,526861 0,473992 0,498582 0,651936 0,748076 0,709196

Holt‐Winters exponential smoothing (HW) Holt‑ Winters Smoothing or Triple Smoothing involves applying exponential smoothing to the seasonal

Journal of of Automation, Automation, Mobile Journal Mobile Robotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME 2020 VOLUME 14,14,N°N° 3 3 2020

component, trend, and level. Thus, to implement it, we must formulate the equation of these last three.

Fig. 13. Naive prediction

Fig. 11. Triple Exponential Smoothing (HW) On the graph, we notice that modeling closely mi‑ mics our learning series. But by zooming in on the trial period, we observe that the modeling does not adapt well to the real values. By closely analyzing the graph, it can be seen that the exponential smoothing of HW does not correctly imitate the behavior of the test series:

Fig. 12. HW zoomed on the test period The table below presents the results of this fore‑ cast: Tab. 5. Holt‐Winters Prediction Results Date 04‑01 04‑02 04‑03 04‑04 04‑05 04‑06 04‑07

Real Value 0,551838 0,559496 0,576744 0,610526 0,761290 0,750789 0,633952

HW 0,784246 0.779285 0,727776 0,794183 0,765640 0,732398 0,875425

To compare between the two types of naive pre‑ diction, several KPIs were used: Tab. 6. KPIs of both types of naive prediction KPI MSE MAE RMSE MAPE WAPE SMPE

OS naive forecast 0,008579 0,069330 0,092617 10,77% 0,1092 0,55005

Naive forecast 0,013474 0,083110 0,116077 11,7520% 0,1309 0,07003

We note that all the performance indicators calcu‑ lated for the step‑by‑step forecast are minimal com‑ pared to the dynamic naive forecast. This allows us to conclude that stepwise forecasting is better than dyn‑ amic forecasting. But our goal is a 7‑day forecast (that is, estimate the next 7 days at one time). This forces us to choose Dynamic Forecasting as a benchmark for our benchmark. Model Comparison To examine the performance of each forecast model, a forecast analysis was perfor‑ med. In this analysis, a 7‑day forecast was made for transatlantic route �lights after 31 March (taken into account 2 years of history). From these predictions, a comparison can be made between different forecas‑ ting methods. Below is a graph of all the forecasting techniques used:

In the next section, we will perform a benchmark between the models as well as calculate the perfor‑ mance indicators of each one.

6. Performances, Discussion and Recommen‐ dations

This section presents the synthesis of this work. We summarize the results of each of the models stu‑ died and compare them to conclude which model will be the most appropriate for our problem. 6.1. Model Performance

Naive forecast In order to measure the performance of the models used, a �irst comparison with the naive forecast will be made. We have considered two types of naive forecast: a dynamic and a step by step:

Fig. 14. The different forecasting models On this graph, the blue line represents the test se‑ ries containing the actual values while the other lines represent the predicted values according to a model. By visualizing the graph, it is assumed that the simple exponential smoothing gives results closer to the rea‑ lity compared to the other models. Let’s recap the results of each model: Articles

Journal Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

VOLUME N°33 2020 2020 VOLUME 14,14, N°

Tab. 7. Comparison of the results of the forecasting models Date 01/04 02/04 03/04 04/04 05/04 06/04 07/04

Real Value 0,551838 0,559496 0,576744 0,610526 0,761290 0,750789 0,633952

ARIMA 0,798146 0,804509 0,806367 0,806910 0,807069 0,807115 0,807128

SES 0,666563 0,602323 0,586976 0,601106 0,697217 0,729360 0,672115

By analyzing the table, we notice that under no ci‑ rcumstances does the ARIMA model (1,1,1) correctly mimic the real behavior of our series, with values that are very out of step with reality, followed by the smoo‑ thing of Holt‑ Winters. The values closest to the actual values are shown in bold, followed by the values repre‑ sented in blue, which are mainly derived from single and double exponential smoothing. Thus, to decide the choice between the different models and choose the most adequate to our problem, we compare in the following part the performance in‑ dicators of the models used. Interpreting KPIs KPIs or performance indicators al‑ low us to measure the accuracy of our estimators by comparing them with actual values. Several KPIs have been calculated to evaluate the models and to check if the use of different KPIs will result in different results: This table indicates that all indicators are mini‑ mal for simple exponential smoothing. It also con�irms that ARIMA (1,1,1) represents the model with the hig‑ hest error value compared to the smoothing models. 6.2. Summary and Recommendations

The purpose of this study is to evaluate the ARIMA univariate time series prediction method and compare it to the exponential smoothing models (the three ty‑ pes of smoothing) to predict the baggage ratio. The goal is to �ind a model that �its the data correctly and could predict the behavior of our data. This is done by �irst differentiating to remove both the seasonal and trend components and to make the series stationary, then estimate the ARIMA models and adapt them to our data set. The ARIMA models as well as the exponential mo‑ dels were used in time series analysis and the best performing models were selected according to the in‑ formation criteria and comparing the error measure‑ ments. The best performing models were used for data

Tab. 8. KPIs of different forecast models KPI MSE MAE RMSE MAPE WAPE SMPE 82

Articles

Naïve 0,013474 0,083110 0,116077 11,75% 0,1309 0,07003

ARIMA 0,03532 0,17037 0,18794 28,75% 0,2683 0,1183

DES 0,674789 0,526861 0,473992 0,498582 0,651936 0,748076 0,709196

HW 0,784246 0.779285 0,727776 0,794183 0,765640 0,732398 0,875425

naïve 0,776358 0,776358 0,776358 0,776358 0,776358 0,776358 0,776358

forecasting. The data �low is represented by a non‑stationary time series. There is a trend and a seasonality. The pre‑ diction can be simpli�ied by studying the original dif‑ ferentiated series. ‑ Taking into account the different models of ARIMA, the ARIMA model (1,1,1) seems the most suitable for the data �low studied based on the information cri‑ teria AIC and BIC. ‑ After identifying and estimating the parameters of ARIMA (1,1,1), a diagnosis of the model was made. Having satis�ied all the assumptions of the validity of the model, this model is considered to be the most appropriate ARIMA model for forecasting. ‑ In addition to ARIMA, in order to determine which method achieves the best results, an exponential smoothing prediction has been made. ‑ The three types of exponential smoothing (Simple, Double and Holt‑Winters) have been implemented by de�ining functions according to the algorithm cor‑ responding to each type.

‑ To choose the model that mimics our data as accu‑ rately as possible, several performance indicators have been calculated. The most accurate model will be the model minimizing the value of the different calculated KPIs. ‑ Compared with all the models studied, the simple exponential smoothing allowed us to obtain a better result with an error rate (MAPE) of 6.32%.

7. Conclusion

Baggage demand forecasting becomes a very im‑ portant task for optimizing the management of ULDs. In this �ield, little research has been done and others are still improving. In this paper we applied the methods of predictive analysis taking into account the recommendations of SES 0,003030 0,042298 0,055045 7,00% 0,06769 0,0334

DES 0,008128 0,042981 0,090157 12,98% 0,12545 0,0638

HW 0,023707 0,109434 0,153971 15,55% 0,1723 0,0936

Journal of Journal of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

the research already done. The results obtained throughout our study and the steps carried out, showed that the ARIMA model re‑ mains far from reality even if it correctly imitated our dataset during the learning, and the simple exponen‑ tial smoothing model is the a model that minimizes KPIs and therefore is considered the best performing model for our forecast.

AUTHORS

Mounia Mikram∗ – LRIT, Faculty of Science, Moham‑ med V University in Rabat. Meridian Team, LYRICA Laboratory, School of Information Sciences, Morocco, e‑mail: mikram.mounia@gmail.com. Maryem Rhanoui – IMS Team, ADMIR Laboratory, Rabat IT Center, ENSIAS, Mohammed V Univer‑ sity in Rabat. Meridian Team, LYRICA Laboratory, School of Information Sciences, Morocco, e‑mail: mr‑ hanoui@gmail.com. �iham �ou��i – SIP Research Team, Rabat IT Center, EMI, Mohammed V University in Ra‑ bat. Meridian Team, LYRICA Laboratory, School of Information Sciences, Morocco, e‑mail: siha‑ myous�i@research.emi.ac.ma. Houda Briwa – Meridian Team, LYRICA Laboratory, School of Information Sciences. ∗

Corresponding author

REFERENCES

[1] AIRBUS. “Global Market Forecast. Growing Hori‑ zons 2017/2036”, 2017. https://airinsight. com/wp-content/uploads/2018/07/2017Airbus-GMF.pdf, Accessed on: 2020.12.17.

[2] ANA CARGO. “ULD & Aircraft Specs | ANA Cargo”. www.anacargo.jp/en/int/speci�ication/, Acces‑ sed on: 2020.12.17.

[3] E. Andersson, E. Housos, N. Kohl, and D. Wedelin. “Crew Pairing Optimization”. In: G. Yu, ed., Ope‑ rations Research in the Airline Industry, Interna‑ tional Series in Operations Research & Manage‑ ment Science, 228–258. Springer US, Boston, MA, 1998. [4] C. Barnhart, A. M. Cohn, E. L. Johnson, D. Klabjan, G. L. Nemhauser, and P. H. Vance. “Airline Crew Scheduling”. In: R. W. Hall, ed., Handbook of Transportation Science, International Series in Operations Research & Management Science, 517–560. Springer US, Boston, MA, 2003.

[5] B. Bokern. “Improve workload prediction in the �ield of Cargo Operations”, MSc Thesis, 2015. VU University Amsterdam, The Net‑ herlands. https://beta.vu.nl/nl/Images/ stageverslag-bokern_tcm235-701817.pdf. Accessed on: 2020.12.17. [6] G. E. P. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung, Time series analysis: forecasting and cont‑ rol, John Wiley & Sons, Inc: Hoboken, New Jersey, 2016.

VOLUME 2020 VOLUME 14,14, N°N° 3 3 2020

[7] R. G. Brown, Smoothing, Forecasting and Pre‑ diction of Discrete Time Series, Dover Publicati‑ ons: Mineola, NY, 2004.

[8] N. H. Chan, Time Series: Applications to Finance, John Wiley & Sons, Inc, 2004.

[9] S. Cheng, Q. Gao, and Y. Zhang, “Comparative study on forecasting method of departure �light baggage demand”. In: Proceedings of 2014 IEEE Chinese Guidance, Navigation and Cont‑ rol Conference, 2014, 1600–1605, 10.1109/CG‑ NCC.2014.7007431.

[10] C. Deb, F. Zhang, J. Yang, S. Lee, and K. W. Shah, “A review on time series forecasting techniques for building energy consumption”, Renewable and Sustainable Energy Reviews, vol. 74, 2017, 902– 924, 10.1016/j.rser.2017.02.085. [11] R. D’Engelbronner, Air Cargo Handling Demand Forecasting; ’As a support tool for short‑term de‑ cision on manpower deployment at World Port’, MSc Thesis, Delft University of Technology, The Netherlands, 2012. [12] M. Dunbar, G. Froyland, and C.‑L. Wu, “Robust Ai‑ rline Schedule Planning: Minimizing Propagated Delay in an Integrated Routing and Crewing Fra‑ mework”, Transportation Science, vol. 46, no. 2, 2012, 204–216.

[13] C. C. Holt, “Forecasting seasonals and trends by exponentially weighted moving averages”, Inter‑ national Journal of Forecasting, vol. 20, no. 1, 2004, 5–10.

[14] Hua‑An Lu and Chien‑Yi Chen, “Safety Stock Es‑ timation of Unit Load Devices for Internatio‑ nal Airline Operations”, Journal of Marine Science and Technology, vol. 20, no. 4, 2012, 431–440, 10.6119/JMST‑011‑0322‑1. [15] P. S. Kalekar. “Time series Forecasting using Holt‑ Winters Exponential Smoothing”, 2004.

[16] A. Khan, X. Yan, S. Tao, and N. Anerousis, “Wor‑ kload characterization and prediction in the cloud: A multiple time series approach”, 2012 IEEE Network Operations and Management Sym‑ posium, 2012, 1287–1294.

[17] S. Lan, J.‑P. Clarke, and C. Barnhart, “Plan‑ ning for Robust Airline Operations: Optimi‑ zing Aircraft Routings and Flight Departure Ti‑ mes to Minimize Passenger Disruptions”, Trans‑ portation Science, vol. 40, no. 1, 2006, 15–28, 10.1287/trsc.1050.0134. [18] Z. Li, J. Bi, J. Zhang, and Q. Li, “Analysis of Ai‑ rport Departure Baggage Check‑in Process Ba‑ sed on Passenger Behavior”. In: 2017 10th In‑ ternational Symposium on Computational Intelli‑ gence and Design (ISCID), vol. 2, 2017, 204–207, 10.1109/ISCID.2017.149. [19] S. Limbourg, M. Schyns, and G. Laporte, “Auto‑ matic aircraft cargo load planning”, Journal of the Operational Research Society, vol. 63, no. 9, 2012, 1271–1283, 10.1057/jors.2011.134. Articles

Journal Journal of of Automation, Automation,Mobile MobileRobotics Roboticsand andIntelligent IntelligentSystems Systems

[20] G. R. Newsham and B. J. Birt, “Building‑level occupancy data to improve ARIMA‑based electricity use forecasts”. In: Proceedings of the 2nd ACM Workshop on Embedded Sensing Systems for Energy‑Ef�iciency in Building ‑ BuildSys ’10, Zurich, Switzerland, 2010, 13–18, 10.1145/1878431.1878435.

[21] M. Rhanoui, S. Yous�i, M. Mikram, and H. Me‑ rizak, “Forecasting Financial Budget Time Se‑ ries: ARIMA Random Walk vs LSTM Neural Network”, IAES International Journal of Arti�i‑ cial Intelligence (IJ‑AI), vol. 8, no. 4, 2019, 317, 10.11591/ijai.v8.i4.pp317‑327. [22] P. R. Winters, “Forecasting Sales by Exponen‑ tially Weighted Moving Averages”, Management Science, vol. 6, no. 3, 1960, 324–342.

[23] W. H. Wong, A. Zhang, Y. Van Hui, and L. C. Leung, “Optimal Baggage‑Limit Policy: Airline Passenger and Cargo Allocation”, Transporta‑ tion Science, vol. 43, no. 3, 2009, 355–369, 10.1287/trsc.1090.0266. [24] S. Yan, Y.‑L. Shih, and F.‑Y. Shiao, “Optimal cargo container loading plans under stochastic de‑ mands for air express carriers”, Transporta‑ tion Research Part E: Logistics and Transpor‑ tation Review, vol. 44, no. 3, 2008, 555–575, 10.1016/j.tre.2007.01.006. [25] G. Yu and J. Yang. “Optimization Applications in the Airline Industry”. In: D.‑Z. Du and P. M. Par‑ dalos, eds., Handbook of Combinatorial Optimiza‑ tion, 1381–1472. Springer US, Boston, MA, 1998.

Articles

VOLUME N°33 2020 2020 VOLUME 14,14, N°

VOLUME 14, N° 3, 2020 www.jamris.org

Indexed in SCOPUS

Journal of Automation, Mobile Robotics and Intelligent Systems pISSN 1897-8649 (PRINT) / eISSN 2080-2145

(ONLINE)

WWW.JAMRIS.ORG • pISSN 1897-8649 (PRINT) / eISSN 2080-2145 (ONLINE) • VOLUME 14, N° 3, 2020

logo podstawowe skrót

Publisher: ŁUKASIEWICZ Research Network – Industrial Research Institute for Automation and Measurements PIAP

logo podstawowe skrót

ŁUKASIEWICZ Research Network – Industrial Research Institute for Automation and Measurements PIAP