ARCI 2022 Proceedings
ARCI 2022 Proceedings
net/publication/375423446
CITATIONS READS
0 198
1 author:
Sergey Y. Yurish
International Frequency Sensor Association (IFSA)
125 PUBLICATIONS 594 CITATIONS
SEE PROFILE
All content following this page was uploaded by Sergey Y. Yurish on 07 November 2023.
Copyright © 2022
by International Frequency Sensor Association (IFSA) Publishing, S. L.
E-mail (for orders and customer service enquires): [email protected]
Visit our Home Page on http://www.sensorsportal.com
All rights reserved. This work may not be translated or copied in whole or in part without the written permission
of the publisher (IFSA Publishing, S. L., Barcelona, Spain).
Neither the authors nor International Frequency Sensor Association Publishing accept any responsibility or
liability for loss or damage occasioned to any person or property through using the material, instructions, methods
or ideas contained herein, or acting or refraining from acting as a result of such use.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not
identifies as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary
rights.
ISSN: 2938-4788
ISBN: 978-84-09-37741-1
BN-20220201-XX
BIC: TJFM
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Contents
Foreword ........................................................................................................................................................... 5
3
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
4
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Foreword
On behalf of the ARCI’ 2022 Organizing Committee, I introduce with pleasure these proceedings with
contributions from the 2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0
(ARCI’ 2022), 2-3 February 2022.
According to the modern market study, the global Industry 4.0 market will reach US$ 165.5 Billion by 2026
growing at the CAGR of slightly above 20.6% between 2021 and 2026. The Industry 4.0 means the usage of an
integrated system, which consists of an automation tool, robotic control and communications. The key factors
fuelling the growth of the industry 4.0 market include rapid adoption of Artificial Intelligence (AI) and Internet
of Things (IoT) in manufacturing sector, increasing demand for industrial robots, rising government investments
in additive manufacturing, and growing adoption of blockchain technology in manufacturing industry.
Industry 4.0 represents the 4th industrial revolution that marks the rising of new digital industry. It is defined as
an integrated system that comprises numerous technologies such as advanced robotics control, automation tools,
sensors, artificial intelligence, cloud computing, digital fabrication, etc. These technologies help in developing
machines that will be self-optimized and self-configured. It helps in enhancing asset performance, technology
usage, material usage and other industrial processes that are involves in various industries. Numerous benefits are
offered by these technologies such as low operational cost, improved productivity, enhanced customer
satisfaction, improved customization, and increased efficiency. The Industry 4.0 holds a lot of potentials and is
expected to register a substantial growth in the near future. There are several conferences on automation, robotics
and communications, but they are not meet the Industry 4.0 challenges.
The series of annual ARCI Winter IFSA conferences have been launched to fill-in this gap and provide a forum
for open discussion of state-of-the-art technologies related to control, automation, robotics and communication -
three main components of Industry 4.0. It will be also to discuss how to adopt the current R&D results for Industry
4.0 and to customize products under the conditions of highly flexible (mass-) production.
The conference is organized by the International Frequency Sensor Association (IFSA) - one of the major
professional, non-profit association serving for sensor industry and academy more than 20 years, in technical
cooperation with media partners – journals: MDPI Processes (Switzerland), MDPI Machines (Switzerland) and
Soft Measurements and Computing (Russia). The conference program provides an opportunity for researchers
interested in signal processing and artificial intelligence to discuss their latest results and exchange ideas on the
new trends.
I hope that these proceedings will give readers an excellent overview of important and diversity topics discussed
at the conference.
We thank all authors for submitting their latest works, thus contributing to the excellent technical contents of the
Conference. Especially, we would like to thank the individuals and organizations that worked together diligently
to make this Conference a success, and to the members of the International Program Committee for the thorough
and careful review of the papers. It is important to point out that the great majority of the efforts in organizing the
technical program of the Conference came from volunteers.
5
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(002)
Summary: Formal determination of the complexity indicators of the environment of a mobile object operating in three-
dimensional space in the presence of obstacles is considered. The mathematical substantiation of the method for calculating
the complexity is given. The concepts of local and integral complexities of the environment are introduced. Formulas for
calculating the complexity are given.
6
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
not such, we will assume that it is replaced by a simple, if there are no obstacles at all on the scene, up
minimal ball containing this obstacle inside. to maximum, if obstacles block the target and do not
It is assumed that the MO has range sensors that allow the MO to move to the target. In order to
allow scanning a spatial area in the direction of reasonably choose a strategy for the formation of the
movement. Geometrically, the scanned area is a trajectory, it is necessary to have a quantitative
pyramid S with the top that coincides with the MO. measure (S ) of the complexity of the scene S .
The pyramid is always symmetrical about the axis
of movement of the MO. Moreover, it has a fixed apex
angle equal to 2 h in the horizontal plane, and equal 3. General Requirements for the Measure
to 2 v in the vertical plane (Fig. 1). The pyramid with of Complexity
the obstacles indicated in it and, possibly, the target,
we will sometimes call the scene and also denote S . The choice of a measure of complexity is, generally
speaking, not so much a mathematical question as a
question reflecting the physical content of what is
meant by the complexity of a real scene. It is possible
to present different requirements to the complexity of
Z this measure, meaning various parameters of the scene,
characterized and/or taken into account by this
measure. Therefore, the introduction of such a measure
can be carried out in many ways. In particular, taking
Y into account that dynamically changing scenes are
generally considered, the degree of complexity can
v depend on time t , i.e., S (t ) S , t .
h
It should also be noted that such a measure will
only relatively characterize the complexity of the
Scene S scene, in the sense that the attainability of a given
MO
target will depend not only on the parameters of the
scene itself, in particular, the size and number of
X obstacles, but also on the size of the mobile object
itself. The scene, which is relatively simple for a small
Fig. 1. Scene S and scanning angles. MO, can be difficult for a large mobile object, since
the target may be unattainable for it due to the lack of
sufficiently wide passages between obstacles.
We will assume that the faces forming the pyramid
At the same time, it is possible to formulate a
are always flat, while the base opposite to the top of the
general requirement, the fulfilment of which is
pyramid in reality can have an arbitrary shape,
naturally required from any measure of complexity.
including kinks and breaks, since it is the result of
The measure should be positive and preferably
scanning the real space in front of MO. Nevertheless,
for the sake of simplicity, we will assume that it has normalized, i.e., 0 S(t) 1 for any S and t . In
the shape of a flat rectangle (Fig. 1), at a distance H this case, equality to zero should correspond to the
from the vertex (the length of the centerline of the simplest scene, on which there are no obstacles, and
pyramid). the trajectory can be a segment of a straight line. The
We also assume that there is an external coordinate most difficult scene, on which it is impossible to lay a
system that allows to accurately position the object and trajectory to the target, should have a difficulty equal
the target. However, further constructions will be to one.
considered in the coordinate system associated with
the MO, in which the origin of coordinates coincides
with the MO center, the ordinate axis – with the MO 4. Triangulation and Spatial Partitioning
axis in the direction of its movement, and the abscissa
axis is directed to the right, so that together with the We will assume that there are n obstacles
applicate axis, the coordinate axes form right triplet B1, B2 ,..., Bn on the scene S . Each obstacle in space
(Fig. 1).
We will assume that the space outside the scanned will be characterized by four real numbers
pyramid is completely occupied by obstacles, i.e. it is Bi xi , yi , zi , ri , where xi , yi , zi are the coordinates
impossible to build trajectories outside the pyramid. In
of the center of the obstacle, and ri is its radius. For
other words, all possible trajectories of the object's
movement lie strictly inside the pyramid. convenience, the center of the i -th obstacle will also
Depending on the position of the target and MO, as be denoted pi (xi , yi , zi ) . We will also designate by
well as the parameters of the scene (position, number,
size of obstacles), the difficulty of achieving the target P p1, p2 ,..., pn – a set of obstacle centers located
may vary. This "difficulty" can vary from extremely in space.
7
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Consider the vertices of the pyramid S as new dimensional space of a four-dimensional polyhedron
virtual obstacles BA , BB , BC , BD , BE and construct the of a special type. Linear relations for the number of
faces of different dimensions of this polyhedron are
spatial Delaunay triangulation on expanded set of described by the Dehn-Sommerville equations [11]. If
obstacles [10]. With this construction, all internal
we denote by f 1 the number of edges (one-
regions of the resulting partition will be tetrahedra, and
its surface will be the convex hull of the triangulation dimensional faces) of the Delaunay polyhedron, then
(Fig. 1). with the help of some additional constructions it can be
shown that the following relations hold:
BC M 2N 3 0 ,
(1)
BB n f1 N 7 0 .
8
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
where [] is the vector product, – radius of the For any i and j all paths from x i to x j fall into
mobile object, i are the angles in radian measure at disjoint classes of paths of topological length 1, 2, …
the corresponding vertices of the face, 0 is a However, some of these classes may be empty.
certain "gap" of safety. The more , the safer the Therefore, finding the width of the safest path is
passage. reduced to choosing the safest path among the safest
paths in each class, i.e., max rij( k ) .
k I
Fig. 3. The region of passage of the object through the face Note that the matrix R has size N N , and by
of the two regions of the partition.
virtue of relations (1), equality N f1 n 7 holds.
Thus, to determine the exact size of the matrix R , it is
The entered widths can be considered as a weight necessary to know the number f1 of edges of the
function defined on the edges of the graph. Thus, we resulting partition. However, it can be seen that the size
consider a weighted topological graph G X , U , w of the matrix is proportional to f1 . Meanwhile, an
with a marked set X MO of vertices corresponding to asymptotic estimate is known [12], namely, any
the position of the object, and a vertex xT triangulation in a space of dimension d contains
corresponding to the target.
The width of a path on a graph is a value equal to
O n
d 2
simplices, where is the nearest integer
the smallest width among all the edges that form this greater than or equal to that element, so in our case it
path. Consider all the paths on the graph that connect is O n 2 . Thus, for large n , there is a quadratic
the object to the target. Reachability of the target dependence of the matrix size on the number of
refers to the maximum width among all such paths. obstacles. In this case, the number of edges of a
topological graph is related to the number of vertices
by the relation M 2N 3 , therefore, the number of
7. Matrix Procedure for Computing nonzero elements of the matrix R is exactly
Reachability determined by its size and by virtue of symmetry is
equal to 4N 6 . Therefore, at large N , the matrix R
To calculate the reachability xT , one can use a is very sparse, which can be used to optimize
simple matrix procedure. Consider the adjacency computations.
matrix R rij of a graph G X , U , w , where Reachability as a characteristic of a scene is
inconvenient, since, on the one hand, it is expressed in
w(uij ), if xi and x j are adjacent , absolute units of distance, i.e., depends on the chosen
rij scale, and, on the other hand, it does not satisfy the
0, otherwise.
axiomatic requirements for a measure of complexity.
Next, we will consider an expression for a complexity
Now define the maximin composition R R R
(2)
measure associated with the reachability and devoid of
of the matrix R with itself according to the rule the indicated disadvantages.
R(2) R R rij(2) , where rij(2) max min rik , rkj ,
kI N
(2) 8. Local Complexity Measure
where I N {1, 2,..., N }. Then rij is equal to the
maximum width among all widths of all paths of length Note that with growth it is natural to assume that
2 connecting the vertices x i and x j . In particular, this the complexity ( ) should decrease. Moreover, if
means that there is at least one maximally "safe" path is large enough, and accordingly ( ) small, then
(2)
of length 2 from x i to x j , the width of which is rij . further increase will only slightly reduce the
complexity. On the contrary, for small reachability, the
Next, inductively define R R R rij
(3) (2) (3)
, complexity of the scene should be large, and small
increases of should lead to a rapid decrease in
…, R( n1) R(n) R rij(n1) , …
complexity. Thus, we can assume that the rate of
9
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
change of complexity at any value of reachability is are reported to the mobile object in the process of
inversely proportional to the value of complexity at the moving it on the scene, or they can change at different
same value, i.e., ( ) ( ) , where is some times. An integral measure can be constructed as
follows.
positive constant characterizing this dependence.
For simplicity, we consider the case when all
Hence, we obtain that ( ) Ce , where C is the obstacles are point, i.e. have a zero radius. Let us
constant of integration, which can be established on the assume for definiteness that the tetrahedron in which
basis of the following considerations. the mobile object is located has number 1. Let us
When 0 the target is unattainable for the MO, denote by 1 ( p ) the complexity of reaching the point
and the complexity of such a scene is maximal. Taking p ( x, y ) S by the mobile object located at the top of
into account the axiomatically imposed requirement
that the complexity measure be normalized, i.e., the pyramid. Then, as an integral characteristic of the
complexity of the scene, we can take the averaging
0 ( ) 1 , we obtain (0) С 1 and, therefore
over the local measures of the complexity of all points
( ) e . of the scene:
The parameter determines how quickly the
complexity of the scene decreases with increasing . 1
S
m S S
1 ( p)dp ,
The choice of can be made in various ways, for
example, as follows. Since the limiting value
( ) 0 is formally reached only at , then for where m S is the volume of the pyramid S .
real problems it can be assumed that the complexity is It is easy to see that the function 1 ( p ) is
"practically" equal to 0, if, say, ( ) , where is piecewise constant on S . If we denote by the Sk - k
a small positive number. On the other hand, it can be -th region of the spatial partition of the pyramid, and
assumed that the complexity of the implementation of
the movement of the MO is practically zero if the path by m Sk its volume, then
width (i.e., reachability) exceeds the area of the
diametrical section 2 of the mobile object by more 1 N N
m Sk r1k
S
m S k 1 Sk
1
( p ) dp
k 1 m S
e .
than a factor k . In other words, we can assume that
2 ln
e k , from where . With this in
k 2 Here the quantities m S and m Sk are easily
mind, the expression for the complexity measure takes calculated from geometric considerations.
This formula is obtained under the assumption that
the form ( ) k 2
. Assuming, for example the obstacles are point. In this case, it admits of a
simple probabilistic interpretation. If the appearance of
0 .0 1 and k 20 , we get ( ) 10 10 2
. a point specifying the target is a random variable
m Sk
The reachability value depends on the coordinates uniformly distributed in S , then is the
of the target, so the last expression should be mS
understood as a function probability of the target appearing in the area Sk ,
therefore, the last expression is the mathematical
( p ) e ( p ) , expectation of the local complexity function.
10
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Acknowledgements
This work was carried out at the Research and Design
Fig. 5. LC = 0.5371, IC = 0.6282. Bureau for Robotics and Control Systems with the support
of the Russian Science Foundation, project no. 22-29-00533.
References
[1]. M. Hoy, A.S. Matveev, A.V. Savkin, Algorithms for
collision-free navigation of mobile robots in complex
cluttered environments: a survey, Cambridge
University Press, 2014.
[2]. Xingjian Jing, Behavior dynamics-based motion
planning of mobile robots in uncertain dynamic
environments, Robotics and Autonomous Systems,
Vol. 53, Issue 2, 2005, pp. 99-123.
[3]. Path Planning for Vehicles Operating in Uncertain 2D
Environments, Ed. V. Pshikhopov, Elsevier,
Fig. 6. LC = 0.3670, IC = 0.4503. Butterworth-Heinemann, 2017.
[4]. W. Li, On the Relationship Between Complexity and
Entropy for Markov Chains and Regular Languages,
Complex Systems, Vol. 5, Issue 4, 1991, pp. 381-399.
[5]. M. Feixas, E. del Acebo, Ph. Bekaert, M. Sbert, An
Information Theory Framework for the Analysis of
Scene Complexity, EUROGRAPHICS’99, Vol. 18,
Issue 3, 1999.
[6]. L. Niepel, J. Martinka, A. Ferko, P. Elias, On Scene
Complexity Definition for Rendering, Winter School
of Computer Graphics and Visualization 95
(WSCG’95), Plzen, 1995, pp. 209–217.
[7]. D. Plemenos, M. Sbert, M. Feixas, On Viewpoint
Complexity of 3D Scenes, in Proceedings of the
International Conference GraphiCon, 2004, Moscow,
Russia, (http://www.graphicon.ru)
Fig. 7. LC = 0.9889, IC = 0.9805.
[8]. J. Rigau, M. Feixas, M. Sbert, Visibility Complexity of
a Region in Flatland, EUROGRAPHICS, 2000.
11
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
[9]. A. Karkishchenko, V. Pshikhopov, On Finding the [11]. A. Brøndsted, An Introduction to Convex Polytopes,
Complexity of an Environment for the Operation of a Springer Verlag, 1983.
Mobile Object on a Plane, Automation and Remote [12]. R. Seidel, The upper bound theorem for polytopes: an
Control, Vol. 80, Issue 5, 2019, pp. 897-912. easy proof of its asymptotic version, Computational
[10]. Siu-Wing Cheng, Tamal K. Dey, Jonathan Shewchuk, Geometry, Vol. 5 Issue 2, 1995, pp. 115–116.
Delaunay Mesh Generation, Chapman and Hall/CRC,
2013.
12
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(004)
Summary: In the context of Industry 4.0 and the proclaimed digitalization of production towards autonomous and self-
monitoring systems, an increasing demand for novel monitoring and diagnostic functions can be derived. Especially in the
area of production technology, continuous monitoring and evaluation of the manufacturing processes is of particular interest
for developing autonomous production systems. A contribution can be made by machine internal drive systems and their
integrated sensors and measurement functions. In this paper, a partial aspect of an observer structure for estimating load and
process forces based on motor currents and axis positions of electromechanical drive systems is presented. The key objective
is the identification of the required transfer functions of the mechanical transfer behaviour between load application point and
servomotor. The methodology is based on the exclusive use of drive internal excitation and signal sources. An identification
routine automatically determines the system order based on frequency responses and calculates a corresponding model. The
optimization of the model parameters is carried out by applying a Nelder-Mead-optimization algorithm. Subsequently, the
optimized models are inverted and compared to the original system. In addition, all model equations are discretized to enable
their implementation on discrete sampled computing systems like the machine control.
Keywords: Disturbance observer, Parameter estimation, Transfer function, Discrete inversion, Multi-mass system.
13
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
restricted. Eventually, the observer estimates the torque Ta as well as the transfer function of the
unmeasurable load torque Tl . An approach for the mechanical part Gmech (z) must be determined.
automatic identification and subsequent validation of In the following section, the procedure for the
the models for friction compensation (Tf,m , Tf,i and automatic identification and inversion of the required
Tf,l ) and axis weight correction Tg were already discrete models is presented. Subsequently, the
investigated in [11]. For the complete observer, the experimental validation of the proposed method is
inverted transfer functions in the indirect case Gp,m (z) carried out on a feed axis of an exemplary machining
and direct case Gp,l (z) for estimating the acceleration center. The paper concludes with a summary of the
results and an outlook.
Fig. 1. Signal flow chart of multi-mass system and transfer function-based disturbance observer (TFDOB)
for drive-based disturbance estimation.
2. Methodology the same time, the angular velocities on motor side and
load side are recorded. Subsequently, the signals are
The fundamental sequence for determining the discrete transformed into frequency domain and their
inverse model equations is shown in Fig. 2. In the first magnitude and phase parts are calculated. To avoid
module, the system is excited on the motor side with a distortions of the frequency response in the range of
Pseudo-Random Binary Signal (PRBS). The main the controller bandwidth, the speed controller should
advantage over other established identification be parameterized as a backup controller during signal
routines, e.g. in [9, 12] is the renunciation of external recording.
excitation sources like impulse hammer or shaker. At
Fig. 2. Flow chart of the discrete model estimation and inversion procedure.
14
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Based on the magnitude response, the model values. Therewith, the value of parameter a is set equal
parameters are estimated in module two. The to one (cf. Eq. (3)). If, on the other hand, a phase
frequency responses for the indirect case Gp,m (z) and rotation of -180 ° takes place without subsequent
direct case Gp,l (z) are characterized by the total mass reversal, this indicates a partial oscillator according to
moment of inertia in the range of low frequencies Eq. (3) with the parameter a set to zero. The associated
according to Eq. (1). In addition, depending on the damping values dr,i and df,i are initialized with default
number of resonance and antiresonance frequencies values of 0.01.
ωr,i and ωf,i , a defined number of partial oscillators can In the second part of module two, the optimization
be detected. These are calculated using Eq. (2). For the of the model parameters is performed using Nelder-
direct case, an adjustment of the partial oscillator Mead method [15]. The previously determined
characteristic can be made by adapting parameter a parameters Jtot , ωr,i , ωf,i , dr,i and df,i are substituted
{0,1} (cf. [10, 15]). into the corresponding Eq. (3) or (4) and the deviation
between measured and modeled magnitude response is
1 minimized. In addition, the method is allowed to vary
GJ s = (1) the frequency values in the range of ± 10 %. The
Jtot ∙s
damping values may vary between 0 and 1. Eventually,
1 2 2 2df this results in a continuous-time model for the indirect
a∙ ∙s + ∙s+1 case and direct case (Eq. (3)) as well as the mechanical
ωf ωf
GPO s = (2) part according to Eq. (4).
1 2 2 2dr
∙s + ∙s+1 In module three, the transformation of all
ωr ωr
continuous-time models into discrete-time models is
By multiplying GJ s with the product of all partial performed using the z-transformation [16]. The cycle
oscillators GPO,i s in Eq. (3), the actual model time Tcyc of the installed control system is used for the
equations for Gp,m (s and Gp,l (s are derived. sampling time TA. Furthermore, there may occur large
differences between the largest and smallest exponents
n when calculating discrete transfer functions. This can
Gp,ml s =GJ s ∙ GPO,i s (3) ultimately lead to inaccuracies when determining the
i=1
corresponding poles and zeros. In order to avoid this
effect, the transfer functions are initially transformed
The model of the mechanical transmission into the discrete state space. Subsequently, a
behavior Gmech (s), which can also be calculated as the transformation of the system matrices into an
quotient of Gp,m (s) and Gp,l (s), is calculated solely by equivalent system is performed using a transformation
multiplying all i partial oscillators (Eq. (4)). matrix. The underlying procedure is described in detail
in [17].
n In module four, the inversion of the model
Gmech s = GPO,i s (4) equations is carried out. Especially in case of direct and
i=1 mechanical transfer functions (Eq. (2) and (3)), the
resulting models may have a non-minimum phase
As illustrated in the next section, a separate character. This means that the transfer function has one
estimation for all three model equations is necessary to or more unstable zeros, which is expressed in the pole-
achieve correct model parameters. This can be justified zero diagram by their location outside the unit circle.
by minor differences in frequency and damping values By inverting the model equation, all zeros become
depending on the location of the direct measurement poles and vice versa. Consequently, the inverted
system and the constructive design of the system. transfer function would have the same number of
The underlying methodology for the actual unstable poles, which basically result in an unstable
determination of the model parameters was already system itself. In this case, the transfer function is
presented in [13, 14] and partially extended. The basic decomposed into a minimum phase component and an
idea is to generate a non-parametric model by all-pass component. The basic procedure is explained
connecting several partial oscillators in series and in detail in [18]. For the subsequent inversion, only the
multiply them with the proportion for the total mass product of the minimum phase component and the
moment of inertia. Therefore, in the first part of magnitude of the all-pass component is used. Although
module two, Jtot is determined by calculating the this results in a phase shift of the inverted transfer
gradient at lower frequencies. At the same time, the function (cf. Fig. 5), however, deviations in the phase
resonance and antiresonance frequencies ωr,i and ωf,i are usually tolerable for numerous applications. On the
are determined based the phase characteristic. It is other hand, the course of the magnitude response is not
necessary to distinguish between two cases affecting affected.
the value of parameter a of Eq. (2). If a previously Furthermore, another obstacle arises when
defined phase threshold value (e.g. 50 °) is exceeded in inverting the model equations. Due to the fact that the
the negative or positive direction and there is a order of the denominator n is always larger than the
subsequent reversal of the phase to the original value, order of the numerator m for all model equations, a
the frequencies are close to the crossed threshold system with differentiating character is obtained by
15
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
performing a direct inversion. However, such a system 850V of the company DMG Mori. The tests were
cannot be implemented in reality for reasons of carried out on the horizontally arranged x-axis of the
causality [19]. A possible solution was already machine. Its structural design is illustrated in Fig. 3.
presented in [10] and applied for a second-order The rotational motion of the servomotor is converted
system. The idea is to extend the inverted transfer into a linear movement of the carriage via a belt gear
function with a defined number (n-m) of high- and a ball screw drive. Position sensors are located on
frequency poles until the denominator order the servo motor and on the carriage, whose time
corresponds at least to the numerator order. For a derivation serves as the input signals for the
continuous system, this is calculated as follows: identification procedure. By taking into account the
spindle pitch hsp and gear ratio ig the load speed is
-1
1 1 converted onto the motor shaft. The connection of the
Gp s = ∙ n-m (5) PRBS excitation signal to the motor torque Tm is
Gp s 1+TAT ∙s
implemented by a drive-internal function generator.
Note that the multiplication with high-frequency The results at the output of module two (model
poles also leads to a lowering of the phase response of identification) are illustrated in Fig. 4. All identified
the inverted transfer functions. With respect to the and optimized parameters are listed in Table 1. For all
magnitude, however, an adequate inversion of the three modell equations, a total number of three partial
model equation is obtained. oscillators was detected. Note that all values for
resonance and antiresonance frequency were
converted from ( rad∙s-1 ) to (Hz). Regarding the
frequency responses, the identified continuous models
3. Experimental Results (orange) match the measured signals (blue) over a
broad range of frequencies. Only the cycle time of the
The presented approach is validated on a real controller (Tcyc = 8 ms) limits the bandwidth of the
machine axis of a three-axis machining center DMC recorded signals up to 125 Hz.
10 1 10 1 10 1
10 1 10 1 10 1
measurement identified model optimized model deviation between measurement and optimized model
Fig. 4. Frequency response of measured signals and identified and optimized models.
16
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
kgm2
dr 0.23 dr 0.09 dr 0.03
Gp,l(s) Jtot in 0.0057
s2 ωf in Hz 28.82 ωf in Hz 58.95 ωf in Hz 99.33
dr 0.19 dr 0.06 dr 0.07
ωf in Hz 26.18 ωf in Hz 63.04 ωf in Hz 85.83
kgm2
dr 0.17 dr 0.09 dr 0.26
Gmech(s) Jtot in -
s2 ωf in Hz 23.69 ωf in Hz 47.09 ωf in Hz 85.83
dr 0.16 dr 0.10 dr 0.09
In the next step, all identified and optimized models one all-pass. Due to the phase-minimum character, this
are transferred to their discrete-time representation and procedure is not required for the indirect case as well
the subsequent inversion takes place. The results can as the mechanical part. The magnitude response shows
be seen in Fig. 5. The frequency responses of all three that the all-pass component (purple) has basically no
discrete models (blue), their discrete inversion influence on the minimum-phase component (yellow).
(orange) as well as the compensated value (green) are Only in the phase signal an additional drop is
plotted individually. Initially, a phase drop for all recognizable. Nevertheless, a phase reduction can be
models due to the discretization procedure is observed for all transfer functions. This corresponds to
recognizable. By reducing the sampling time TA, the other methods, for example the closed-loop approach
bandwidth of the models may be further enhanced. For in [9]. Apart from that, it can be stated that especially
the direct case in the center part of the figure, the in the magnitude response, an adequate discrete
algorithm has detected one unstable zero and inversion is performed for all three model equations.
consequently performs an all-pass decomposition with
10 1 101 10 1
10 1 10 1 10 1
discrete model discrete, inverted model allpass part phase-minimal model compensated signal
Fig. 5. Frequency response for discrete modeled and inverted transfer functions.
17
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
18
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(006)
Summary: Production technology is characterized by the use of electromechanical feed axes, for which the concept of cascade
control has become established. The concept is based on linear control engineering. It is not suitable for the control of process
forces, which is associated with nonlinearities. Here, adaptive control algorithms from the field of higher control engineering
represent a promising approach. The basis for adaptation is the estimation of process parameters. In this paper, the development
of a model for online parameter estimation is presented. Key components are data pre-processing, parameter estimation based
on a recursive least squares (RLS)-algorithm, and data post-processing. The functionality of the parameter estimation is
demonstrated in simulation and validated by means of experiments on a test setup with modern industrial motion control. In
addition, various influencing factors are examined and their effects are evaluated.
19
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
20
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Variable load characteristics can be initiated with high points 𝑥 , 𝐹 and new data are provided cyclically.
reproducibility by a movement of the axis against the Averaging is performed as long as the following abort
spring package. The use of spacer elements can conditions are not met:
provoke high steps in the stiffness. The stiffnesses of
the springs are known, which gives a reference value ∣𝑥 𝑥 ∣ 𝜀 𝑜𝑟 ∣ 𝐹 𝐹 ∣ 𝜀 (8)
for parameter estimation. A manufacturing tolerance
of 10 % per spring must be taken into account. When the position limit 𝜀 or the force limit 𝜀 is
exceeded, the process is terminated and the mean value
4. Implementation and Validation of the interval is transferred to the ring buffer of the
input data memory. In addition, force and position
4.1. Structure of the Parameter Estimation Model values in the following cycle 𝑡 1 , are defined as
the new reference data points and averaging starts
As shown in Fig. 2 the parameter estimation model again. This corresponds to a digital FIR low pass
consists of three components. The data pre-processing, filtering whose filter length is controlled by the abort
the actual parameter estimation with the help of the condition (8). Thereby it is ensured that a model
RLS-algorithm and data post-processing. Core excitation is present and no invalid data sets with
component is the parameter estimation, which is 𝛥𝑥 0 are used. The generation of the input data for
explained in chapter 2. The basic structure of this the RLS-algorithm with the cycle time 𝑡 is based on
algorithm is described in more detail in [4], [7] and [8].
the mean values 𝑥 𝑡 and 𝐹 𝑡 and it is shown
Once the algorithm has been chosen, it can be adapted
graphically in Fig. 3(b). Overlapping data points are
by using the forgetting factor, depending on the
specific implementation. The design of the data pre- used for a balance between high 𝛥𝐹 -excitation and fast
response to stiffness changes. This overlap is achieved
processing and the choice of suitable design
parameters is considerably more extensive than the by means of the ring buffer of length 𝐿 in the input data
RLS-algorithm. The purpose of data pre-processing is memory. It corresponds to the difference between the
to improve the signal-to-noise ratio and to provide the newest and oldest mean value, so that 𝐿 2 values are
RLS-algorithm with signals with a high excitation. skipped. In this way, suitable RLS input data 𝛥𝑥 𝑡
Therefore, suitable data sets of 𝛥𝑥 and 𝛥𝐹 for the and 𝛥𝐹 𝑡 can be generated if the measurement data
stiffness estimation must be generated. The discrete are strictly monotonically decreasing or increasing
data of the measured force 𝐹 𝑡 , the actual position over sufficiently long time periods. Typically, this
𝑥 𝑡 , the actual speed 𝑣 𝑡 and the speed setpoint applies to the position and the correlated force. In a
𝑣 𝑡 are made available to the parameter estimation final step, after parameter estimation, data post-
processing can take place by low-pass filtering.
unit as an input signal from the motion control in the
base cycle of the PLC 𝑡 1𝑚𝑠. A sign check of the
4.2. Design of the Parameter Estimation Model
feed rate is performed against the 𝐿 previous
measurement data in the input data memory, which has
First, the effects of individual influencing factors in
the cycle time 𝑡 . This allows the detection of standstill
the data pre-processing are investigated. In the
phases and measurement noise and serves for data pre- following, a suitable design parameter set for the
filtering. Subsequently, the data is low-pass filtered parameter estimation unit will be determined.
and stored in the input data memory. The sequence of Synthetic data are used for this purpose, since they can
filtering in the data preprocessing is shown in Fig. 3(a). be used to generate exact stiffness curves with known
With the initialization of the algorithm, the data points parameters.
𝑥 𝑥 𝑡 and 𝐹 𝐹 𝑡 are defined as reference
21
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
a) b)
Fig. 3. Sequence of averaging in data pre-processing from measured data (a) and generation of the input data for the RLS-
algorithm with an input data memory length of 𝐿 3 (b).
The setpoint and actual force 𝐹 and 𝐹 , the 𝜎 0.513𝑁 . The position signal at standstill shows
actual position 𝑥 as well as the setpoint and actual noise with a variance of 𝜎 1.9585 ∗ 10 𝑚𝑚 .
velocity 𝑣 and 𝑣 are required in the estimation Furthermore, the actual velocity shows a noise with a
model. These signals have to be provided simulatively
variance of 𝜎 3.4985 ∗ 10 . To simulate
with a correct sampling frequency, a correct
quantization and a correct noise by the machine model. this noise using Matlab Simulink, a stochastic process
Furthermore, 𝐹 is calculated in the test setup from of mean-free, Gaussian-distributed pseudo-random
the moving average of the current and the last nine numbers with a sampling time of 1𝑚𝑠 is added to the
measured values. The measured values show noise and actual values. The test data for the investigation of
are quantized to 1 𝑚𝑁. This is also taken into account influencing factors consist of movements at a feed rate
in the model. To simulate the behavior of the parameter of 0.2 , 0.4 and 1 into a spring with a
estimation unit, a realistic modeling of this signal is stiffness of 100 . The stiffness is set to 500
particularly relevant. In order to determine the variance
after one second. Then the feed rate is kept constant for
of the force signal, the machine is loaded from 0𝑁 to
500 𝑚𝑠. The sampling frequency is 1 𝑘𝐻𝑧. The results
5𝑘𝑁 at a typical stiffness value 𝐾 500 . The are illustrated in Fig. 4. It can be seen that varying the
force set points of 1𝑘𝑁, 2𝑘𝑁, 3𝑘𝑁, 4𝑘𝑁 and 5𝑘𝑁 are filter length 𝐿 of the ring buffer has little effect and the
held for 2𝑠 each. From the steady state within these estimation behavior improves with higher values. A
holding periods, the variances were determined. The similar effect can be observed for the position limit 𝜀 .
mean variance of the quantized measurement signal is
a) b)
-80 -80
500 500
-100 -100
400 -120 400 -120
300 -140 300 -140
Fig. 4. Influence of different settings of the parameter estimation unit on the behaviour of the estimation. a) Variation of the
length 𝐿 of the input data memory; b) Variation of the position limit 𝜀 ; c) Variation of the force limit 𝜀 ; d) Variation of the
average filter length 𝑚.
22
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
The parameter estimation model reacts faster and results with recorded measurement data from the
there is a better approximation to the nominal value for machine are illustrated in Fig. 5. A ramp-shaped force
smaller force limits 𝜀 . This can also be observed for profile with standstill phases is used. The influence of
the average filter length 𝑚 . The estimation of 𝐾 the adjustment factor 𝛾 is investigated here. Larger
agrees after 40-50ms in nearly all cases with the values lead to a faster estimation of the stiffness (see
setpoint jump of 𝐾 to 90%. Fig. 5(b)), but this also increases noise effects (see
Fig. 5(c)).
a)
600 0
-1000
400
-2000
200 -3000
-4000
0
-5000
4 5 6 7 8 9 10 11
Time [ms] 10
4
b) c)
200 5 510
0 -2420
150 500
-5
-2440
100 -10 490
-15
50 480 -2460
-20
0
-25 470 -2480
4.75 4.752 4.754 4.756 4.758 4.76 6.14 6.145 6.15 6.155 6.16
Time [ms] 4 Time [ms] 4
10 10
Fig. 5. Comparison of the parameter estimation 𝐾 for different adjustment parameters 𝛾 using real measurement data of the
machine (simulation with pre-recorded measurement data). a) Overview of the entire stiffness estimation; b) Parameter
estimation at the first stiffness jump 𝐾 ; c) Noise of the parameter estimation 𝐾 .
23
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Fig. 6. Parameter estimation on the test setup. [1]. J. M. Allwood, et al., Closed-loop control of product
properties in metal forming, CIRP Annals –
Manufacturing Technology, Vol. 65, 2016, pp. 573-
596.
A comparison between the estimated stiffness 𝐾 and
[2]. X. Yao, et al., Machining force control with intelligent
the actually calculated stiffness 𝐾 using the values in compensation, The International Journal of Advanced
Table 2 shows very good agreement. Open-loop tests Manufacturing Technology, Vol. 69, No. 5-8, 2013,
with several abrupt changes in stiffness were also pp. 1701-1715.
carried out on the test setup. Here, as in the simulation, [3]. W. Leonhard, Control of Electrical Drives - Power
the estimation model provided reliable values for Systems, 3rd edition, Heidelberg, Germany: Springer,
stiffness steps. Berlin Heidelberg, 2012.
[4] I. D. Landau, et al., Adaptive control - Algorithms,
analysis and applications, Communications and
Table 2. Data for verification of the parameter estimation control engineering, 2nd edition, London, Springer,
unit with real measured values on the test setup. 2011.
[5] K.-P. Schulze and K.-J. Rehberg, Entwurf von
adaptiven Systemen - Eine Darstellung für Ingenieure,
Data Plateau 𝑷 Vol. 1, Berlin, VEB Verlag Technik, 1988.
Start position, [6] R. Neugebauer, et al., Non-invasive parameter
𝑃 10, 𝑃 26.37, identification by using the least squares method,
End position, 𝑚𝑚
𝛥𝑥 𝛥𝑥 16.37 Archive of Mechanical Engineering, Vol. 58, 2011,
pp. 185-194.
Start force, End 𝐹 809, 𝐹 [7] R. Isermann, Identification of dynamic systems: An
force, 𝛥𝐹 𝑁 introduction with applications, Springer, Berlin, 2011.
3000, 𝛥𝐹 2191 [8] K. J. Åström and B. Wittenmark, Adaptive Control,
Stiffness 𝐾 2nd edition, Dover Books on Electrical Engineering.
determined from 𝑁 Newburyport, Dover Publications, 2013.
𝐾 133.84
𝛥𝑥 and 𝛥𝐹 𝑚𝑚 [9] S. Beineke, Online-Schätzung von mechanischen
Approximated Parametern, Kennlinien und Zustandsgrößen
geregelter elektrischer Antriebe, Dissertation,
estimation 𝐾 of 𝑁
𝐾 130 … 140 Universität Paderborn, Fortschritt-Berichte VDI Reihe
the parameter 𝑚𝑚 8, Meß-, Steuerungs- und Regelungstechnik,
estimation unit
Düsseldorf, VDI Verlag, 2000.
[10] R. Neugebauer, et al., Time-Based Method for the
Combined Identification of Velocity-Loop Parameters,
5. Summary and Conclusion Archive of Mechanical Engineering, Vol. 58, 2011,
pp. 175-184.
In this paper, a model for estimating the process [11] D. W. Clarke and Ch. J. Hinton, Adaptive control of
stiffness was presented. It is composed of the three materials-testing machines, Automatica, Vol. 33,
Issue 6, 1997, pp. 1119–1131.
components data pre-processing, parameter [12] A. Sewohl, et al., Performance analysis of the force
estimation, which is based on an RLS-algorithm, and control for an electromechanical feed axis with
data post-processing. The verification of the model industrial motion control, in Proceedings of the 17th
was done in three steps by simulations and International Conference on Informatics in Control,
experiments. The influence of different design Automation and Robotics 2020, Paris, France, Vol. 1,
parameters was also investigated. It has been found 2020, pp. 667-674.
that the estimated stiffness 𝐾 agrees sufficiently [13] D. W. Clarke, Adaptive control of servohydraulic
accurate with the actual stiffness 𝐾 for both real and materials-testing machines: a comparison between
black- and grey-box models, Annual Reviews in
synthetic data. In addition, the developed model Control 25, 2001, pp. 77–88.
provides reliable results even in the face of various
challenges, such as stiffness steps, reversal of the
direction of motion and standstill phases. With the
methodology described in this publication and the
developed parameter estimation unit, the basis for the
24
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(007)
Summary: In this paper, we simulate implementing automated handling for transferring fabric part pieces throughout the
construction process of US military-relevant textile products. The current infrastructure of the textile industry would benefit
from integrating new automated technology in the textile production line to replace manual operations. Using Simio software,
we simulated the current production process for the cummerbund in the US Army’s Modular Scalable Vest (MSV). The current
production is done through batch production with manual transferring of the batches of parts to the next workstation and the
loading/ unloading of part pieces into sewing workstations. Using SIMIO simulation software, we can model the implications
on time and labor of replacing the manual operations with automated robotic handling to enable continuous flow
manufacturing. Incorporating automation in the manufacturing of the cummerbund increases efficiency, throughput, and
quality of production.
Keywords: Automated handling, Materials handling, Textile manufacturing, Digital simulation, Industry 4.0.
25
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
26
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
seconds added to the previously mentioned robotic components to enter the operation and those exact
assembly operations to represent the loading and components to be in a completed cummerbund. The
unloading of parts. long time in the system is attributed to the bottlenecks
within the assembly. The modeled output aligns with
the manufacturer’s current production capabilities that
3. Results and Discussion were provided. Even optimizing the system by adding
an additional operator and station at two bottleneck
3.1. Simio Model of Current Production operations, the number produced per day can only
of Cummerbund Assembly increase 419.8 units/day with the average time in the
system 150.8 minutes.
The current factory layout and results from the
simulation are shown in Fig. 1 and Table 3,
respectively. In this simulation model, each assembly 3.2. Scenario 1: Automated Handling with Batch
step was given a 10 % variance in the production time Production
using a Beta Pert distribution. The variance is added to
account for the variability of a human operator having Robotic arms were added to the model to simulate
some faster and some slower production times. The automated loading and unloading of the fabric pieces
variance of a human operator can range depending on into the automated equipment shown in Fig. 2.
the operator’s experience, the difficulty of the
operation, and fatigue.
D E A B/C
F
G
Fig. 2. Simio model of current factory facility with an
integrated robotic arm for handling fabric.
H
JJ
K The implementation of robotic arms reduced the
I average production times to the times provided in
Table 2. The variance was reduced to 5 % for the
automated processing steps to reflect the consistency
Fig. 1. Current production layout. and reliability that automated handling offers. The
production remained in batch production, where 20
units had to be completed before transferring to the
Table 3. Production output of current factory layout
in simulation. next workstation. The automated handling enabled an
increased workday of sixteen-hours to account for the
Current Production Output reduced labor that could be redistributed to two shifts.
Number of Operators 11 This is assuming the manufacturer is able to employ
Number of Robots 0 one additional operator, or reallocate an operator from
Hours in Work Day 8 a different production line. The results of automated
Average Time in System 660.43 minutes handling within the workstation are shown in Table 4.
Number Produced per Hour 39.72 units/hour
Average Numbers 317.79 units/day
Table 4. Production output of current production with
The simulation of the current production accounted inserted robotic handling to load/unload automated
equipment.
for an eight-hour workday. For this production line, the
manufacturer employs eleven operators, or one shifts Automated Batch Production Output
worth, due to the labor shortage of seamstresses in the Number of Operators 6
United States. Currently, the manufacturer is not able Number of Robots 5
to hire two shifts worth of sewing operators due to the Hours in Work Day 16
scarcity. The current production line produces 317.79 Average Time in System 628.38 minutes
cummerbunds a day, or 39.72 units per hour. On Number Produced per Hour 39.17 units/hour
average, the cummerbund is in the system for 660.43 Average Numbers Produced 626.67 units/day
minutes. This is the time it takes for the fabric
27
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
28
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
improve continuous production. We modeled having equipment that can operate continuously without
two operators at operations G, I, and J. However, with breaks. There is one operator at K, which can be filled
employing more workers on the production assembly, by three shifts to account for the 24-hour workday. The
the assembly operation returned to one shift, or an results of continuous automated production are shown
eight-hour workday. This is assumed due to the in Table 8.
likelihood of not being able to hire or reallocate seven
additional operators. The results are shown in Table 6.
Table 7. Production time of automated assembly steps of
cummerbund when using robotic handling and automated
Table 6. Production output of continuous production with equipment.
inserted robotic handling at automated workstation and an
additional operator at operations G, I, and J. Cummerbund Production Time
Assembly Step (seconds)
Automated Continuous Production A (A) 5
with Redistributed Labor Output B (A) 5
Number of Operators 9 C (A) 28
Number of Robots 5 D (A) 27
Hours in Work Day 8 E (A) 10
Average Time in System 971.83 minutes F (A) 6
Number Produced per Hour 47.28 G (A) 15
Average Numbers Produced 378.28 H (A) 7
I (A) 20
K 33
Redistributing the labor showed improvements in Total Time 221.6 seconds
production output and a reduction in bottlenecks. The
cummerbunds produced per hour increased to
47.28 units and the average time in systems decreased Table 8. Production output of continuous production with
to 971.83 minutes. However, due to the eight-hour robotic handling and automated assembly at operations A-I.
workday, the daily output fell to 378.23 units a day, an
All Automated Processing Output
improvement from the current output, but a regression
Number of Operators 1
from both automated assemblies with sixteen-hour Number of Robots 9
workdays. Hours in Work Day 24
Average Time in System 4.23 minutes
Number Produced per Hour 97.2 units/hour
3.5. Scenario 4: All Automated Processing Average Numbers Produced 2332.8 units/day
29
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
enabling two shifts of operators, doubling the remaining four assembly steps, autonomous
production hours. Batch production serves as a more continuous flow production is possible in producing
suitable production process when there are longer 97.2 units per hour and 2332.8 units per day due to the
human operations still in the assembly process. If capability to run twenty-four hours. Automated
available, labor should be redistributed to the longer handling of textiles is essential to automating the
operations to decrease the bottleneck effects. However, assembly of textile products. Automating the handling
if redistributing the operators reduces production to of textiles improves efficiency for production by
one shift, the daily output will decrease despite an reducing assembly times and enabling longer
increase in the hourly production. An all automated production hours. Automation in textile handling is
production process has the most efficient assembly and most beneficial when the entire process can be
operates for twenty-four hours thus significantly automated, alleviating the bottlenecks that come with
increasing the daily output. manual operations in a continuous production line.
However, automated equipment, including machinery
and robotic arms, requires a significant investment, so
Table 9. Comparison of the production output of continuous it essential to assess the value of adding automation to
production with robotic handling and for the current the production process before investing.
manufacturing process and the four alternate scenarios.
Scenario Current 1 2 3 4
Operators 11 6 6 9 1
Acknowledgments
Robots 0 5 5 5 9
Units/Hour 39.72 39.17 34.97 47.28 97.2 This work was funded through the University of
Units/Day 317.79 626.67 559.54 378.28 2332.8
North Carolina Systems Office and the United States
Army Combat Capabilities Development Command.
4. Conclusions References
Simio software was used to simulate the [1]. International Labour Office- Geneva, The future of
implications of inserting automated handling in the work in textiles, clothing, leather and footwear,
assembly of an MSV cummerbund. When automated International Labour Office, Sectoral Policies
Department. – Geneva: ILO, 2019. Working Paper:
handling is employed at the automated equipment
No. 326.
operations in the current batch production process, the [2]. R. Handfield, H. Sun, L. Rothberg, Assessing Supply
daily production of cummerbunds nearly doubles to Chain Risk for Apparel Production in Low Cost
626.67 units from improved efficiency and introducing Countries Using Newsfeed Analysis, Supply Chain
a sixteen-hour work day. Using automated handling in Management, Vol. 25, No. 6, 2020, pp. 803–21.
continuous flow production is less effective producing [3]. K. Kulkarni and R. Adivarekar, Developments in
559.54 units per day. Reallocating the saved labor can Textile Continuous Processing Machineries, in
improve the bottlenecks in continuous manufacturing Advances in Functional Finishing of Textiles, Springer
by increasing hourly production from 34.97 units per Singapore, 2020.
hour to 47.28 units per hour, but due to the increased [4]. M. Suh, Automated Cutting and Sewing for Industry
4.0 at ITMA 2019, Journal of Textile and Apparel,
labor, an eight-work hour day must be employed, Technology and Management, Special Issue: ITMA,
reducing daily production to 378.28 units per day. 2019.
With the creation of automated equipment for the
30
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(008)
Summary: Simultaneously Mapping And Localization (SLAM) systems are essential in robotics systems. Using a single robot
to map a large scene is time-consuming. Therefore, a map can be constructed using a group of robots working together. Visual-
based loop detection in cooperative SLAM is essential, which can help to efficiently and accurately merge and construct global
maps among robots working independently. However, most visual SLAM algorithms focus on loop detection along the same
trajectory direction. These algorithms cannot handle multi-robot cooperative systems well because, in such systems, multi-
robot trajectory direction may often form loops along opposing or perpendicular directions. This paper proposes a multi-robot
cooperative SLAM system based on panoramic images. Each robot uses a camera as a sensor and a local map algorithm to
construct a local map. All local maps are merged into a global map when loops are detected. The panoramic images provide
scene information in all directions for loop detection. Experiments show that the loop detection accuracy improves in the
proposed system, and the time for constructing a global map is significantly reduced using a multi-robot cooperative system.
31
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Afterward, keyframe-based approaches were FAB-Map 2.0 [16] can use a vocabulary as large as
proposed. It only uses the selected frame as the 10,000 words.
keyframe to estimate the map. PTAM [4] was the first Some neural network-based methods have been
keyframe-based SLAM algorithm proposed to separate applied to place recognition. NetVLAD [17] is
the two functions of tracking and mapping in the modified based on VLAD [18] and directly embedded
SLAM system as two threads. ORB- SLAM2 [5] is one in the trainable CNN(convolutional neural network)
of the most popular visual SLAM. It adopts the multi- architecture. [19] used the Edge Boxes object proposal
threaded framework of PTAM, but the ORB[6] method combined with a mid-level CNN[20] to
features are extracted. Loop detection thread is added, identify and extract landmarks to complete place
which uses an efficient search algorithm BoW(Bag-of- recognition.
Words) [7]. In the past few years, feature-based methods have
become one of the main technologies in place
recognition. Its low computing cost and effectiveness
2.2. Fisheye Based Visual SLAM can be applied to low processing and storage capacity
devices [21]. The application scenarios of the multi-
Because a single camera’s FoV (Field of View) is robot cooperation visual SLAM system are complex
limited, extracting feature points and detecting the and huge. It increases the difficulty of pre-training of
correct loop in some indoor and outdoor scenes is not the neural network-based methods. The neural
easy. The fisheye camera has a wide FoV and can network-based approach requires large computation.
collect more scene information. The challenge of using In a centralized system, the central server needs to
the fisheye camera is how to solve the fisheye camera’s process a large amount of data from the client robots,
distortion effectively. In [8], cubemap is used to and the use of neural network-based methods will
improve the distortion of the fisheye camera. The increase its burden.
method uses five planes: up, down, left, right and front,
and then projects the spherical image separately. [9] is
a multi-fisheye camera SLAM system. This method 3. Methodology
requires the calibration of a polynomial model of the
fisheye camera. The camera parameters are used for 3.1. System Framework
connection transformation between multiple cameras.
[10] uses a multi-fisheye camera to obtain 360-view This cooperative SLAM system uses a visual
images and input them into the SLAM system for odometer based on a panoramic camera as a sensor and
mapping. a global place recognition based on panoramic images.
The use of panoramic images solves the problem of
loop detection in different trajectory directions in a
2.3. Visual Based Cooperative SLAM multi-robot system.
As shown in Fig. 2, it is the workflow of the multi-
Generally, multiple-robot cooperative SLAM can robot cooperative SLAM system. The system contains
be a centralized system [11]. In a centralized system, a central server and multiple robots. Suppose robots in
multiple robots equipped with sensors, communication the system are equipped with a panoramic camera and
components, and processors are used to complete the wireless communication interfaces. Each robot starts
mapping task in a centralized system. Each robot can from its initial position and independently constructs
run visual SLAM independently. The computation for its local map. The robot’s initial position is random,
some tasks, such as map merging, is performed by a which depends on the user. The robots do not know the
specific robot on the team or by an external agent [12]. position and coordinate system of other robots. The
The specific robot or the external agent communicates visual SLAM algorithm used by each robot is similar
with other robots as the system’s center. Other robots as the algorithm described in [10]. The difference is
are independent and do not communicate with each that we have added a data transmission part in the loop
other. The central server receives data sent by other detection. The central server communicates with
robots and provides required data and feedback. [13], multiple robots in the system. Each robot will send its
[14] are cooperative systems that use this structure. map data to the central server through its wireless
communication channel in robot movement. The
central server maintains a globally consistent map and
2.4. Place Recognition a corresponding sub-map for each robot. After the
central server receives the data of each robot, it will
Place recognition [15] is of great significance to reconstruct the keyframes and map points consistent
visual SLAM. One method is to extract local features. with each robot and add them to the sub-map
Hundreds of local features may be extracted from each corresponding to each robot. And the central server
image. BoW [7] clusters discrete extracted features and will do place recognition among the sub-maps. When
uses many images to train offline to generate a it is detected that different robots have come to
vocabulary containing multiple words. A typical common areas, the sub-maps corresponding to these
vocabulary usually contains thousands of words, but robots will be merged to get a globally consistent map.
32
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Each robot is equipped with a LadyBug3 camera as We transmit data mainly about keyframes and map
a sensor. The panoramic images taken by the points and their corresponding relationships. We use
LadyBug3 camera are fed to local panoramic SLAM keyframes instead of all captured frames. For example,
which has been modified based on [10]. Every robot a dataset containing 498 frames, a map constructed by
will generate and maintain its own map. See an a single robot only contains 181 keyframes. Only
overview Fig. 3, there are four main threads in the local transmitting keyframes greatly reduces the data that
panoramic SLAM, tracking, local mapping, loop needs to be transmitted and ensures the map
closing, and global optimization. The first three information’s integrity.
threads typically run in parallel. Global optimization is We use LCM (Lightweight Communications and
enabled after a loop is detected. Marshalling) [22] to complete data transmission. LCM
The tacking and the local mapping work together can encapsulate the data to be transmitted into a whole,
to complete the visual odometer function, which uses facilitating data transmission and management.
continuous panoramic images to produce a sparse map. Our goal is to continuously rebuild the map of each
This map shows the robot’s trajectory and is used to robot using the transmitted data in the central server,
achieve localization. The loop detection is in charge of so there is no tracking module in the central server. The
place recognition. It detects whether the robot is at the loop detection module only uses the information of
same place it has been before. If the loop is detected, keyframes and map points and does not update those
the local map near the current keyframes is optimized information. Therefore, we chose to fetch the relevant
using the two keyframes of place recognition matches. data in the loop thread of each robot and transmit it.
Next, the global BA (Bundle Adjustment) thread is In the transmission process, packet loss is inevitable
enabled. It will apply the optimization to all the due to the bottleneck of network communication. The
keyframe poses, and map point coordinates in the map. central server can associate newly received data with
These processes are used to reduce the accumulated the current map to reduce the impact of packet loss.
errors in the robot movement.
33
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
distance metrics M calculated for the two keyframes is covisibility keyframes, loop keyframes, and non-loop
less than the threshold, then they are similar enough to keyframes. We can get that the calculated distance
be considered as a loop. This loop candidate pairs are range of the loop keyframe is closer to the covisibility
then further verified using geometry-based validation. keyframes than the non-loop keyframes varies greatly.
We set 20 different thresholds for the loop detection Based on these observations, we believed that
of monocular images and panoramic images in the applying a fixed threshold to all keyframes for loop
Kashiwa dataset [23], respectively. The process of loop detection is not optimal to deal with complex scenes.
detection does not include geometry-based validation. Therefore, we propose a distance filtering method
Fig. 4 describes the comparison of recall obtained from based on an adapted threshold.
panoramic and monocular images. It can be seen that
the recall obtained by using panoramic images is
higher, which means more loops can be detected by
using panoramic images.
34
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
greatly thus further proved the fixed threshold method maps of the robots are merged into the global
to be sub-optimal. consistent map, the above RT matrix is applied to add
the subsequent keyframes and map points of the robot
into the global map. In this way, the global consistent
map can be built continuously.
4. Evaluation
4.1. Datasets
Fig. 8. The comparison of adaptive threshold and fixed Table 1. Comparison of detection results between adaptive
threshold. The blue line represents the calculated adaptive threshold and fixed threshold.
threshold of the keyframes with loops. The orange line
represents the fixed threshold. Detected Precisio
Recall
loop n
Adaptive
51 64.56 % 100 %
3.4.3. Map Merging threshold
Fixed threshold 13 16.46 % 100 %
Loop detection is in charge of detecting loop
matched pairs between different sub-maps. After
detecting a loop, a 4 4 matrix of the loop matched 4.3. Cooperative SLAM System Performance
pairs are obtained by RANSAC iterations, like (3). It
represents a similar transformation between the pose of Fig. 9 shows the designed cooperative SLAM
the current keyframe and the pose of the detected loop system based on two robots and merging two sub-maps
keyframe. Using M we can obtain the relationship in the same trajectory direction (a) and different
of rotation and translation between the poses of these trajectory direction (b). The red arrow indicates the
two keyframes. direction of movement of robot1. The yellow arrow
indicates the direction of movement of robot2. The
R T purple circle represents the loop detected between
M (3) robot1 and robot2. The central server uses the
0 1
transmitted data to reconstruct the sub-map (1) and (2)
Then, we can use the (3) to merge the sub-map corresponding to each robot. Then loop detection with
corresponding to the two robots by a rigid adaptive threshold will detect loops in each sub-map.
transformation, like (4). It transforms every keyframe In the experiments, both of the loops in the same
and map point in two local coordinate systems to a trajectory direction and different trajectory directions
global coordinate system. are detected. Then, the two corresponding sub-maps
are merged, and the global consistent maps are
M M 𝑀, (4) constructed (a)-(3) and (b)-(3).
Table 3 shows the comparison of time-cost between
where M and M are the poses of the keyframes and single robot mapping and cooperative mapping. The
the coordinates of the map points in the two coordinate mapping of a single robot used the frames from 1 to
systems respectively. 280, and it consumed 161.915s. In the cooperative
The global consistent map can be obtained by system, robot 1 used the frames from 1 to 140 and
merging the sub-maps of the two robots. Once the sub- consumed 84.9719s. Robot 2 used the frames from 141
35
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
to 280 and consumed 76.751s. The central server seen that the central server without feature extraction
received the data from robot-1 and robot-2 and and tracking has very high computational efficiency. It
reconstructed the map, run loop detection, and merged is more efficient to use multiple robots to map in the
sub-maps. It consumed a total of 10.084s. It can be same scene.
Fig. 9. Cooperative SLAM system based on two robots and the merging of two sub-maps
in the same trajectory direction (a), and different trajectory directions (b).
In terms of accuracy, we use APE (Absolute Pose map. We improved the method by keyframes, and map
Error) as the metric. Table 2 shows the APE points were added. Our system can associate
comparison between single robot and multi robots. We keyframes and map points with existing maps even
can see that compared with single robot mapping, when packets are lost in network communication. We
cooperative mapping can reduce the cumulative error also improved the loop detection part of panoramic
and increase the accuracy of mapping. SLAM. We propose three adaptive thresholds instead
of fixed thresholds to filter candidate loop keyframes.
Experiments show that the filtering effect of the
Table. 2. Various parameters of APE(m). adaptive threshold is better than that of the fixed
threshold. The adaptive threshold can be used to find
rmse mean std sse more loop matching pairs and correct loops. Both of
Single Robot 0.83 0.69 0.46 45.23 them improve the efficiency and robustness of loop
Multi Robots 0.24 0.21 0.11 3.28 detection. Experiments show that our system can
complete multi-robot transmission and map merging
well in the same and different trajectory directions.
Table. 3. The comparison of time-cost between single robot
mapping and cooperative mapping.
36
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
the International Symposium on Mixed and Augmented [15]. Arroyo R. et al., Bidirectional Loop Closure Detection
Reality (ISMAR), IEEE, 2007, pp. 225-234. on Panoramas for Visual Navigation, in Proceedings
[5]. Raúl Mur-Artal, and Juan D. Tardós, ORB-SLAM2: an of the IEEE Intelligent Vehicles Symposium (IV), 2014,
Open-Source SLAM System for Monocular, Stereo 14466969.
and RGB-D Cameras, IEEE Transactions on Robotics, [16]. M. Cummins and P. Newman, Appearance-only
2016. SLAM at large scale with FAB-MAP 2.0, Int. J. Robot.
[6]. E. Rublee, V. Rabaud, K. Konolige, and G. Bradski, Res., Vol. 30, No. 9, 2011, pp. 1100-1123.
ORB: an efficient alternative to SIFT or SURF, in [17]. R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and
Proceedings of the IEEE International Conference on J. Sivic, NetVLAD: CNN architecture for weakly
Computer Vision (ICCV), Barcelona, Spain, November supervised place recognition, in Proceedings of the
2011, pp. 2564-2571. IEEE Conf. on Computer Vision and Pattern
[7]. D. Galvez-López and J. D. Tardós, Bags of binary Recognition, 2016, pp. 5297-5307.
words for fast ´ place recognition in image sequences, [18]. H. Jégou, M. Douze, C. Schmid, and P. Pérez,
IEEE Transactions on Robotics, Vol. 28, No. 5, 2012, Aggregating local descriptors into a compact image
pp. 1188-1197. representation, in Proceedings of the IEEE Computer
[8]. Wang Y., Cai S., Li S.-J., Liu Y., Guo Y., Li T., Cheng Society Conference on Computer Vision and Pattern
M.-M., CubemapSLAM: A Piecewise-Pinhole Recognition, 2010.
Monocular Fisheye SLAM System, in Proceedings of [19]. N. Sünderhauf, S. Shirazi, A. Jacobson, F. Dayoub,
the 14th Asian Conference on Computer Vision, Perth, E. Pepperell, B. Upcroft, and M. Milford, Place
Australia, December 2–6, 2018, pp. 34-49. recognition with ConvNet land-marks: Viewpoint-
[9]. Urban, S., Hinz, S., MultiCol-SLAM-A modular real- robust, condition-robust, training-free, in Proceedings
time multi-camera SLAM system, 2016. of the Robotics: Science and Systems Conf., Rome,
arXiv:1610.07336. Italy, Jul. 2015.
[10]. Ji, S. et al., Panoramic SLAM from a multiple fisheye [20]. Krizhevsky, I. Sutskever and G. Hinton, ImageNet
camera rig, ISPRS Journal of Photogrammetry and classification with deep convolutional neural networks,
Remote Sensing, 159, 2019, pp. 169-183. in Proceedings of the Annual Conference on Neural
[11]. Leung K. Y. K., Cooperative Localization and Information Processing Systems, 2012, pp. 1090-1105.
Mapping in Sparsely-Communicating Robot Networks. [21]. Garg S., N. Suen De Rhauf and M. Milford, Don't Look
Ph. D. Dissertation, University of Toronto, Toronto, Back: Robustifying Place Categorization for
ON, Canada, 2012. Viewpoint- and Condition-Invariant Place Recognition,
[12]. Saeedi, Sajad, Trentini, et al., Multiple - Robot in Proceedings of the IEEE International Conference
Simultaneous Localization and Mapping: A Review, on Robotics and Automation (ICRA’ 2018), 2018.
Journal of Field Robotics, Volume 33, Issue 1, 2016, [22]. Huang, A. S., E. Olson, and D. C. Moore, LCM:
pp. 3-46. Lightweight Communications and Marshalling, in
[13]. Forster S. Lynen, L. Kneip and D. Scaramuzza, Proceedings of the IEEE/RSJ International
Collaborative monocular slam with multiple micro Conference on Intelligent Robots and Systems (IROS),
aerial vehicles, in Proceedings of the IEEE/RSJ Int. 2010.
Conf. on Intelligent Robots and Systems, 2013, [23]. Kashiwa (http://gpcv.whu.edu.cn/data/panslam.html)
14002391.
[14]. P. Schmuck and M. Chli, Multi-UAV collaborative
monocular slam, in Proceedings of the IEEE Int. Conf.
on Robotics and Automation, 2017, 17058387.
37
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(009)
Summary: In this paper we will supplement the declining bee population by the development of a small drone that is
able to autonomously pollinate flowers. The proposed solution uses a DJI Tello drone carrying a Maix Bit processing board
capable of running all processing on-board. Additionally, the drone is equipped with a small color camera and a distance sensor
to detect and approach the flower. We developed a two stage algorithm that is able to detect the flower, steer the drone towards
the flower and makes the drone touch the flower to pollinate it.
38
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
drone can autonomously fly towards the flower and low power Kendryte K210 hardware platform which is
precisely pollinate the flower. able to run quantized Mobilenet based networks
We investigated many hardware platforms that mounted on a commercially available DJI Tello Talent
would fit our criteria for on board computation of Deep drone. A two stage flower approaching technique is
Learning based computer vision tasks. Our developed to successfully touch and pollinate the
requirements for the platform were the following: flower. In the first stage, we detect the flower from a
1. The platform should be light weight; distance between 8 and 0.8 m using a convolutional
2. The platform should be low-power; detection network. In the second stage (0.8 - 0 m) we
3. Easy to deploy and integrate with an existing use an end-to-end network to predict the center of the
drone; flower in the camera feed. To guide the drone towards
4. Able to run state-of-the-art computer vision the center of the flower, a PID control loop is used to
algorithms with a sufficiently high frame rate. convert the location of the flower to steering
For the most part this rules out higher power mini- commands for the drone. In this section we explain all
computer devices such as the NVIDIA Jetson series. of these components one by one.
The STM32F746 or other ARM Cortex M7 based
platforms provide a deep learning library (CMSIS-NN) 3.1. Flower Dataset
to run deep learning applications on the CPU
architecture. The Ambiq Apollo3 Blue development To train the computer vision neural networks, we
board is able to run Tensorflow Lite models mostly created an artificial dataset based on images of real
targeted at voice and gesture recognition not computer sunflowers, artificial (synthetic) sunflowers and
vision. The Greenwave GAP8 is a RISC-V processor virtually rendered sunflowers. An important aspect we
optimized to run deep learning models in a multi take into account is that for the drone to successfully
threaded way. The Kendryte K210 processor is also pollinate a sunflower, it should approach the sunflower
based on a RISC-V core, but in addition to that also in a line perpendicular to the sunflower. In order to be
contains a deep learning accelerator that is able to run able to do this, we need to estimate the angle of gaze
YOLO [13] models at up to 20 fps. Because of its orientation of the sunflower to enable us to correct for
performance and relative light weight, we decided that this when we are planning the drone's trajectory. This
the Kendryte processor best fits our use case. means that in each image the angle of the flower w.r.t.
From a machine learning point of view, we need the camera should also be annotated.
models that are able to do two types of tasks: detection For the virtual flowers, we simply render them at
and classification (direct steering). These areas are a different frontal angles. For the real and synthetic
very active field of research, and many pipelines are sunflowers, we made a recorder device that is able to
available, each with their own trade-offs. Discussing rotate the flower automatically while taking pictures of
all current techniques would lead us too far from the the flower. The generated images are then pasted on
scope of this paper. Instead, we will discuss the images from the Places dataset, more specific on
techniques that are available on the chosen platform images from Japanese gardens [17] to use as
here. The Kendryte processor has support for the background. A total of 5660 annotated images were
Mobilenet [11] family of backbones. This is a generated this way. An example of these images can
backbone especially optimized for low-power mobile be seen in Fig 1.
devices, and is able to save a tremendous amount of
computing power by using depthwise convolutions.
Since the publication of the Mobilenetv1 paper, many
additions to the network have been proposed as well
[12]. However, due to the platform we choose, we are
limited to Mobilenetv1. The classification head uses a
simple cross entropy loss. In the detection stage we use
a modified detection network, together with a
Mobilenet to do fast inference. The detector makes
bounding box predictions with only a single pass over
the network, as opposed to region based detectors like
Faster-RCNNN [14] and Masked-RCNN [15]. Due to Fig. 1. Left: 3D generated sunflowers for end-to-end
their nature, generally speaking, single stage detectors approach. Right: Generated images for direct visual
are faster than region based ones. servoing approach.
The authors in [16] did a complete survey on the
current state of object detection algorithms. 3.2. Hardware Platform
3. Autonomous Drone Pollination Since the precise steering the drone has to perform,
processing of the video feed should happen on-board
In this section we go further into detail about the such that the delay between receiving images and
inner workings of our pollination drone. We create an correcting the drone's position is minimal. This implies
artificial sunflower dataset containing images of real, the need for a small and light-weight processing board
as well as artificial and virtual sunflowers. We use the capable of performing real-time image processing. We
39
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
choose to use a Maix Bit development-board which dataset. We altered the CNN architecture to
contains a Kendryte K210 (RISC-V) processor. This additionally output the angle of the flower. This
board measures 53 mm x 25 mm and weights 25 g. predicted angle is used to steer the drone towards a 0°
Furthermore, this board is equipped with a camera, angle w.r.t. the flower. The position of the flower is
resulting in a minimal image receiving delay. used to steer the drone such that the flower is always
The camera feed is directly read by the Maix Bit on in the center of the drone’s view. In this stage the size
which we either run the end-to-end network or the of the detected flower is used as a distance
detection network. The Maix Bit communicates with measurement. When the drone approaches the flower
the Tello drone via its on-board ESP32 and the size of the flower exceeds 60 % of the frame-
microcontroller. In addition to the camera, we also height, the flower becomes too big to be reliably
have a ToF depth sensor (VL53L1X) to measure the detectable and we switch to the second visual servoing
distance from the flower in the final approach. A stage.
detailed overview of our setup is shown in Fig. 2.
Fig. 3 shows the drone carrying the Maix Bit. 3.3.2. Direct Visual Servoing
40
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
41
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(010)
“They got my keys!”: On the Issue of Key Disclosure and Data Protection
in Value Chains
A. Mosteiro-Sanchez 1,2, M. Barcelo 1, J. Astorga 2 and A. Urbieta 1
1
Ikerlan Technology Research Center (BRTA), Arrasate/Mondragón, Spain
2
University of the Basque Country, Bilbao, Spain
E-mail: [email protected]
Summary: Value chains exchange massive volumes of data. Interconnection between companies makes them vulnerable to
value chain attacks and data breaches. Information is encrypted before uploading it to shared databases, and its hash is stored
in distributed data solutions like Distributed Ledger Technology (DLTs) to ensure data auditability. Key secrecy is vital for a
security system based on an encryption algorithm. However, keeping encryption keys secret is vital to the system’s security.
Thus, accidental disclosures compromise the system security, as does the lack of a periodic key renovation system. This paper
presents a solution based on Ciphertext-Policy Attribute-Based Encryption (CP-ABE). Our proposal enforces key renewal
after each security event and guarantees it via time-based encryption, performed by Smart Contracts. We thereby ensure users
refresh their keys periodically, attackers do not access new information once the key disclosure is discovered, and that original
information does not need to be re-encrypted.
Keywords: Value chain, CP-ABE, DLT, Smart contract, Industry 4.0, Key disclosure.
42
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
obtained by nefarious means, accidentally disclosed Encryption ensures information confidentiality but
keys or legitimate keys that should have been renewed. does not procure user identification. Therefore, even if
Traditional key revocation systems require lists the data breach is discovered, there is no mechanism to
identifying secret keys that cannot be used again. prevent attackers from reusing the stolen keys. Current
However, this implies that at some point, the encryptor key revocation systems rely on lists to record revoked
and decryptor must negotiate new keys to continue the keys. Each time a data consumer uses their key, it is
information exchange. In addition, data protected with compared to those in the list. Although this prevents
the old keys have to be re-encrypted. Hence, the key the use of disclosed keys, it implies list management.
revocation system designed for a value chain must In addition, the keys are linked to the CTs, which must
provide a solution where IIoT devices do not need to be re-encrypted. A new symmetric key must also be
renegotiate keys or re-encrypt information. generated and securely transmitted in symmetric
This paper proposes a time-based encryption encryption systems. In the case of asymmetric
system based on CP-ABE to protect the system from encryption algorithms, the encryptor and decryptor
key misuse. Users must own secret keys generated have to renegotiate the key pair.
after a particular timestamp that reflects the last Every key disclosure implies the risk of attackers
detected security event. This way, users with invalid using stolen keys to access the system. Therefore, there
keys or old keys are automatically denied access to is a need for a revocation system compatible with one-
information and have to authenticate themselves again. to-many encryption systems and which does not have
The solution foregoes the revoked key lists used by to re-encrypt every CT generated during the system
traditional revocation systems. In addition, the IIoT lifetime. The system must also not renegotiate keys
devices do not need to reencrypt the information, nor between the IIoT device that encrypted the information
do they have to renegotiate secret keys with value and the chain member that retrieved it. To this end,
chain partners. applying a new encryption layer whenever data is
This paper has been structured as follows. retrieved can limit key abuse. In this paper we propose
Section 2 of this paper defines the identified issue with to link this encryption to a security timestamp
traditional key revocation methods. Section 3 reflecting the last known security event. This
describes the proposed solution, and Section 4 guarantees that the system rejects expired or leaked
summarizes the conclusions of the proposal. keys without maintaining a revocation list. The
solution, based on CP-ABE, is explained in the next
section.
2. The Issue of Key Disclosure
43
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
the shared database, they request a secret key (SK) to the CT, the PPs, and the system restoration timestamp
an authority. The authority authenticates the partner, (ETMSP). The SC uses ETMSP to define an access policy
timestamps the request, and considers the timestamp such that AP = (TKTMSP > ETMSP). The SC uses this AP
(TKTMSP) as the partner’s attribute. Thus, the system and the PPs to encrypt the CT requested by the
generates each user’s secret key based on the TKTMSP consumer. The newly encrypted CTtime is then sent to
of their key request. Previously, the system has the consumer. Consumers who do not comply with the
generated another timestamp reflecting the last system AP defined by the SC cannot obtain the original
restoration after a security event (ETMSP). Using CP- information. The encryption and decryption processes
ABE properties, the requested information can be are detailed below and are depicted in Fig. 3.
encrypted at the time of the request according to a
policy that requires a secret key more recent than the
security event. This way, users with old keys or known
disclosed keys cannot access the information. Instead,
they have to identify themselves again and authenticate
their identity to the authority to obtain a new key.
44
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
4. The Smart Contract sends CTtime back to the disclosed key misuse. Finally, IIoT devices are not
consumers. overloaded since they do not have to re-encrypt
5. Consumers decrypt CTtime using their key. If TKTMSP information or renegotiate keys.
fulfils the condition expressed in the AP, they
access the information. Meanwhile, they recover ⊥
if the key is older than the last event. Acknowledgments
Thus, consumers must request a new key from the This work was financially supported by the
AA whenever they recover ⊥. This strategy ensures European commission through ECSEL-JU 2018
that consumers keep their keys up to date, and the program under the COMP4DRONES project (grant
authentication system prevents attackers from agreement Nº 826610), with national financing from
obtaining valid keys. Furthermore, IIoT devices are France, Spain, Italy, Netherlands, Austria, Czech,
exempt from further interactions beyond the original Belgium, and Latvia. It was also partially supported by
encryption, as they do not have to renegotiate keys or the Ayudas Cervera para Centros Tecnológicos grant
re-encrypt information. And since at no point is the of the Spanish Center for the Development of
original CT decrypted by anyone other than the Industrial Technology (CDTI) under the project
consumer, data confidentiality is guaranteed during the EGIDA (CER-20191012), and by the Basque Country
whole process. Government under the ELKARTEK program, project
TRUSTIND - Creating Trust in the Industrial Digital
Transformation (KK-2020/00054).
4. Conclusions
45
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(011)
Summary: Industry 4.0 has established new standards in manufacturing and quality processes, which are mainly based on the
type and efficiency of the different components of an industrial system. Virtual commissioning (VC) is an approach that allows
the user to propose a commissioning solution using computer models of industrial systems to speed up and improve the
traditional processes. This paper presents a VC solution for an industrial automotive assembly line process using toolchains of
standard and off-the-shelf software and hardware components, including physical PLC and HMI. The virtual assembly line
environment was developed using digital mock-ups of industrial robots, smart conveyors, controllers and sensors. The
toolchains share most of their components, virtual and physical, and can be classified as pure software or software and hardware
applications. The connection between the elements of the toolchain is done using industrial protocols like Modbus TCP and
APIs, ensuring complete data transfer. After several tests, the results have shown that by performing the simulated process
with the VC platform, it is possible to verify automation hardware and software, with a representative digital robotic cell,
thanks to the applied simulation technique.
46
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
on the phases within the workflow from engineering the real world is expected to be in the control
until real commissioning, one might need to implement components area of the shopfloor.
different simulation technologies. Because of the
previous, we show an implementation of three
simulation approaches permitting to validate
requirements during different project phases. The
targeted architecture can be shown in Fig.1. The
modeling and simulation tools are based on:
ControlBuild: For functional and behavioral
modeling of the mechatronic components of the
automotive assembly station
Delmia: For modeling 3D kinematized
components.
47
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
To V&V the automation process, we focused on the communication is done via a coupling driver
use of two validation environments: Siemens and respectively, using the PLCSimAdv API of Siemens,
Schneider automation software. According to Fig. 4, for TIA Portal, and Modbus TCP for ControlExpert.
there are two toolchains permitting to perform both Additionally, in order to associate the variables
SiL, and HiL techniques. On the upper right of the between the PLC and ControlBuild, output and input
toolchain in Fig. 4, the system to be commissioned is a mapping tables are created in order to properly match
Siemens virtual PLC 1515F-2 PN, and a virtual HMI the variables, ensuring thus, a bidirectional
KTP700F. These automation components are communication. It is worth noticing that safety and
connected to the simulation platform via shared control signals, as well as the information generated by
memory, permitting to exchange data bidirectionally, simulated sensors in Delmia and Controlbuild provide
to V&V the programs and routines computed in the the required data for closing the control loop with the
virtual PLC and HMI through SiL. On the bottom right PLCs (physical or simulated), meaning that the proper
of the Fig. 4, a physical HMI is connected to the mapping of these variables along the toolchain is in
simulation platform. In this case a virtual PLC high regard and data loss must be minimized.
Modicon 580 and a physical HMI Harmony STU855
from Schneider Electric are used. The communication
between the simulation platform and the execution
layer based on Schneider automation software is done
via Modbus TCP. From the system setup, it has been
accomplished the end-to-end communication and
digital continuity from the simulation to the execution
layer, permitting thus to improve the commissioning
process for an automotive scenario.
4. Conclusions
48
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
49
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(012)
Summary: With the possibility to interconnect almost every systems, the Industry 4.0 has brought a lots of possibilities to
improve lead-time and cost management while dealing with the design or the evolution of assembly lines. New tools and
concepts are developed, such as Virtual Commisionning or digital twins, to improve the process to upgrade an existing
assembly line or create a new one. These tools allows to design, simulate and validate changes without implying costly and
time-consuming tests on real infrastructures. This paper describe an approach, and its implementation, to facilitate the use of
Virtual Commisionning tools. It uses standard OPC UA communication and information modeling to implement a model
driven and hardware agnostic approach to be the boundary between the virtual and the physical world. The solution can reuse
models, such as AutomationML, from the design phase, can be used with simulation tool to develop and test the
implementation in a virtual environment, and be used in the target hardware infrastructure for integration tests.
Keywords: Cyber-physical Systems, Model transformation, Design principles in Industry 4.0, Smart Factories, Smart
Manufacturing and Technologies, Digital Production and Virtual Engineering.
50
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
4.0), as the preferred protocol for its communication the same manufacturer, they will all define the type
layer [7]. representing this robot. The first approach would be to
Modern factories can now use OPC UA as differentiate types depending on their source server.
communication protocols at every level of the process. Using this approach, there is no risk to mix different
Recent PLC are OPC UA servers, gateways exist to versions of the same type, or 2 types named the same.
transform classic industrial communication protocol to The second one would be to merge common types,
OPC UA, and software can be fed with data using their because the type of the 3 robots is the same, it should
OPC UA clients. For those reasons, data acquisition be defined and handled so. Both approaches are
using OPC UA looks like a foregone conclusion. possible based on configuration of the aggregator.
OPC UA aggregation servers have been studied for
a few years now, and several sample exist to serve as a
starting point for data collection. As described in [8] 1.2. Data Modeling
several challenges must be faced to successfully setup
an OPC UA aggregation server. The steps identified IEC62541 defines OPC UA’s meta model [4, 9].
for a generic approach of aggregation are: This meta model defines the way to represent data
1. Type aggregation: consists in getting all the types using an object oriented model, see Fig. 1. In OPC UA
from an underlying server. To do so, the Types model, objects are composed of objects, variables and
node of the underlying server is browsed methods. The model describes types as well as
recursively, and the result is merged with already instances.
known types in the aggregation server.
2. Instance aggregation: consists in getting all
instances in the underlying server. Nodes
aggregated from source servers are then evaluated
with specified rules to be put in the right place in
the aggregation server’s address space.
3. Service mapping: In the proposed architecture, the
service calls on the aggregation server are
forwarded to the source server on which the node
concerned by the call is located.
This generic approach lays the foundations of the
construction of an aggregation server, and is used in
implementations, such as the sample application of the
OPC foundation. However, it has some drawbacks:
In a vertical multi-level aggregation, the Fig. 1. OPC UA Object model.
propagation of the service call from the client to the
base source server might be long and cause latency The set of existing types and instances are
or timeout. represented in a server as a graph, with nodes (types
There is no local version of nodes. Every request and instances) bound by references, see Fig. 2. Nodes
must be forwarded to the underlying source server. have attributes, that define what they are, and
If a group of nodes must be read by several clients, references that link them to other nodes.
it would generate a lot of unnecessary Read
requests to the underlying server. That can be a
problem on some industrial networks, which
should not be saturated.
To deal with these problems, a second kind of
aggregation (Data Warehouse mode) has been
implemented, based on the OPC foundation’s sample
aggregation server. Basic steps defined previously still
remain true, the server starts by the discovery of types
contained in underlying servers, then it gets instances,
and services are mapped with underlying server’s
services when it is necessary. Nodes are copied in the
aggregation server’s address space, and subscriptions
to corresponding nodes are created. Doing so, some
services, such as Read do not need to be mapped, and
propagated. For some services like method call, there
is no other way than propagating the call to obtain the Fig. 2. OPC UA Node model.
desired behaviour.
Another limit to be tackled with is the redundancy An OPC UA server can load models using 2
of namespaces. Some namespaces are defined by standard methods. The first consists in loading a
vendors, and duplicated in each of their devices. For standard Nodeset2 file that describe the address space
instance, if a shop floor is equipped with 3 robots from of a server. The second is based on code. UA server
51
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
solutions generally comes with a code generator that paradigm. The use of AML format by OPC UA server
converts models to usable code (specific to the server). has been standardized [12, 13]. It defines a way to
Both ways are complementary and have pros and cons. transform AML file in OPC UA model loadable by an
Modeling is an important point of OPC UA OPC UA server.
specification and its meta model allows users to model However, the wide variety of existing models,
any of their industrial assets. On top of the meta model whether a standard translation exist or not, are only
defined by the OPC UA specification, some working models and do not define how the data and the model
groups have formed to standardize the use of OPC UA are bond.
modeling in their fields. Those initiative gave birth to
companion specifications. The list of these
specifications [10] includes several information 2. Architecture
models aiming at the description of machines such as
OPC UA for Machinery, OPC UA for Robotics, OPC To try to go a step further and solve issues
UA for MachineTools, etc. Some of these models are mentionned in Part 1, using a generic approach, the
dedicated to modeling language adaptations such as I4SSM (Industry 4.0 Semantic and Security
OPC UA for I4 Asset Administration Shell, Management) has been developed. The main piece of
ISA-95 Common Object Model, OPC UA for the solution consists in an OPC UA server, which
AutomationML, etc. implements an architecture with 3 functions, described
AutomationML (AML) is a data exchange standard in Fig. 3:
[11] used for production system engineering. It gives a Aggregation.
way to exchange data between applications to model OPC UA operation server.
the system and represents it using an object oriented Semantic layer.
2.1. OPC UA Operation Server The operation server also manages the loading of
models. They can be loaded using 2 formats: Nodeset2
The operation server is a standard OPC UA server or compiled libraries.
which exposes and manages interfaces to configure the The main difference with state of the art servers
aggregation module and the semantic layer. [14] relies in the fact that an interface and reflection
The operation server handles OPC UA client calls mechanism is used here so that a compiled model can
and handles it with the help of other components if be used without having to rebuild the whole server.
needed.
Acquisition of data is managed by the aggregation
module, which discovers and updates nodes from 2.2. Aggregation Layer
underlying servers. These nodes are exposed to clients
using the operation server. The link between both The aggregation layer of I4SSM is designed to be
modules is done via an internal event and messages generic. The implementation of the aggregation
mechanics. module is based on the conclusions drawn from
52
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Part 1.1. This article focuses on the Data warehouse particular aspect, a collaboration with Intel has enabled
mode as default implementation, which creates a local the use of different hardware configuration enhanced
equivalent to nodes discovered from underlying data with software or hardware add-ons and accelerators.
sources. Moreover, the aggregation module exposes an Some common actions have been implemented on a
interface to develop specific acquisition drivers, in test environment facilitating test execution. This
case a non OPC UA connection to data source is environment enabled the possibility to test the
required. solution’s scaling, both horizontally (more aggregated
As for previous implementation [8], the servers) and vertically (more layers of aggregation
aggregation server is an OPC UA client, and uses servers) to ensure the possibility to use the solution
Browse standard service to discover its underlying with an architecture as close as possible from real-life
source server address space. It begins with the applications. The test of the functionalities of I4SSM
discovery of types and continues with instances. both separately and in parallel has given feedbacks on
Discovered types are added to aggregation server’s the implementation of the solution. For instance
types, i.e. they are added in the aggregation’s server Table 1 shows improvement of performances between
Type tree. Then the server creates a copy of underlying campaign 1 and 2, concerning the number of servers
servers instances, and organizes it in the aggregator’s that can be aggregated. As a result, the feedback on the
address space by putting them in folders corresponding first 2 tests campaigns has reinforced the need to
to their source server, as shown in Fig. 4. include in the project life cycle the stress tests to
anticipate performances issues and find bottlenecks in
order to facilitate code optimization by design.
5K VARIABLES
Source servers 1st campaign 2nd campaign
1 100 % 100 %
2 100 % 100 %
3 60 % 100 %
4 0% 100 %
5 0% 100 %
10 0% 100 %
15 0% 100 %
53
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
link and implementing its associated behavior. This still be up and working in case of a server restart.
mechanism is implemented so that a simple However, this matter is not trivial to solve because of
configuration of the server is needed to add a type of the automatic discovery of underlying servers address
link to the server, no rebuild of the server is required. spaces.
Nodes discovered in underlying sources are
mapped with a local node, which NodeId is uniquely
generated in the server. The mapping table and the
NodeIds must be persisted, so that it can be retrieved
after a restart or a crash. It is important, so that existing
semantic links can find their sources and targets back.
The same is also true for external client applications
that often store NodeIds and base their internal
mappings on that.
To be able to set the server in its prior shutdown
state, the following items of interest have been
identified:
Nodes added with the AddNode service.
Nodes updated with the Write service.
References added with the AddReferences service.
User roles.
Sematic links.
Aggregated nodes’ NodeId mappings.
These elements are identified as key elements that
Fig. 5. SemanticLinkSet in I4SSM address space. need to be persisted. Nodes added to the server must
be recreated, as well as the references. Updates on
nodes must be persisted to be able to retrieve
Table 2. SemanticLinkType definition. parameters as they were before the stop of the server.
For example a parameter that defines an alarm
SEMANTICLINKTYPE ATTRIBUTES threshold should not be rewritten by the user when the
Attribute Value server restarts. The same idea is also true with user
BrowseName SemanticLinkType roles configuration.
IsAbstract True
SEMANTICLINKTYPE REFERENCES
Modelling
3. Use case: Virtual Commissioning
References Node Class BrowseName
Rule and Automotive Cell
SubtypeOf ObjectType BaseObjectType
3.1. Automotive Cell Components and Process
HasComponent Method AddSource Optional
HasComponent Method AddTarget Optional
The Virtual Commissioning simulation platform,
HasComponent Method RemoveSource Optional described in [15], is composed of two software
HasComponent Method RemoveTarget Optional allowing to perform functional models of mechatronic
components of the station (ControlBuild), and a 3D
visualization environment for the kinematized
The modeling and implementation using OPC UA components (3DX Delmia). Additionally, these two
meta model offers 2 major advantages: a standard way simulation tools are connected via OPC UA to the
to design and model semantic links, and the possibility I4SSM server introduced previously. Fig. 6 shows an
to use a simple OPC UA client to add, remove and overview of the proposed toolchain.
modify semantic links in the server. Indeed, any OPC
UA client that can perform method call is able to
manage semantic links. That way, scripting can be
used to massify the creation of links, to fit better with
real use case in which thousands of semantic links are
needed to map data from a production line to its high
level model.
54
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
The automotive cell consists in several components This use case demonstrates the possibility to create
allowing to emulate an assembly process for the body a digital twin of an automotive assembly line, using
sides of a car. The assembly station is composed of OPC UA and taking advantage of a unified model. In
mechatronic components, such as robots, sensors and this scenario, Delmia is responsible for the 3D
safety components, as shown in Fig. 7, and listed as visualization of the line, ControlBuild manages its
follows: simulation. I4SSM is responsible for implementing the
Car body hardware abstraction layer, by the exposition of a
Car conveyor model that the simulation tool can understand, and by
Rotational buffers linking this model to the real data. With that
6-Degree-of-Freedom robots abstraction, simulation can be performed and
Presence sensors visualized in 3D, moreover tests can be done using
Safety components, such as safety gates and light virtual or real PLCs, without changing the
barriers. configuration of the simulator. Using this, engineers
can virtually simulate a change of parameters, validate
a new PLC program using virtual PLC and even test a
modification on the physical environment.
4. Conclusions
55
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
56
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(013)
Summary: The relevance of Artificial Intelligence (AI) for manufacturing and service operations in the industrial sector is
growing rapidly. Various economic actors, including small and medium-sized enterprises (SME) that are the backbone of the
European economy, are already well aware of specific use cases of AI in the industrial context and demand effective AI
solutions to their challenges. However, existing approaches and models for identifying the appropriate AI use case for
efficiency improvement potentials can cover the growing market demand to a limited extent only. Uncertainties about the
actual capabilities, a lack of a strategic focus, missing success metrics, poor data quality, or false expectations of AI systems
can all lead to failure of AI projects. To counter this, an AI maturity assessment can assist companies to overcome challenges
on the way to a sustainable planning, deployment and adaptation of AI in industrial applications. This paper examines the need
for a holistic method and presents the Fraunhofer IFF AI Maturity Model – a method based on the needs of SME in the
manufacturing industry.
Keywords: Artificial intelligence, AI-supported business model innovation, Use case-specific assessment, Maturity level,
Feasibility analysis, Decision support.
57
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
As one of the objectives of application-oriented which existing or easily captured customer data or
research, it is necessary to identify the gap between the process data combined with AI methods form the basis
state of knowledge and state of implementation of AI for the development of new value propositions. If these
and its economic possibilities within the SME are integrated into corresponding revenue mechanics
landscape. This gap is due to existing barriers to use and value chains, innovative, AI-based business
AI. A survey of AI experts by The Scientific Institute models can be developed.
for Infrastructure and Communication Services However, to realize such business models, in
revealed that a lack of expertise and skilled workers is addition to domain knowledge – the knowledge of how
a strong (36 %) or very strong (64 %) obstacle to the the industry works, what customers want, and the
implementation of AI solutions in SME established processes – methodological knowledge in
(Wissenschaftliches Institut für Infrastruktur und the context of AI is required above all. Therefore, it is
Kommunikationsdienste 2019). This reflects the extent necessary to identify the status quo of the resource
of the general shortage of skilled workers in SME: IT endowment and requirements typical for the company
specialists in particular are rare on the labor market. size class. Therefore, the Fraunhofer IFF developed a
SME are often unable to keep up with the salaries paid third-party assessment of the current maturity level in
by large companies. order to derive necessary capacities and resources for
Inadequate availability of data in sufficient empowering employees and exploiting potential by
quantity and quality is another major obstacle. Due to means of AI. The assessment was designed with the
the smaller size of the company, the possibilities for claim to be objective, holistic and realistic. The
data collection and data utilization are more limited maturity model as a component of a use case analysis
than in large companies. Many of the SME are not is predominantly aimed at companies in the
sufficiently digitized to ensure an adequate collection manufacturing sector.
of production data and machine data, among other First of all, it is important to understand why the
things, by means of software and sensor technology. maturity model approach was preferred for making
Since the presence of suitable datasets along the value companies capable of acting in the field of AI. On the
chain is a key prerequisite for training AI systems, one hand, maturity models can be used to describe the
successful deployment of AI is directly dependent on change in the object under consideration (assessment)
the company's level of digitization. Data security in order to derive recommendations for action based on
concerns, lack of acceptance among employees, and this, which are aimed at achieving the next higher
limited financial resources are further obstacles that level. In addition to performance evaluation and
were also identified and analyzed by AI experts and improvement, maturity models are also used for
trainers. internal and external comparisons. They provide a
benchmark for best-in-class companies or the entire
industry (Bruin et al. 2005). In the corporate context,
3. The AI Maturity Model maturity models can be helpful tools, especially for the
responsible managers, to be able to classify the current
Based on these findings, concrete needs were development status of the company in relation to the
defined, which are to be covered by a methodical past and the future. Maturity models focus the
approach. The goal is to derive a standardized, holistic assessment on a specific area, but can be applied to a
procedure for identifying relevant use cases for AI broad field (Häberer et al. 2017). Figure 1 depicts the
deployment and providing a decision template. Fraunhofer IFF AI Maturity Model that represents an
Existing AI applications mostly focus on incremental development trajectory required to
increasing efficiency. However, AI offers SME greater achieve a networked and fully automated production
potential through service or product innovations, in and logistics environment with the help of AI use.
Fig. 1. The AI Maturity Model – five stages to AI maturity and their respective characteristics.
58
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
The model consists of five stages of AI integration: technologies; AI experts and interdisciplinary project
Stage 1: Manual data processing structures.
Stage 2: Networked data utilization Example: Commercial vehicle manufacturer
Stage 3: Integrated intelligence collects process data at every step of the manufacturing
Stage 4: Smart data process. Data pre-processing and initial analysis takes
Stage 5: Factory of the future place on the edge. Through standardised data exchange
The following stage characteristics were interfaces, information is exchanged between the shop
established to make a clear differentiation between the floor and central PPS system. The collected data is then
individual maturity levels so that the differences to the processed in the cloud. AI is used to forecast
next higher or next lower stage are distinct: completion times, delays and possible disruptive
1. Manual data processing: Recurring and events in production. The algorithms suggest the
manual data entry; Distributed, local data storage and optimal production planning and exchange the data
availability; Unstructured, non-value-adding data; with upstream and downstream processes (supply
Lack of a scalable system for deploying AI chain management, transport of finished products).
applications; Lack of AI strategy, skills, and resources Thus, the control of these processes is also data-based
in-house. on the actual state of production and existing
Example: A paper-based, non-automated data production state forecasts for defined forecast
capture and paper-based process control system in horizons.
screw manufacturing is present. In production 5. Factory of the future: Quantifiable value
planning and control, data evaluation is dominated by creation through data; Trustworthy and explainable AI;
humans and data availability is only given in individual Independent system monitoring and self-calibration;
processes. Enabling, interoperable and performant systems;
2. Networked data utilization: Structured and Industry-leading AI competencies and diversified AI
partially automated data acquisition; partially teams.
networked data storage; context-related and Example: High-tech electronic chip manufacturer
retrospective data evaluation; standards and optimized has fully automated production lines. Production
ICT; embryonic AI initiatives and resource allocation. planning is carried out by AI algorithms depending on
Example: A building materials manufacturer order situation, deadline chain, material availability.
considers its data to be relevant to the business because Quality control is carried out inline using current
data is evaluated on a rules-based approach and used production data, evaluation of images of the products
to support employees in their decision-making. The and semi-finished products by AI-based systems.
team leader in the construction department can Quality deviations are automatically compensated for
evaluate his shift planning in a partially automated and by the production control system. Human intervention
algorithm-supported manner and store it in a is only required in cases that are not clearly defined and
knowledge database, which the management can in exceptional circumstances. The supporting
access. processes (logistics, infrastructure) are also data-
3. Integrated Intelligence: Continuously driven and depend directly on the main process
networked data storage; Networked product and (production). Through the external data exchange,
process data; Central data management and production capacities as well as external service
availability; AI frameworks and scalable IT providers and suppliers are controlled across all
infrastructure; Internal AI competencies, specific AI locations.
budgets and project groups. Based on the assumption that one level of AI
Example: Data scientists at a turbine integration is characteristic for each stage of
manufacturing company use data mining analysis development, the individual AI levels of a use case,
techniques to evaluate and interpret extensive and with increasing maturity, merge into one another, but
complex data sets for error detection in production are not mutually exclusive. Therefore, the distinction
series. As part of a dedicated AI strategy, a data between AI levels for the respective use case levels
warehouse has already been built as an should be regarded as intersecting.
interdepartmental data storage and analytics platform. The main purpose is that an organization should not
Analysts in development, production, quality only discover at which level its areas of application are,
assurance and warranty processing have access to this but in particular what kind of solutions are goal-
data warehouse to perform ad hoc analyses and create oriented in order to get to the next higher level. This
standard reports supported by machine learning gives the organization an indication of what it needs to
algorithms. After particularly complex calculations, achieve if it wants to reach the next higher level of AI
the company can now forecast for each part and each maturity. Thus, it should not be the goal of an SME to
product variant the frequency with which certain achieve stage 5 but rather the next stage of
defects occur as a rule. development.
4. Smart data: Automated pipelines for data The maturity model does not fulfil the claim of
evaluation; Communicable, consistently available evaluating the entire company in terms of a uniform AI
data; Networked product and process structure and maturity, substantiated by a single numeral degree.
infrastructure; Configurable interfaces and cloud Instead, it aims to evaluate individual use cases within
59
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
the company, representing a business process or What areas of application are there for AI in
function that describes a sequence of individual my company and where can AI specifically be
activities that are carried out successively to achieve a helpful?
business or operational result with a discernible value Which AI tools exist and which of them can fit
(IBM 2017). As process-related components of an my business use cases?
organization, use cases are seen as complex workflows What data do I need at all and how do I need it
that are intended to contribute to business success. For to train my AI?
this reason, the objective is to identify the Secondly, operational use cases are to be identified
economically most prospective business use cases for and prioritized. On-site inspections (plant visits) and
ensuring a successful and structured introduction of AI personal interviews form the basis for a structured
within a company. approach and solution development. The goal is to
successfully carry out a company-specific use case
4. The Fraunhofer AI CheckUp determination for various alternative actions. For this
purpose, a questionnaire based on the fields of action
The maturity model is one component of an overall of AI in the industrial context was developed to enable
service that aims to make manufacturing companies a systematic evaluation of AI use cases. Since the
ready for the use of artificial intelligence – the so- maturity model is adapted for the requirements, goals
called AI CheckUp. The AI CheckUp is a method and challenges of each individual use case to be
developed for the analysis of processes to identify assessed, an assessment is always use case-specific and
suitable use cases for AI-based solutions within SME. benchmarking of use cases does not appear to be
The AI CheckUp arose from the development need of purposeful. Therefore, use case-specific and general
a modular structure to cover different scientific and fields of action were developed that are relevant to
economic requirements for AI in an industrial context: cover a company’s needs for actions in a holistic way.
Evaluation of the existing database (quality, The AI CheckUp focuses on the following seven fields
availability, etc.); of action of AI in the industrial context:
1. Quality of the collected data
Identification of improvements in the existing
2. Handling / Processing of data
data management and utilization (problem
3. Infrastructure and access to technologies and
monitoring);
tools
A credible evaluation system of use cases and
4. AI strategy and leadership
realistic references of AI solutions
5. Employees and AI acceptance
implemented to date;
6. AI evaluation and impact assessment
Manufacturer-independent orientation and
7. Data sovereignty and collaboration
consulting;
Figure 2 depicts the seven fields of action that are
Third-party assessment of the AI and data to be inquired and assessed by the experts. Each field
management competencies available in the of action has one or several central questions, intended
company through a holistic multi-method to guide the experts' data collection, but do not limit
approach (interviews, workshops, analysis of the field of action to these key questions alone. Based
the database, etc.); on the corresponding answers from the interviews with
Estimation of quantitative benefits; management, skilled workers and specialists, as well
Checklist with minimum requirements for a as the information gathered through observations and
project; process analyses during the plant tours, each of the
Identify new research needs. seven fields of action can be evaluated and classified
The approach is to compare the as-is situation with according to the level characteristics of the maturity
the requirements that an AI-based solution requires, model at level 1 to 5 for the respective use case.
identify existing data sources, and design the to-be data Hence, the AI maturity level is determined. The
workflows for the AI-based solution. Eventually, an application areas are classified objectively in terms of
identification of required resources (professional, the requirements and prerequisites for AI. The aim is
technological, methodological) and well defined to realistically assess the potential of AI to achieve
success criteria is to be established. these goals. An effort-benefit assessment of the
The first step is to raise awareness and educate the measures is carried out so that problem definitions can
employees of a company around the topic of AI and be derived and aligned with the corporate strategy. The
clarify the potentials and restrictions of AI claim is a realistic assessment of where the identified
implementation. There is also a focus on the business and prioritized use cases of the company stand in
goals and strategy in the sense of concretizing what a relation to AI and which AI-supported solutions make
company wants to achieve independently of AI. economic sense. Based on the maturity degree of the
Experts in the field impart fundamental knowledge respective use case, necessary development steps can
about AI and provide answers to questions such as: be derived and prioritized, taking into account an
Why should I get involved with AI? integrated risk assessment (such as dependencies on
How do I find out whether AI is the right partners and suppliers).
technology for me?
60
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Fig. 2. Seven fields of action of AI and their key questions for information acquisition.
What follows is an AI use case development. Based forces) (Osterwalder 2010). Approaches for
on the information intake and the determined maturity optimizing the economic mechanics of a specific
degree, efficiency improvement potentials can be business use case can thus be determined and
evaluated and individual measures for increasing the developed. As a result, new product and service ideas
AI maturity of the respective use case can be derived. as well as efforts towards process innovation can be
The measures can be of technical, organizational or initiated. The AI Use Case Canvas consists of 14
methodical nature. Together with the employees, the elements and aims to holistically visualize a specific
experts derive operational use cases with potential for use case in the company:
increasing efficiency and reducing costs, and match the 1. Value proposition: It describes the essence of
problems with possible solutions (problem-solution- the use value that the use case provides to the user.
fit). 2. Internal customer segments: Involved,
Ultimately, a decision template is developed and affected or interested internal parties and
provided. For the purpose of decision support of the organizations, which exert influence on the use case
SME in possible investment projects, the Fraunhofer and influence each other (in particular own employees,
IFF developed a management tool for systematically shareholders and service providers).
developing identified use cases further, visualizing its 3. External customer segments: Target groups
elements, and serving the decision makers as a decision outside of the company to whom the use case is
support template – the so-called AI Use Case Canvas. intended to add value (in particular customers, the
Its development is based on the Business Model general public, authorities and lawmakers).
Canvas by Alexander Osterwalder which is an 4. Business model: The as-is logic of how a
established “strategic management tool to describe company generates added value and secures earnings.
how an organization creates, delivers and captures The revenue mechanism are in the foreground here.
value” (Osterwalder 2021). Comparable approaches 5. Data situation / data quality: The data
with similar structuring principles for business process situation / data quality describes how well data are
modeling of AI applications are available on the suited to fulfil their purpose for the use case.
market, such as OwnML on the website 6. Technology and infrastructure: The set of
www.ownml.co (OwnML 2022). The Business Model tools, technologies (software and hardware),
Canvas arranges the business model elements in nine infrastructure, and methods used to provide the use
basic building blocks and establishes a uniform case with the required resources and to achieve (create)
structure, with the aim to show the logic of how a the value proposition.
company intends to make money. Similarly, the AI 7. Domain and method knowledge: Existing
Use Case Canvas was developed to describe business domain knowledge of how the industry works,
use cases in a structured way. Decision makers can customer needs, and established processes, as well as
discuss, evaluate and develop the building blocks, and methodological knowledge in the context of AI, data
discover correlations between them. This can serve as management and business models.
a template for decision support for the further 8. Key partners: The partnerships that are
development of existing and future use cases. The tool created through networking with individuals and
can be integrated into the company's strategic planning organizations that can contribute to the business
and provides an overview of a company’s business use success of the use case.
case within its external environment (market forces, 9. Key resources: The key assets for the
industry forces, key trends, and macroeconomic successful implementation of the use case, or the
61
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
resources available for the use case with which measures and solutions that employees can trust and
operations can be executed effectively. accept. By means of the identification of business use
10. Key activities: Processes that are essential for cases for AI, the analysis of fields of action and a use
the success of the business model or use case. Key case-specific classification, an evaluation of efficiency
activities describe which activities are of particular improvement potentials can be made so that individual
importance for the success of the use case. measures for increasing the AI maturity can be
11. Risks: Risks associated with the AI use case derived. In the end, an SME can benefit from decision
project, possible negative outcomes, possibility of loss support including an effort-benefit analysis as tangible
and failure. The risk-based assessment of AI describes added value. Thus, in the context of digitization, SMEs
the extent and quality of addressing system criticality can undertake an incremental development of their
as an expression of the damage potential of an business model, with manufacturers evolving from a
algorithmic system. pure product sale to an as-a-service business model, for
12. Cost structure: The main expenses and costs example, and be able to keep up with multinational
associated with this use case (in particular fixed costs companies.
such as salaries, wages, rent, and infrastructure
expenses, variable costs (proportional to the volume of
goods or services produced), economies of scale, and Acknowledgements
economies of scope).
13. Revenue streams: Revenues that can be The development of the Fraunhofer IFF AI Maturity
generated from each customer segment, such as sales Model was supported by Christian Löwke, Andreas
of goods or services, user fees, membership fees, Herzog and Lina Lau. We would like to thank you for
leasing, renting, lending, royalties, and advertising your professional contribution.
fees.
14. Key performance indicators for success of
failure: Business metrics that reflect the success of the References
use case in various areas.
[1]. Bitkom. Industrie 4.0 – so digital sind Deutschlands
Fabriken, May 19, 2020, p. 8. https://www.bitkom.org/
sites/default/files/2020-05/200519_
bitkomprasentation_industrie40_2020_final.pdf,
Accessed: 14.11.2021, in German.
[2]. De Bruin, T., Rosemann, M., Freeze, R., Kulkarni,
U., Understanding the Main Phases of Developing a
Maturity Assessment Model, in Proceedings of the 16th
Australasian Conference on Information Systems
(ACIS), 2005 p. 3.
[3]. German Engineering Association (VDMA), Leitfaden
Fig. 3. The Fraunhofer IFF AI Use Case Canvas Künstliche Intelligenz – Potenziale und Umsetzungen
im Mittelstand, VDMA Bayern, Munich, Bavaria,
2020.
The AI Use Case Canvas is developed and https://www.vdma.org/documents/34570/1052572/Le
illustrated together with the managers and specialists itfaden+K%C3%BCnstliche+Intelligenz-
Potenziale+und+Umsetzungen+im+Mittelstand.pdf/c
of the company, so that decision makers always have e38a591-68cb-9775-101e-
the possibility to continue working with this structure d7cad064b149?t=1615364023575). Accessed:
and use it as a strategic decision template. The 07.01.2021 (in German).
objective is to accompany the SME in its [4]. Engineering Lifecycle Management / 6.0.4.
transformation into a future- and AI-capable company Anwendungsfälle definieren, IBM Corporation, 2017.
and to make it capable of acting in the field of AI in a https://www.ibm.com/docs/de/elm/6.0.4?topic=requir
number of steps. ements-defining-use-cases. Accessed: 07.01.2022
(in German).
[5]. Häberer S., Kujath M., Flechtner E., Industrie 4.0-
5. Conclusions CheckUp, wt Werkstattstechnik online, wt Jahrgang
107, 2017, Heft 03 , pp. 141-142.
[6]. OwnML, The Machine Learning Canvas,
In order for companies to profitably exploit the https://www.ownml.co/machine-learning-canvas
potential of new AI technologies, it is essential to Accessed: 13.01.2022.
determine the current state of its business use cases. [7]. Osterwalder, A., Strategyzer AG, Löwenstrasse
This is where the maturity model as a component of the 2 CH, 8001 Zürich, Switzerland 2021,
AI CheckUp can provide support for SMEs. The aim https://www.alexosterwalder.com/,
of the AI Maturity Model is to support companies in Accessed: 21.12.2021.
taking advantage of the economic opportunities [8]. Osterwalder, A., Pigneur, Y. Business Model
offered by AI and machine learning in particular. In Generation. A Handbook for Visionaries, Game
order to make the most of the potential, it is important Changers, and Challengers, John Wiley & Sons, 2010,
pp. 200-210.
to pursue a human-centric AI approach and to develop
62
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
[9]. The German Federal Government. Strategie Mittelstand – Relevanz, Anwendungen, Transfer.
Künstliche Intelligenz der Bundesregierung: WIK GmbH, Rhöndorfer Straße, 6853604 Bad
Fortschreibung 2020, 2020. https://www.ki-strategie- Honnef. 2019. https://www.mittelstand-digital.de/
deutschland.de/files/downloads/201201_Fortschreibu MD/Redaktion/DE/Publikationen/kuenstliche-
ng_KI-Strategie.pdf. Accessed: 07.09.2021 intelligenz-im-
(in German). mittelstand.pdf?__blob=publicationFile&v=5,
[10]. Wissenschaftliches Institut für Infrastruktur und Accessed: 14.11.2021 (in German).
Kommunikationsdienste. Künstliche Intelligenz im
63
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(014)
Summary: The growth of personal vehicle ownership has had a direct impact on the need to accommodate vehicles into
structures that haven’t been conceived to handle such requirements of usage. One impacted area by this change is the optimal
usage of existing parking facilities, considering an increase in parking space needs, existing facilities will be filled faster than
before. While navigating empty or near-empty facilities results in a reduced effort in finding available spaces, the moment a
driver must navigate the existing infrastructure almost completely or multiple times has negative effects both economically
and for the environment. This paper presents a method of managing the position of vehicles in parking facilities based on a
multivariable approach with a load balancing component. The presented solution provides minimal functionality even if data
sources are unavailable thus allowing permanent service availability.
Keywords: Parking, Route optimization, Reduced travel time, Load balancing, distance, Dijkstra, SUMO.
64
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
2.1. Definition of Input Data lseg and 𝑠 having the same meaning as before.
The load factor is expressed using eq. 3:
The basic data inputs were considered the ones
with the least technological implication: the road ∑
infrastructure consisting of entry and exit points, the , 𝑟𝑒𝑠 1
𝑙 , (3)
connecting roads with characteristics regarding length, 0, 𝑟𝑒𝑠 1
and the total amount of parking spaces and allowed
maneuvering. where:
More complex types of data are considered for the lveh = length of the vehicles traveling on the segment;
advanced feature of the algorithm, those data types res = usage condition, the computed values is used
include access and entry times for dedicated only if the vehicle demand exceeds the available
infrastructure nodes, the number of vehicles targeting infrastructure.
the parking facility, the destination when exiting the This approach allows to inflate the cost in case the
parking facility, and the congestion of each traffic road structure usage is above the road capacity, the
element. One important factor in regards to this type of factor is only applied once the strain is above the
dynamic data is that their acquisition will increase both capacity.
the initial cost of the facility due to the sensors needed Additionally to this, to better capture the parking
or infrastructure needed to acquire them and the specific maneuver downtime on the road network, a
general computing requirements. time penalty is added for vehicles that would be
parking or exiting the parking space Δtm [s].
Thus the actual cost function becomes the one
2.2. Algorithm Description
described in eq. 4:
As previously mentioned, the developed algorithm _
can be seen as an extended version of the Dijkstra
algorithm, in the following ways: the cost function of 𝑐 𝑡 𝑑 ∆𝑡 (4)
the starting point is not zero as with the standard
approach as it reflects the entry time to the facility, it
considers subdivisions of the graph without generating Another specificity added to the algorithm is the
new point, it features the balancing component by fact that the check for capacity and infrastructure entry
increasing the cost for paths that already have traffic time is done before allowing vehicle access to the
assigned to them. given infrastructure.
In a common navigation system, the minima is The complete flow can be seen in Fig. 1.
computed based on the physical distance between the
start and the trip destination, the proposed algorithm
deviates from this definition and defines the primary 3. Simulation and Results
minima condition as the smallest amount of time spent
in traffic. This change is supported as map providers A medium-sized parking lot was designed to be
include besides the road length also information about used as a demonstrator for the deployment, in the
various speed limits, thus the first component used for initial conception phase a dedicated computer-aided
evaluation becomes the one described in eq. 1 design software (LibreCAD) was used, the output of
the projection phase is visible in Fig. 2.
𝑡 𝑙 /𝑠 , (1) The parking facility features the following
elements: a single one-way street connecting the two
where: entry and exit points. The parking lot is divided into
tseg = necessary time to traverse a map segment [s]; parking areas consisting of the parking spaces
lseg = legnth of the segment [m]; connected to a road segment. The parking space
sseg = the maximun allowed speed to traverse the dimensions were chosen to be 2.5 m for with and 5m
segment [m/s]. for length. A general speed limit of 15Km/h is
The information obtained by this step is purely considered for the traffic inside the parking facility
static and doesn't take into account the real-world while the access road has a speed limit of 50 Km/h.
environment like infrastructure use resulting in a The simulation necessities, network definition,
degradation of the projected arrival cost. static and dynamic information extraction, route
Adding the dynamic characteristic data about the generation, and vehicle control were done using the
other traffic participants is needed, this can be either SUMO (Simulation of Urban MObility) package
sourced from measurement systems or is known in features, developed by the German Aerospace Center,
case the path planning is done by a centralized this is “an open-source, highly portable, microscopic
mechanism. This component is presented in eq 2. and continuous traffic simulation package designed to
handle large networks” [10], while its primary scope is
𝑑 𝑙 𝑙 /𝑠 , (2) not the simulation of parking facilities the features it
offers were of great help in implementing the
where: simulation scenarios and solution by extending and
lf = load factor for the segment; relying upon it in the developed code.
65
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
As the SUMO tooling requires a specific input used by the solution scripting for parameters like
format for being able to simulate the environment, the junctions, direction change restrictions, segment
initial schematic was recreated in the netedit utility length, max segment speed, and the association
denoting the network nodes and the associated between the parking area and the access road. Specific
restrains. The network provides direct information to the simulation is the computation of a parking space
66
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
offset, due to the association being made on parking using the TraCI API, Traci allows this by creating a
area level with each parking space the with of the TCP-based connection to the SUMO application and
individual space needs to be accounted for when using that connection to pass data. The high-level
computing the distance travelled. Another specificity overview of the SW components used is visible in
of the simulation is the slit of the parking area due the Fig. 4. Application internals where specifically
presence of a junction, this is visible in the lateral developed for this study while existing elements are
parking areas thus increasing the number of areas that marked separately.
need to be monitored. The parking facility netedit The simulation phase considered 3 different
specific design is visible in Fig. 3. scenarios for choosing a parking space for a vehicle
next in queue:
a) Guidance based solely on proximity to the target;
b) A randomized allocation of parking spaces with
guidance instructions;
c) The proposed extension to the algorithm to balance
the network usage.
The evaluation criteria considered was the least
time spent idling, this value is taken from the
simulation and as per TraCI API [11] is defined as
“The waiting time of a vehicle is defined as the time
(in seconds) spent with a speed below 0.1 m/s since the
last time it was faster than 0.1 m/s.”
Fig. 3. Parking facility design translation for simulation.
As a general path target, vehicles always originate
in the right outmost extremity of the access road and
Currently, this step has been done manually, the their destination is a parking space within the facility.
available scripting allows the direct import only of In a direct comparison, the 3 scenarios have clear
OSM (Open Street Map) format, the authors are differences from a performance point of view, the
considering the development of a script that would time-based criteria show the ability of the developed
make it possible to do type conversions DXF file solution to reduce the general time spent in traffic, by
format to the XML file format used by SUMO. having the same demand this is visible in Fig. 5.
The control of the simulation is done from a
python-based implementation of the parking manager
67
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
The individual max idle time per solutin are presented 4. Proposals for the Physical Realization
in Fig. 6, Fig. 7 and Fig. 8. of the System
5. Conclusions
Allocation based on the minimal driving distance,
this approach generates excessive strain on the first The solution can determine based on a given
connection points to the parking area and results in infrastructure the necessary parking lot information
congestion. Waiting times observed during these and provide access control features in case the capacity
simulations were the highest while choosing a random is reached while outperforming classic navigation
parking area among the available structures yields solutions in terms of time spent in traffic.
better results than the closest available area. Future extensions should consider more simulation
Higher waiting times are still visible this is due to the scenarios and the inclusion of more complex driving
bottlenecking of the entryway. behavior (like take over).
68
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
The authors are planning after the extended weight-Dijkstra algorithm, IET Intelligent Transport
validation phase to also realize a physical Systems, 10, No. 7, 2016, pp. 495-502.
implementation of the solution. [7]. Katyal, Mayanka, and Atul Mishra, A comparative
study of load balancing algorithms in cloud computing
environment, International Journal of Computer
References Applications, 117, 24, 2014, pp. 33-37.
[8]. Kim, Oanh Tran Thi, Nguyen H. Tran, Chuan Pham,
[1]. Todd Litman, Parking management: strategies, Tuan LeAnh, My T. Thai, and Choong Seon Hong,
evaluation and planning. Victoria, BC: Victoria Parking assignment: Minimizing parking expenses and
Transport Policy Institute, 2016. balancing parking demand among multiple parking
[2]. E. Hassania Rouan and A. Boumezzough, RFID Based lots, IEEE Transactions on Automation Science and
Security and Automatic Parking Access Control Engineering, 17, no. 3, 2019, pp. 1320-1331.
System, Business Intelligence, 2021, pp. 434–443. [9]. A. Souza, Z. Wen, N. Cacho, A. Romanovsky,
[3]. M. Owayjan, B. Sleem, E. Saad and A. Maroun, P. James and R. Ranjan, Using Osmotic Services
Parking management system using mobile application, Composition for Dynamic Load Balancing of Smart
in Proceedings of the Sensors Networks Smart and City Applications, in Proceedings of the IEEE 11th
Emerging Technologies Conference (SENSET’ 2017), Conference on Service-Oriented Computing and
2017, pp. 1-4. Applications (SOCA’ 2018), 2018, pp. 145-152.
[4]. IEEE Spectrum: Technology, Engineering, and [10]. P. A. Lopez et al., Microscopic Traffic Simulation
Science News, 2021. using SUMO, in Proceedings of the 21st International
https://spectrum.ieee.org/computing/ Conference on Intelligent Transportation Systems
hardware/your-navigation-app-is-making-traffic- (ITSC’ 18), 2018, pp. 2575-2582.
unmanageable (accessed Jun. 12, 2021). [11]. Python: module traci._vehicle, Sumo.dlr.de, 2021.
[5]. Madkour, Amgad, et al., A survey of shortest-path https://sumo.dlr.de/daily/pydoc/traci._vehicle.html
algorithms, International Journal of Applied (accessed Jan. 2, 2022).
Engineering Research, Vol. 13, No. 9, 2018, [12]. Sandru F. D., Nanu S., Silea I. and Miclea R. C.,
pp. 6817-6820. Kalman and Butterworth filtering for GNSS/INS data,
[6]. Zhang, Jin-dong, Yu-jie Feng, Fei-fei Shi, Gang Wang, in Proceedings of the 12th IEEE International
Bin Ma, Rui-sheng Li, and Xiao-yan Jia, Vehicle Symposium on Electronics and Telecommunications
routing in urban areas based on the oil consumption (ISETC’2016), 2016, pp. 257-260.
69
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(015)
Summary: Different propulsion mechanisms have been suggested for a variety of chemical micromotors. Their high efficiency
and thrust force enable several applications in the fields of environmental remediation and biomedicine. In particular, bubble-
recoil based motion has been modeled by three different phenomena: capillary forces, bubble growth, and bubble expulsion.
However, these models have been suggested independently based on a single influencing factor (i.e., viscosity), limiting the
understanding of the overall micromotor performance. Here the combined effect of medium viscosity, surface tension, and
fuel concentration on the micromotor swimming performance is analysed. Hence the dominant propulsion mechanisms
describing its motion are accurately identified. Using statistically relevant experimental data, a holistic theoretical model is
proposed for bubble-propelled tubular catalytic micromotors that includes all three above-mentioned phenomena. The
proposed model provides deeper insights into their propulsion physics toward optimized geometries and experimental
conditions. These findings pave the way for highly optimized catalytic micromotors for various applications.
70
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
71
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
active catalytic area of the micromotor, which causes 9, 4, 2018, pp. 303–316.
an increase in the oxygen production rate. An increase [2]. Wang, B., Kostarelos, K., Nelson, B. J. and
of both the semi-cone angle and the length of the Zhang, L., Trends in micro-/nanorobotics:
micromotors leads to an increase in the catalytic materials development, actuation, localization,
surface area. Particularly, an increase of the semi-cone and system integration for biomedical
angle from 2.5° to 5° for a conical micromotor with a applications, Advanced Materials, 33, 2021,
length of 50 µm results in a surface area decrease of 20 pp. 1–44.
%. Additionally, larger semi-cone angles increase the [3]. Srivastava, S. K., Guix, M. & Schmidt,
drag force acting on the micromotor. In summary, an O. G., Wastewater mediated activation of
increase in bubble expulsion frequency and bubble micromotors for efficient water cleaning, Nano
radius is predicted for increasing micromotor length Letters, 16, 2016, pp. 817–821.
and semi-cone angle. At the same time, an increased [4]. Jurado-Sánchez, B. & Wang, J., Micromotors for
speed is predicted for smaller semi-cone angles and environmental applications: A review.
longer micromotors. Environmental Science: Nano, 7, 2018,
pp. 1530–1544.
3. Conclusion [5]. Zhang, Y., Yuan, K. & Zhang, L., Micro/
Nanomachines: From functionalization to
A physical explanation of the bubble-propelled sensing and removal, Advanced Materials
micromotor motion is performed using the established Technologies, 4, 2019, pp. 1–22.
capillary force, bubble growth, and jet-like propulsion [6]. Mei, Y. F., Naeem, F., Bolaños Quiñones,
mechanisms. Switching of propulsion mechanisms is V. A., Huang, G. S., Liao, F., Li, Y., Manjare,
for the first time unveiled at certain values of the H2O2, M., Naeem, S., Solovev, A. A. and Zhang,
MC, and SDS concentrations. Besides, it is observed J., Tubular catalytic micromotors in transition
that for certain experimental parameters, a concerted from unidirectional bubble sequences to more
action of all three propulsion mechanisms is observed. complex bidirectional motion, Applied Physics
A theoretical model is proposed to calculate the speed Letters 114, 2019, pp. 1–16.
of the bubble-propelled micromotors, including the [7]. Wrede, P., Medina-Sánchez, M., Fomin,
contributions of all three known propulsion V. M. and Schmidt, O. G., Switching propulsion
mechanisms, resulting in good agreement with the mechanisms of tubular catalytic micromotors,
obtained experimental data. This offers new Small, 17, 2021, pp. 1–11.
possibilities to optimize and predict the performance of [8]. Klingner, A., Khalil, I. S. M., Magdanz,
bubble-propelled micromotors for different V., Fomin, V. M., Schmidt, O. G. and Misra,
applications in various working environments, S., Modeling of unidirectional-overloaded
including new applications, such as, in-situ sensing of transition in catalytic tubular microjets, Journal
the medium viscosity, as well as for fundamental of Physical Chemistry C, 121, 2017, pp. 14854–
understanding of the impact of various geometrical 14863.
parameters and experimental conditions (length, semi- [9]. Manjare, M., Yang, B. and Zhao, Y. P. Bubble-
cone angle, surface properties) on the propulsion of propelled microjets: Model and experiment.
conical catalytic micromotors. Journal of Physical Chemistry C, 117, 2013,
pp. 4657–4665.
References [10]. Li, L., Wang, J., Li, T., Song, W. and Zhang,
G., Hydrodynamics and propulsion mechanism
[1]. Medina-Sánchez, M., Xu, H. and Schmidt, of self-propelled catalytic micromotors:
O. G. Micro- and nano-motors: The new Model and experiment, Soft Matter, 10, 2014,
generation of drug carriers, Therapeutic Delivery, pp. 7511–7518.
72
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(019)
Summary: Singular Perturbation Margin (SPM) and Generalized Gain Margin (GGM) are Phase Margin (PM) and Gain
Margin (GM) like stability metrics. In this paper, the problem of SPM and GGM assessment for Linear Periodically Time-
Varying (LPTV) systems is formulated. Chang transformation makes it possible to reduce the SPM analysis for Hill equations,
which is essentially a stability problem of higher order LPTV systems due to the SPM gauge introduced dynamics, to the
second order LPTV systems. Based upon Floquet Theory, the SPM and the GGM assessment methods for the second order
and the general order LPTV systems are established, respectively.
Keywords: Stability margin, Singular perturbation, Regular perturbation, Linear system, Periodically time-varying systems,
Hill equation.
73
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Anom (t ) Anom (t T ) (2) Design the SISO LTI SPM gauges at the perturbation
sites, which might have the equation forms as [20]
minimum-phase gauge, zero-error gauge or
with T the system period, is referred as to the
approximately-linear phase gauge, etc., and the
quantitative assessment of the capability of the
singularly perturbed system is given by
exponentially stable null equilibrium (origin) in
accommodating singular perturbations and regular
perturbations: x A 11 (t ) x A12 (t ) z
, (4)
(1) Design the SPM gauge and GGM gauge for the z A21 (t ) x A22 z
specific perturbations;
(2) Determine the stability of the perturbed system where the singular perturbation parameter describes
using stability criterion when different singular
the time-scale separation, A11 (t ), A12 (t ), A 21 (t ), A22
perturbation parameters and gain parameters are
given, respectively; are function matrices and parameter matrix that are
(3) Obtain the SPM and GGM by the marginally dependent on the perturbation site and the SPM gauge
design, and z
m
stable situation of the perturbed system; is state vector of the fast
(4) Reveal the relation between the characteristics of dynamics.
the nominal system (1) and its exact SPM and
GGM.
3.2. Decoupled Slow and Fast Systems
3. Perturbed Hill Equation The singularly perturbed system (4) can be totally
decoupled to slow and fast system by Chang
3.1. Stability Criterion and SPM/GGM Gauges transformation.
We first introduce a necessary and sufficient Lemma 2 With the SISO LTI SPM Gauge, the
exponential stability condition for LPTV systems that decoupled slow system of the singularly perturbed Hill
will be used in the sequel. equation (4) is a Hill equation with the same period to
Lemma 1 [4] The natural response of an LPTV the nominal system.
system is stable, if and only if no eigenvalues of the Proof. Consider the two perturbation positions,
Discrete Transition Matrix (DTM) have magnitudes Site 1 and Site 2, in the loop of Fig. 1.
greater than one and that an eigenvalue with unity Site 1: With the SISO LTI SPM Gauge, the system
magnitude is not degenerate. matrices of the singularly perturbed system in form of
The eigenvalues of the DTM is also called Floquet (4) are given by
characteristic multipliers of the periodic system. The
... 0 ,
A11 0 1 , A12 (t ) 0 a 2 q (t ) C
transition matrix of a general order LPTV system is
analytically intractable in Lemma 1, but for the second- 0 0 f
order cases, the general form
A21 B f 1 0 , A22 A f ,
x 1 (t ) x 0 (t ) 0
where A f , B f , C f are the parameter matrices of the
can be converted into an expression without a first SISO SPM gauge. Chang transformation [22] is given
derivative, and the classical form is by
74
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
which can be denoted as x 2 (t ) x 1 (t ) x 0 .
L ( t ) j L j ( t ), H (t ) j H j (t )
j 0 j 0 Then, by transformation
1 t
where x ( t ) exp 2 (t )dt y (t )
2 0
L0 A221 A21 , H 0 A121 A22
Lk 1 A221 L k Lk A11 A12 Lk , k 0,1, 2,... Eq. (7) is transformed to be
H k 1 H k H k Lk A12 A11 A12 Lk H k A221
y ( t ) 1 (t ) 2 (t ) 22 ( t ) y (t )
1 1
k 0,1, 2,... 2 4
C L
x C f L1 C ff L22 x C f L1
C f L2 C f L1
C f L2
(7) Fig. 2. Stability diagram for the Mathieu equation. Blank
2 q (t ) a C f L2 x 0 regions correspond to stable solutions and shaded regions to
unstable solutions. The diagram is symmetrical about a axis.
75
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
In the expression of the slow system, matrix L(t) order LPTV system to obtain the singular perturbation
can be obtained by the iterative expression and the margin.
slow system is given by 3.5. GGM Assessment of Hill Equations
0 1
0
0 0 2 q cos(2t ) a 02
0
0
L(t ) when the function (t ) is unit rectangular waveform
in (3), which is shown by Fig. 3, and can be viewed
as the pair of constant coefficient equations
L( t ) j L j ( t )
j 0
76
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
In Fig. 4, the blank area represents the stable solutions equations, which is essentially a stability problem of
of Meissner equation with the parameter couple higher order LPTV systems due to the SPM gauge
(q, a) . When the nominal system with parameter introduced dynamics, to second order LPTV systems.
Based upon Floquet Theory, the SPM and GGM
couple ( q0 , a0 ) , is perturbed by regular perturbation assessment methods for the second order and general
parameter k, then ( q , a ) will move along the line order LPTV systems are established, respectively.
a / q 3 , until the couple ( q f , a f ) first touches the
boundary of the blank area, i. e. the regularly perturbed
Acknowledgements
system becomes marginally stable, which is shown by
the legend square. Then, the GGM k m ax and k m in are
The author is grateful of the financial support from
obtained. National Natural Science Foundation of China No.
61803010 and Aeronautical Science Foundation of
China No.2018ZA51002.
References
[1]. M. Faraday, On a peculiar class of acoustical figures;
and on certain forms assumed by groups of particles
upon vibrating elastic surfaces, Proceedings of the
Royal Society of London, 3, 1830, pp. 229-318.
[2]. E. Mathieu, Memoire sur le movement vibratoire d'une
membrane de forme elliptique, J. Math. Pure Appl., 13,
1868, p. 137.
[3]. G. Floquet, Sur les equations differenctiales lineaires.
Ann. L'Ecole Normale Super., 12, 1883, p. 47.
[4]. J. A. Richards, Analysis of Periodically Time- Varying
Systems. Springer, 1983.
Fig. 4. Stability Diagram for Meissner Equation and GGM [5]. G. W. Hill, On the part of motion of the lunar perigee
Assessment. which is function of the mean motions of the sun and
moon, Acta Mathematica, 8, 1886, pp. 1-36.
[6]. E. Meissner, Uber Schuttelerscheinungen im System
Case 2. When Meissner equation is regularly mit periodisch veranderlicher Elastizitat, Schweiz.
Bauztg., 72, 1918, pp. 95-98.
perturbed on parameter q. The parameters d and c [7]. J. R. Carson, Notes on theory of modulation, Proc. IRE,
in the DTM of the regularly perturbed system become 10, 1922, pp. 57-64.
d a 2qk , c a 2qk . In Fig. 4, when
[8]. H. Jeffreys, Approximate solutions of linear
differential equations of second order, Proc. Lon. Math.
a 20 , the legend circles show boundaries of stability Soc., 23, 1924, p. 428.
solutions, and with q 1 as the nominal system [9]. B. van der Pol, M. J. O. Strutt, On the stability of
solutions of Mathieu's equation, Philos. Mag., 5, 1928,
parameter (point A), the GGM of Meissner equation is pp. 18-39.
k m a x 1 .6 , k min 1.6 . [10]. W. Y. Yan, R. R. Bitmead, Decentralized control
techniques in periodically time-varying discrete-time
Case 3. When Meissner equation is regularly control systems, IEEE Transactions on Automatic
Control, 37, 10, 1992, pp. 1644-1648.
perturbed on parameter a, the parameters d and c [11]. M. S. Chen, Y. Z. Chen, Static output feedback control
in the DTM become d ak 2q , c ak 2q .
for periodically time-varying systems, IEEE
Transactions on Automatic Control, 1999, 44, 1,
In Fig. 3, the legendtriangle represent the boundaries pp. 218-222.
of stable solutions. [12]. S. Kalender, H. Flashner, A new approach for control
of periodically time varying systems with application
to spacecraft momentum unloading, in Proceedings of
3.6. GGM Assessment of General Order Equations the AIAA Guidance, Navigation, and Control
Conference and Exhibit, 2006.
Due to the numerical methods for DTM, the [13]. J. Zhang, C. Zhang, Robustness of discrete periodically
algorithems of section 3.5 can be generalized to any time-varying control under LTI unstructured
perturbations. IEEE Transactions on Automatic
order LPTV system to obtain the GGM.
Control, 45, 7, 2000, pp. 1370-1374.
[14]. P. Brunovski, Controllability and linear closed-loop
controls in linear periodic systems, Differential
4. Conclusions Equations, 6, 4, 1969, pp. 296-313.
[15]. Y. T. Hagiwara, Periodically time-varying controller
The problem of SPM and GGM assessment for synthesis for multiobjective control of discrete-time
LPTV systems is formulated. Chang transformation systems and analysis of achievable performance."
makes it possible to reduce the SPM analysis for Hill Systems & Control Letters, 6, 2, 2011, pp. 296-313.
77
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
[16]. S. Bittanti, 30 years of periodic control-form analysis with bounded parameter variation, Iet Control Theory
to design, in Proceedings of the 3rd Asian Control & Applications 3, 10, 2009, pp. 1448-1461.
Conference, Shanghai, 2000, pp. 1253-1258. [20]. X. Yang, J. J. Zhu, A. S. Hodel, Singular perturbation
[17]. X. Su, M. Lv, H. Shi, Q. Zhang, Stability analysis for margin and generalized gain margin for linear time
linear time-varying periodically descriptor systems: a invariant systems, International Journal of Control,
generalized Lyapunov approach, in Proceedings of the 88, 1, 2015, pp.11-29.
IEEE World Congress on Intelligent Control & [21]. X. Yang, J. J. Zhu, Singular perturbation margin and
Automation, 2006. generalized gain margin for nonlinear time invariant
[18]. X. Xie, J. Lam, P. Li, A novel h∞ tracking control systems, International Journal of Control, 89, 3, 2016,
scheme for periodic piecewise time-varying systems, pp. 1-18.
Information Sciences, 484, 2019, pp. 71-83. [22]. P. V. Kokotovic, H. K. Khalil, J. O’Reilly, Singular
[19]. R. C. L. F. Oliveira, M. C. De Oliveira, P. L. D. Peres, perturbations methods in control: analysis and design,
Special time-varying Lyapunov function for robust Academic Press, New York, 1986.
stability analysis of linear parameter varying systems
78
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(020)
Summary: The movement of materials in industrial assembly lines can be done efficiently using Automated Guided Vehicles
(AGVs). However, visual perception of industrial environments is complex due to the existence of many obstacles in pre-
defined routes. With the INDTECH 4.0 project, we aim to develop an autonomous navigation system, allowing the AGV to
detect and avoid obstacles based on the processing of depth data acquired with a frontal depth camera mounted on the AGV.
Applying RANSAC algorithm and Euclidean Clustering to the 3D point clouds captured by the camera, we are able to isolate
obstacles from the ground plane and separate them into clusters. The clusters give information about the location of obstacles
with respect to the AGV position. In the experiments conducted in outdoor and indoor environments the results revealed that
the method is effective, returning high percentages of detection for most tests.
Keywords: Autonomous navigation, Obstacle avoidance, AGV (Automated Guided Vehicle), ROS (Robot Operating
System).
79
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
environment that can be static or dynamic, indoor or approximated by their centroid. This process results in a
outdoor, in a process called natural navigation. point cloud with less points, but that still describes the
The remainder of this paper is structured as follows. original surface accurately.
Section 2 is a description of our methodology, detailing In our tests, we defined the limits of the distance
the various processing phases of our method. Section 3 window in the interval [1.5, 5] m along the z-axis for the
presents and analyses the results according to various pass-through filter and a voxel edge length of 0.07 m for
metrics relevant for this kind of algorithm. Section 4 the voxel grid.
presents the conclusions of this paper and suggestions of
future work.
2.2 RANSAC Segmentation
80
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
performed in the premises of Active Space Technologies For each ROS bag, the point from each cluster closest
and consisted in recording ROS bags using the D435i to the camera is obtained and published in a ROS topic
depth camera with various types of obstacles located at (cluster_closest_point) for analysis. This topic contains
different distances. We tested the algorithm for both the xyz coordinates of the closest point of each object.
static and mobile obstacles and in indoor and outdoor The total number of point clouds analysed by the
environments. algorithm is also counted.
To represent static obstacles, chairs placed at 1.8 m, A python script is used to subscribe to the
2.5 m and 4.5 m were used. The distance to the chairs (cluster_closest_point) topic and the information is
was measured taking the closest part of the chair into assembled and plotted to a 3D graph. This graph gives us
account. This set up allows us to evaluate the algorithm a sense of the position of obstacles in the corresponding
performance for different distances across the defined test.
depth window ([1.5, 5] m). For each of the three In addition to visualising the data, a clustering
distances, a ROS bag is recorded with two chairs side- algorithm is applied to this set of points, originating
by-side, indoors and outdoors. clusters of points that correspond to detections of the
Another test was made with three chairs, one at each same object. Using the clusters of closest points, we
distance mentioned above, to assess simultaneous count the number of points that represent each obstacle
detection of various obstacles at different distances. The detected by the camera. That value determines how many
final test with chairs consisted in moving the camera times each obstacle is detected.
perpendicularly to the three chairs, in the direction of the By comparing these values with the number of point
chair furthest away. This test simulates object detection clouds processed by our object detection algorithm we
of static obstacles with a circulating AGV. are able to calculate the effectiveness of the object
For mobile obstacles, we performed tests in which detection.
people moved in front of the camera in various Each ROS bag has a duration of approximately
directions. A final test was made in which the camera was 15 seconds, which at 30 fps represents a total of
also moving among wandering people, simulating a more 450 frames for each test run. The process described
complex scenario that would be common for an industrial above is repeated 10 times for each ROS bag and the
AGV. results are presented in Tables 1 and 2.
Every test mentioned above has been done indoors
and repeated outdoors.
This batch of tests enables us to get a sense of the Table 1. Indoor detection effectiveness.
algorithm’s performance in different conditions
considered relevant for the project. Figure 2 represents Distance (m) Indoor Detection Effectiveness [%]
the typical outputs of this algorithm concerning the three 1.8 92.72 ± 5.04
major stages of processing. 2.5 70.59 ± 6.98
4.5 51.02 ± 0.70
81
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Thus, detection success decreases with distance to the whole test regardless of the movement of the AGV and
obstacle, but does not appear to change for obstacles at of the person.
the same distance. The percentage of obstacle detection
is also significantly more stable in outdoor conditions, as
can be concluded by the lower standard deviation values.
4. Conclusions
For distance measurements, we are able to obtain
reasonable values, with an expected drop in accuracy for In this work, an object detection algorithm using
farther objects. These measurements are entirely point cloud data from an Intel RealSense D435i is
dependent on the accuracy of the Intel RealSense D435i proposed. This algorithm is meant to be applied in an
depth camera, and, as demonstrated, we can rely on the autonomous navigation system with obstacle avoidance
good accuracy provided by the depth sensor for our capabilities for AGVs in dynamic industrial
application. environments.
The presented results show this method can detect
obstacles and measure their position relative to the AGV
3.3 Moving Object Detection in various environmental conditions. This is very
important in a SLAM (Simultaneous Localization and
For moving obstacles, the performance of the Mapping) context to enable robust autonomous
algorithm shows favourable results, although the analysis navigation of robots. This method is also able to function
of these tests has been made merely by visual inspection. in real-time, which is crucial for autonomous navigation.
The proposed method is able to segment and follow The algorithm detects objects with high efficacy and
moving obstacles smoothly with high degree of accuracy returns results with inconsequential frame drops from the
and with high frame rate without apparent fail in camera’s native frame rate. The resulting frame rates of
detection. detection are adequate for a slow moving AGV, as
Figure 3 shows two frames from the mobile intended for our application.
AGV/wandering person test where it is clear that the The obstacle position measurements returned by our
person is segmented accurately and there are no false method also present high degrees of precision that
positives. The detection remains accurate during the improve with illumination and as obstacles get closer to
82
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
the camera, characteristics that are innate to the Intel [3]. Zhou Y. and Tuzel O. Voxelnet: End-to-end learning for
RealSense D435i depth camera. point cloud based 3D object detection, in Proceedings of
Further developments of this algorithm may provide the IEEE Conference on Computer Vision and Pattern
the capability to create a system that generates odometry Recognition, 2018, pp. 4490-4499.
[4]. Qi C. R., Liu W., Wu C., Su H. and Guibas
measurements using depth camera data, such as L. J., Frustum pointnets for 3D object detection from
implementing a visual odometry system for SLAM. Such RGB-D data, in Proceedings of the IEEE Conference on
algorithms are useful to perform sensor fusion with an Computer Vision and Pattern Recognition, 2018,
inertial measurement unit or can even replace this kind pp. 918-927.
of sensors. Finally, the methodology can be extended to [5]. Shi S., Wang X. and Li H. PointRCNN: 3D object
process RGB images, acquired simultaneously by the proposal generation and detection from point cloud, in
Intel camera, for object detection and classification. Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition, 2019, pp. 770-779.
[6]. Kircali D. and Tek F. B. Ground plane detection using an
Acknowledgements RGB-D sensor, in Proceedings of the 29th International
Symposium on Computer and Information Sciences,
This work was funded by Project “INDTECH 4.0 – 2014, pp. 69-77.
new technologies for smart manufacturing”, POCI- 01- [7]. Habermann D., Hata A., Wolf D. and Osório F. S.
0247-FEDER-026653, financed by the European 3D point clouds segmentation for autonomous ground
vehicle, in Proceedings of the III IEEE Brazilian
Regional Development Fund (ERDF), through the Symposium on Computing Systems Engineering, 2013,
COMPETE 2020 - Competitiveness and pp. 143-148.
Internationalization Operational Program (POCI). The [8]. Vosselman G. Point cloud segmentation for urban scene
work is also funded by national funds from FCT - classification, Int. Arch. Photogramm. Remote Sens.
Portuguese Foundation for Science and Technology, Spat. Inf. Sci, Vol. 1, No. 257-262, 2013, p. 1.
under the project UIDB/04033/2020. [9]. Nguyen A. and Le B. 3D point cloud segmentation: A
survey, in Proceedings of the 6th IEEE Conference on
Robotics, Automation and Mechatronics (RAM), 2013,
References pp. 225-230.
[10]. Rusu R. B. and Cousins, S. 3D is here: Point cloud library
[1]. Pires M. A., Natural navigation solutions for AMRs and (PCL), in Proceedings of the IEEE International
AGVs using depth cameras, MSc diss., Universidade de Conference on Robotics and Automation, 2011, pp. 1–4.
Coimbra, 2021. [11]. Fischler M. A. and Bolles R. C., Random sample
[2]. Lynch L., Newe T., Clifford J., Coleman J., Walsh consensus: a paradigm for model fitting with applications
J. and Toal D. Automated ground vehicle (AGV) and to image analysis and automated cartography,
sensor technologies - a review, in Proceedings of the 12th Communications of the ACM, Vol. 24, No. 6, 1981,
IEEE International Conference on Sensing Technology pp. 381–395.
(ICST’18), 2018, pp. 347-352.
83
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(023)
Summary: The changes in industrial processes that are being introduced by the new industrial revolution called "manufacturing
4.0" have broad repercussions on engineering education. Engineering education depends on the competencies required by the
industrial sector; however, the technological transition is not constant and homogeneous in companies or in productive regions,
especially in developing countries such as Mexico, so there are still many productive systems that operate under the philosophy of
manufacturing 3.0. It is necessary to design an adequate profile and determine an appropriate educational model for the mechatronic
engineer in the context of the transition between manufacturing 3.0 and industry 4.0, which allows him/her to have the necessary
competences to face the current and future challenges of the companies. This paper proposes Competency Based Education (CBE)
as the basis for training mechatronic engineers.
84
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
as manufacturing cell or CIM (Computer Integrated service industries [3]. One of the crucial infrastructures
Manufacturing) have evolved to cyber-physical systems that must be valued in the training of the mechatronics
and digital twins, integrated with artificial intelligence, engineer is computational as it is constantly changing and
advanced simulation and autonomous and cooperative has a wide influence on the technologies supporting
robots. Similarly, the design and manufacture of various Industry 4.0, so constant updating of curricula can help
parts and components are now performed by advanced the computational infrastructure not to be overtaken by
software and simulation, as well as by additive changes in processing speed or become obsolete in a
manufacturing systems. Information is stored and short period of time. In fact, mechatronics has gone from
processed by Big Data and cloud computing techniques, being purely associated with essentially autonomous
and protected by cybersecurity methods. Training is systems, such as robots, to providing the intelligent
aided by virtual and augmented reality techniques. To objects and systems that are the building blocks of cyber-
these challenges must be added the great improvements physical systems and, therefore, of systems based on the
in interconnection and communication provided by the Internet of Things and the cloud, and all this has been
Internet of Things and the improvement of IT integration made possible due to the evolution of computer
systems in companies. All these technological technologies, so the importance of these technologies in
improvements bring great challenges only on the the education of the mechatronics engineer is
technical side for mechatronics engineers. transcendental.
With regard to the various companies, especially in Mechatronics was born at the height of the third
Mexico, the technological transition (from industrial revolution and, consequently, together with
manufacturing 3.0 to industry 4.0) has become a serious computing, informatics and robotics, it brought a
problem, as they face the challenge of deciding whether significant technological impact to the industrial world
a technological upgrade (almost total change of over a period of at least 50 years. Today, the education
production systems) that implies a considerable of the mechatronics engineer faces the major challenges
economic investment or a technology reconversion that of transforming CIM systems (core of manufacturing
involves a substantial improvement to the production 3.0) to modern cyber-physical systems (CPS) that
systems they already have and, therefore, less economic integrate various Digital Twins (DT) and are the basis of
investment, is relevant. This type of analysis and decision Industry 4.0.
making undoubtedly involves mechatronics engineers, Cyber-physical entities are described as any entity
so it will be necessary to use reverse engineering composed of physical and cyber-physical elements that
techniques for the evaluation of the technologies interact autonomously with each other, with or without
available to the companies if a technological human supervision [4]. The main characteristic of a
reconversion criterion is to be applied. Likewise, it will Cyber-Physical System is its ability to merge the physical
be necessary to be involved in the technology transfer and virtual worlds, in particular, through its embedded
process, which implies knowledge of the purchase, software.
installation and start-up of the systems acquired and One of the technologies on which cyber-physical
knowledge of intellectual property for the study of systems are supported is simulation, described as the
patents and licenses in the event that a technological imitation of the behavior of the properties of a system.
update is chosen. One of the most important concepts within CPS are
The transition between two industrial revolutions Digital Twins that use simulation as part of their
generates an important series of scientific and processes.
technological challenges, not only for companies, but A digital twin is defined as "a formal digital
also for governments, the education sector, trade and representation of some asset, process, or system that
strategic competition between regions or countries. captures attributes and behaviors of that entity suitable
for communication, storage, interpretation, or processing
within a given context" by the Industrial Internet
3. Mechatronics Education in Industry 4.0 Consortium [5].
In general CPS are understood from the point of view
The synergic integration of Mechanics, Control, of integration that combines systems or processes that
Electronics and Computing in the development of occur in reality with computational processes, while
products, machines, processes and production systems Digital Twins or digital copy are used in real-time
has always had great challenges for the mechatronics improvement and optimization processes. This implies
engineer, so now we must consider that to the traditional that the Digital Twin relates to its physical counterpart in
challenges of mechatronics we must add various two ways: 1) It receives from it information generally
requirements and technical requirements demanded by from sensors and 2) The valuable or processed
industry 4.0, especially the various techniques of information from the digital copy is sent to the physical
artificial intelligence, digitization and the internet part to achieve specific objectives. Cyber-physical
of things. systems and Digital Twins are two of the central concepts
The fourth industrial revolution requires that must necessarily be taken into account for the design
multidisciplinary knowledge (as in the case of or redesign of mechatronics career curricula and for the
Mechatronics) and expertise. Multidisciplinary skills are design of multidisciplinary competencies.
mandatory and real-time collaboration will be more On the other hand, the most significant
important than ever for success in the manufacturing and transformation related to the way products are
85
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
manufactured is digitalization. The fourth industrial evidence of the level of development of the defined set
revolution aims to optimize the computerized third of competencies [9].
industrial revolution (Industry 3.0). This requires the On the other hand, with the purpose of making
development of intelligent equipment with access to students access active and meaningful learning, active
more data, thus becoming more efficient and productive methodologies were introduced. Prince and Felder
by making decisions in real time [6]. In this sense, the compiled six most innovative teaching methods, namely,
different educational models around mechatronics inquiry learning, problem-based learning, project-based
should consider as a priority two fundamental aspects for learning, case-based learning, discovery learning, and
the training of engineers: 1) The formal study of the just-in-time teaching. These teaching methods are
various algorithms and tools of Artificial Intelligence and learner-centered and motivate students with real-world
2) The conformation of intelligent teaching and learning problems, cases, and observations, which triggers
environments. There is a range of tools that support students desire to learn knowledge [10].
Artificial Intelligence, such as Deep Learning, Machine There are several proposed taxonomies for
Learning, Neural Networks, Fuzzy Logic, Genetic engineering education. One of them consists of seven
Algorithms, Natural Language Processing, Knowledge categories [11]: 1) Engineering specific work includes
Engineering, Expert Systems, Data Mining and others. general engineering work, specialized engineering work,
These tools will be commonly used in industries in the and engineering combined with other work, 2) Non-
coming years. engineering work specific work includes general,
Faced with the constant challenges of manufacturing supervision and leadership, and management and
4.0, mechatronics and mechatronics education must administration, 3) Communication 4), Interpersonal
reinvent themselves. To this end, designers of interactions include personal and teamwork, 5) Personal
mechatronic systems of all varieties and types must be dispositions include general and discipline 6), Adaptive
aware of a set of fundamental principles if they are to dispositions include personal development, lifelong
ensure success. Some of these principles are as follows: learning and change management, and 7) Advanced
technology; complexity; operational software; hardware; dispositions include achievement orientation, impact and
reliable and rechargeable power supplies; innocent influence, conceptual thinking, analytical thinking,
human error; connectivity; privacy, dependency, initiative, self-confidence, interpersonal understanding,
ubiquity, and hybrid society, among others [7]. concern for order, information seeking, teamwork and
cooperation, experience and service orientation.
The competency model may be ideal for the
comprehensive training of the mechatronics engineer
4. Competency-Based Approach
whose professional activity will be to solve problems in
the transition from manufacturing 3.0 to industry 4.0, as
The training of the new engineer not only demands this model offers students the possibility to hone their
an update of the curriculum but also requires universities ability to recognize, build and manage their skills. It
to have a strong connection with companies and the allows students to evaluate and improve their
adoption of new educational approaches, such as performance, solve difficulties, plan innovative
Competency-Based Education (CBE) and active learning strategies and interpret diverse situations in an ethical
methodologies, as well as a rational contextualization of and responsible manner. The competencies integrate a
the conditions in which the technological transition is response to society, the culture of quality, globalization
taking place in each country and in each productive and business competitiveness.
region. Competencies are understood as a combination of Although in Mexico, as in other Latin American
skills, abilities, knowledge, aptitudes, attitudes and countries, CBE has been implemented for two decades,
values to perform well in complex and authentic this educational approach has not been homogenized nor
contexts. has it been effective in engineering education. Some
Competency-Based Education is both goal-oriented universities in other countries have been applying the
and outcome-oriented. It is an approach embedded in competency-based approach and have even adopted
adult learning theory, which posits that adult learners are active methodologies such as Project Based Learning
more likely to engage in learning that is focused on a [12] and Learning Factories [13] to train engineers in the
specific goal or outcome. This means that both teaching current industrial context. It is necessary that CBE be
and learning in CBE are oriented toward the development established in an integral way in engineering careers in
of an explicitly stated and described skill, and measured Mexico and in other Latin American countries, since in
by observing how the learner performs this skill [8]. A this way it will be possible to design effective curricula
Competency-Based Education approach focuses on the and learning strategies, as is the case of Mechatronics
development and demonstration of competencies by a Engineering, in accordance with the demands of the new
learner. Thus, a CBE perspective puts the learner at the industrial revolution.
center of the model and provides a comprehensive
education oriented to the competencies demanded in the
world. CBE has strong implications in the teaching- 5. Conclusions
learning process. This approach alters the curricular
organization, the way content is taught, the learning The current profile of the mechatronic engineer must
environments, the assessment instruments and the be designed taking into account the context of each
86
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
country and the implications generated in the transition [4]. DeSmit Z., Elhabashy A. E., Lee J. Wells L. and Camelio
from manufacturing 3.0 to industry 4.0. The objective of J. A., Cyber-Physical Vulnerability Assessment in
their training should be the acquisition of competencies Manufacturing Systems, Procedia Manufacturing, 5,
and technical and social skills that allow them to solve 2016, pp. 1060-1074.
[5]. Jian D., Tian M., Qing Z., Zhen L. y Ji‑Y., Design and
problems in both technological approaches.
application of digital twin system for the blade‑rotor test
The CBE is a good option for the formation of the
rig, Journal of Intelligent Manufacturing, 2021.
mechatronic engineer in the context of the current
[6]. Souza, M. L. H., C. Andre da Costa, G. de Oliveira
technological transition, because it is possible to design Ramos, Rodrigo da Rosa Righi, A survey on decision-
and evaluate the necessary competencies in engineering making based on system reliability in the context of
students for each requirement needed by the industry. It Industry 4.0, Journal of Manufacturing Systems, 56,
is necessary to incorporate in the curricula and study 2020, pp. 133–156.
plans of the mechatronics career special topics of cyber- [7]. Russell D., Reinventing Mechatronics - Final Thoughts,
physical systems, digital twins and artificial intelligence in Reinventing Mechatronics, Xiu-Tian Yan, David
with the aim of making inroads into the base technologies Bradley, David Russell and Philip Moore (Eds.),
of industry 4.0, and promote the development of Springer, 2020, pp.179-186.
[8]. Mace K. L., Welch C. E., The future of health professions
competencies and skills in programming and computing
education: considerations for competency-based
in students, since this topic is of great relevance and education in athletic training, Athletic Training
importance in the current context of technological Education Journal, 14, 03, 2019, pp. 215–222.
applications. [9]. Félix L. C., Rendon A. E. and Nieto J. M., Challenge-
Mechatronics as an area of technological application based learning: an I-semester for experiential learning in
and education of the same, face the difficulties and Mechatronics Engineering, International Journal on
challenges involved in transforming CIM systems that Interactive Design and Manufacturing, 13, 2019,
represent the technologies of manufacturing 3.0 to cyber- pp. 1367–1383.
physical systems that make up the core of industry 4.0. [10]. Prince M. J. and Felder R. M., Inductive teaching and
learning methods: Definitions, comparisons, and research
bases, Journal of Engineering Education, Vol. 95, 2006,
pp. 123-138.
References [11]. Woollacott L. C., Validating the CDIO syllabus for
engineering education using the taxonomy of engineering
[1]. Kaur R., Awasthi A., Grzybowska K. Evaluation of Key competencies. European Journal of Engineering
Skills Supporting Industry 4.0—A Review of Literature Education, 34, 6, 2009, pp. 545–559.
and Practice, in Grzybowska K., Awasthi A., Sawhney R. [12]. Atmojo U. D., Project-based Learning at Aalto Factory of
(Eds.), Sustainable Logistics and Production in Industry the Future - A Flexible Production-based Industrie 4.0
4.0. EcoProduction (Environmental Issues in Logistics Learning Factory, in Proceedings of the 11th Conference
and Manufacturing). Springer, Cham. 2020. on Learning Factories (CLF’2021), 1–2 July, 2021,
[2]. BRICS skill development working group: Skills Graz, Austria, 2021.
development for Industry 4.0. 2018. [13]. Montoya A., Guarín A. and Mora J., Advantages of
[3]. Stankovski S., Ostojić G., Zhang X., Baranovski I., Learning Factories for Production Planning Based on
Tegeltija S., and Sabolč Horvat. Mechatronics, Shop Floor Simulation: A Step towards Smart Factories
Identification Tehnology, Industry 4.0 and Education. in Industry 4.0, in Industry 4.0 - Impact on Intelligent
in Proceedings of the 18th IEEE International Symposium Logistics and Manufacturing, Tamás Bányai (Ed.),
INFOTEH-JAHORINA, 20-22 March, 2019. Chapter 2, IntechOpen, 2020.
87
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(024)
Summary: Industry 4.0 leads to a strong digitalization of industrial processes, but also a significant increase in communication
and cooperation between the machines that make it up. This is the case with intelligent autonomous vehicles (IAVs) and other
cooperative mobile robots which are multiplying in factories, often in the form of fleets of vehicles, and whose intelligence
and autonomy are increasing. The implantation and deployment of IAVs fleets raises several challenges: acceptability by
employees, vehicle location, traffic fluidity, vehicle perception of changing environments. Simulation offers a good framework
for studying solutions for these different challenges. Thus, we propose in this paper the extension of a collision detection
algorithm to deal with the obstacle avoidance issue. The conclusive simulation will allow us to experiment in real conditions.
88
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
concerned. They are important messages that would autonomy. The individual autonomy of the IAVs
complete the range of possible messages to be strengthens their collective autonomy.
exchanged to cooperate and avoid collisions for the The messages exchanged by the different IAVs
IAVs in the Bahnes et al. algorithm [2]. remain consistent with the Bahnes’s algorithm. We just
propose two new message types for the collaborative
perception (added to the three messages defined by
3. Algorithm Improvement Bahnes et al.: Hello_Msg, Coop_Msg and Ack_Msg):
• Obstacle_msg: a message sent by an IAV agent to
The collision avoidance algorithm of [2] makes it the other IAV agents circulating in the warehouse
possible to deal with the priority of different vehicles to indicate the presence of a perceived obstacle.
when approaching an intersection. However, it does not • Alert_msg: a message sent by an IAV agent to the
deal with the problems of detection, communication other IAV agents circulating in the warehouse to
and avoidance of fixed or moving obstacles (e.g. indicate an unavoidable obstacle.
human operators).
We extend the algorithm of Bahnes et al. to handle
the presence of fixed or moving obstacles in Fig. 1. 4. Simulation Results
Then, we simulate the algorithm staying within the
framework of the three scenarios proposed by [2]. The traffic plan presented in Fig. 3. It shows the
These simulations rely on an agent-based model where different scenarios that we consider as a benchmark
IAVs are identified [13, 14]. Indeed, agent-based plan to compare results. Ten IAVs are distributed over
simulation for IAVs is the most common in the same 3 circuits: the red IAVs on the first circuit, the blue ones
way as simulations based on discrete events or robotics on the second and the yellow ones on the third.
software [15]. It involves different intersections, where vehicles
IAV agents have the ability to exchange messages can arrive from different sides like in a warehouse (4
and are equipped with radar. This allows them to detect intersections are shown in Fig. 3.). Thus, it provides the
vehicles in front of them. For instance, given an IAV different characteristics of an industrial environment
agent ai, if another IAV agent aj in front of it is stopped and allows us to realize simulated experimental tests in
or travelling at a slower speed, the IAV agent ai can line with realistic scenarios.
detect it with its radar and stop accordingly to avoid We notice in the simulation that the avoidance is
hitting it, as in Fig. 2. well respected and the obstacles are perceived by the
To improve the collective autonomy of the IAVs it IAVs. Therefore, the simulation validates the extended
is essential that they have a good capacity for individual Bahnes’s algorithm with collision avoidance and fixed
or dynamic obstacle detection processing.
Fig. 1. Improvement of Bahnes’s algorithm to treat the problem of collision and obstacles
89
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Fig. 2. Simulation of radar use: (a) at the top of the picture: one blue and three yellow IAVs arrive near the intersection, (b)
while waiting for the yellow IAV to pass the intersection, the radar of the blue IAV and the two other yellow IAV allow
them to stop and keep their distance to avoid colliding
Fig. 3. Simulation of the scenarios: (a) in the center of the picture: a blue and yellow IAV arrive at an intersection,
(b) the yellow IAV passed the intersection after communicating with other IAVs, (c) on the left side of the picture: a red IAV
perceives a fixed obstacle in front of him, (d) a red IAV avoided the obstacle.
90
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
91
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(025)
Summary: The emergence of artificial intelligence (AI) increases human abilities in solving cognitive tasks, expands the
scope of measurement as well as forces a rethinking of the basic philosophical question about the material and the spiritual
(ideal). The development of AI-based systems will lead to mass production of their various types with relatively weak AI,
which, along with systems with strong AI, will change the life of society but create social problems. The paper describes the
effectiveness of using measurement systems with AI in manufacturing, medicine, transport, assessment of the product and
service quality, determination of human abilities, etc. The directions of activities that can accelerate the positive trends in the
new development stage but will delay the emergence and facilitate the solution to related social problems are highlighted.
Keywords: Artificial intelligence, Measurement, Weak and strong artificial intelligence, Development of civilisation, Voice
assistant.
92
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
outcome of data analysis, the AI can identify socialization. A technical revolution of this kind is
correlations between the data, recognize images and inevitable, as it opens up new prospects for the
measure multidimensional quantities, classify them, development of civilization.
predict changes in the ratios of various parameters over Naturally, there will be a code of ethics for the AI-
time, and make decisions or propose them to a human based systems that should exclude undesirable
operator. consequences of their socialization. Significantly, in
Nowadays, as a rule, the human operator 2021, the Code of Ethics for Artificial Intelligence was
formulates the tasks. In many cases, the AI finds the published in Russia [2]. It establishes “general ethical
algorithm to solve them, sometimes without informing principles and standards of behaviour to guide
the operator about it. However, in a few years, humans participants in AI relationships ... as well as of
will only set strategic tasks. Tactical tasks, e.g., corresponding mechanisms for implementing the
suppression of interfering influences or provisions of this Code.” The Code applies to relations
troubleshooting, will be handled by the AI related to ethical aspects of the development (design,
independently. engineering, piloting), implementation, and use of AI
Thus, while in the past technical systems were technologies at all stages of the life cycle...” The Code
created to increase the capacity of individual human states that “human beings, their rights, and freedoms
organs (hands, feet, eyes, etc.) and to speed up the shall be regarded as the highest value”.
solution to simple mathematical problems, the However, it is hardly possible that such a document
development of AI opens up the prospect of increasing will be able to fundamentally change the direction of
the human abilities to solve tasks of a cognitive nature. robotics development. As the number of robots
Since the level of complexity and variety of tasks capable of emotional responses increases above some
to be performed by systems based on AI is different, it limit, more developed socialization is likely, possibly
is reasonable to produce a number of system types that involving interested humans.
differ in terms of AI level. Then we can expect the spontaneous selection of
It is now common to divide AI into weak (narrow) leaders. They will be able to set tasks for groups of
and strong (general). Systems based on weak AI robots. This does not exclude tacit correction of a robot
extract information from a limited set of data and are firmware with their participation. As a result, there
only able to cope with the specific tasks that they were have been assumptions about the possibility of
trained to perform. They are usually oriented towards conflicts between robots and humans [3]. The history
tasks that, when carried out by a human, do not require of civilization provides numerous examples of the
high qualifications but oblige him/ her to use regularly emergence of similar situations in the relationships
updated reference data. Often, communications of between human communities. Usually, they concern
similar systems with humans, i.e. conversational skills violations of ethical laws established by both religions
with a limited vocabulary, are also required. Such tasks and state laws.
include, e.g., working as a waiter/waitress, security
guard, tour guide, interpreter, storekeeper, etc. It
should be noted that the systems based on weak AI are 3. AI-based Measurement Systems
now relatively widespread, while the number of their
types and manufactured product output are growing The problem of preventing conflicts between
steadily. robots and humans is not yet a hot topic, although it
A system with strong AI is much more expensive. should not be forgotten. Metrology specialists now
It can reason, make judgements in the face of face a different challenge: efficiently apply AI for
uncertainty, plan, learn, integrate prior knowledge into measurement purposes.
decision-making, suggest new ideas, etc. It can be The first direction of work is the identification of
assumed that systems with strong AI will also be multidimensional quantities that characterize the
produced in several types, depending on the specific occurrence of deviations from nominal values,
tasks they solve. measurement of the parameters of these quantities, and
However, it is possible that in the process of the AI- determination of the deviation dynamics over time.
based system operation, such as mobile robots, there The next stage is the development of methods and
will happen a need to coordinate the actions of several means that allow obtaining a desirable result. The
robots to solve a new problem. In particular, it will be essence of the result depends on the task to be solved.
necessary to perform the actions simultaneously or There exist several groups of tasks here.
jointly process information coming from many One of them is to search for irregularities in the
sources, the number of which is greater than that of materials used in critical products, e.g., to detect a
channels a single system with AI has. hidden crack in the support of a construction.
Apparently, the robot firmware will allow for this Conventional methods of ultrasound analysis can
kind of cooperation. For example, a robot will detect the presence of irregularities using a panoramic
“realize” that on its own it cannot solve the task set by scanning. If the results of the multichannel
an operator and will initiate the cooperation. measurements made with the help of an AI-based
Essentially, this proactive request for help will mean system show a correlation of changes, in particular, in
an emotional response, which, if perceived by another many space points, under some impacts on the support,
robot, will mark the beginning of an era of robot
93
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
or over time, then the presence of a crack in the support particular job, the emotional expressiveness of the
is diagnosed. This fact makes it possible to prepare a musical composition performance, etc.
solution aimed at eliminating the hazard. However, in some cases, an expert can be biased,
Another group of tasks is related to medicine, i.e., i.e. interested in an assessment for reasons unrelated to
the detection of abnormalities in human physiological the materials submitted to him.
systems. In [4], on the basis of the electrocardiogram The trustworthiness improvement of
processing, Uspenskiy proposes to use AI for “measurements” being performed with experts’
analyzing the ratios between blood pressure pulse participation up to an acceptable level that excludes
parameters got at several points on the patient’s body crude errors is the second direction of metrological
surface. Statistics of clinical studies give the basis for research and development that relies on the capabilities
the automatic diagnosing of more than forty diseases of AI-based measurement systems.
by characteristic deviations of measured ratios from To minimize the role of subjective factors, expert
the nominal values fixed in healthy people. The judgements should be reduced to a common scale.
diagnosis, in its turn, determines the course of Selecting the experts, one should take into account the
treatment to restore the health of the patient. competence level bringing the expert number to the
This method also opens up additional number that depends on the task to be solved. This
opportunities. As we all know, certain music requirement can be met using AI. The useful method
influences the physiological processes taking place in is, e.g., the correlation analysis of experts’ answers to
the body [5]. By analyzing changes in the above ratios, abstract questions combined with a number of
while a patient is listening to selected music excerpts, indicators characterizing the features of experts’
it is possible to refine the diagnosis and treatment neurophysiological activity (brain biorhythms, voice
recommendations. sound, etc.) while making their assessments.
Diagnosing in this way is also technically relatively The Bayesian approach has been successfully
easy with exposure to force, vibration, heat, or applied in this field too. Examples are given in [7].
otherwise. This feature makes it possible to help The third direction is the measurement of abilities
patients having a significant likelihood of abnormal and properties of intelligence, including the emotional
physiological processes. sphere.
It is sufficient to arrange measurements of a Usually, the level of human intelligence is assessed
specially selected group of indicators, which form, by the intelligence quotient IQ. It enables the
using an AI-based system, an estimate of the value of comparison between individual results and average
the quantity indicating the approach of an attack. The ones for people of the same age. The assessment is
corresponding signal should be transmitted to the made with the help of tests, i.e., a series of
doctor if the patient is in hospital or to the patient progressively more difficult questions chosen by
himself/herself if he/she has a means of preventing a experts. There exist widespread tests, but the authors
seizure. Such methods can diagnose diseases at early are not aware of any standard in this field. The reason
stages too. seems to be the limited interpretation of intelligence in
The considered direction of AI-based measuring this method and its emphasis on logic.
systems can be effective in the automatic translation of However, the role of the emotional sphere in a
emotionally coloured speech [6]. Special speech human intellectual activity is significant. Analysis has
processing can single out infrasound multidimensional shown that AI-based systems make it possible to
modulation components and identify, e.g., the irony of identify abnormalities in intelligence development,
the said. e.g., from the sound of babies’ cries, i.e. his/her
AI can also be efficient in traffic management. The emotional reactions. The abnormalities can be either
relevance of work in this area increases with the negative, which indicates, e.g., the early stages of
appearance of driverless cars. But also for driving autism, or positive related to the accelerated
conventional cars in regular traffic flows, AI-based car development of imaginative thinking.
measuring systems will certainly be in demand. They The ineffectiveness of IQ in investigating human
will be able to track the direction and speed of abilities does not undermine the merits of the test
oncoming and passing cars, calculate the probability of method. The application of tests including not logical
collision with them or with other obstacles, and form components only but emotional ones may be efficient
signals correcting the car movement. in evaluating AI-based systems in case communication
The above examples concern the application of with humans is one of their primary functions.
measurements in areas where they were necessary Performing the duties of a waiter, salesperson,
before but could not provide the required efficiency security guard, or tour guide could be the examples
without AI. where the “voice assistant” function can be very
However, in some fields, the term “measurement” appropriate. An IQ assessment with such a test makes
can so far only be applied conventionally. This it possible to compare the quality of this function in
statement relates to the assessment of systems produced by different companies.
multidimensional quantities in which the role of an In particular, in 2019, Loup Ventures prepared 800
expert is essential. Such quantities include many questions assigned for testing “voice assistants”
indicators of the quality of goods and services, the Amazon Alexa, Apple Siri, and Google Assistant [9].
level of knowledge and skills of applicants for a They checked the understanding of the question and
94
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
95
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(026)
S. V. Prokopchina
Financial University, Russian Federation
E-mail: [email protected]
Summary: Within the framework of the Industry 4.0 concept, intensive development of the processes of intellectualization of
sensor systems is envisaged. Among the most important specific properties of real measuring processes in complex systems
is, first of all, their implementation under conditions of considerable uncertainty. The uncertainty is caused by a priori
incompleteness, inaccuracy, vagueness of information about a complex measuring object and its functioning environment,
which does not allow to build an adequate model of the object before the measurement experiment, to identify and formalize
the influencing factors of the external environment and to develop effective algorithms for the functioning of information and
measurement systems.
The report proposes an approach to the intellectualization of measurement systems in conditions of uncertainty by creating
intelligent sensor networks based on Bayesian intelligent technologies (BIT) and means of their implementation. Typical
modules of such networks are considered, which are integrated sets of various sensors and intelligent measurement information
processing systems. The results of the networks are comprehensive assessments of the state of complex objects and
recommendations for providing of their sustainable functioning. An important part of such systems is the built-in means of a
complete metrological justification of all received solutions. The systems have a hierarchical architecture, according to the
levels of management of complex objects, which has the possibility of self-development based on newly received information.
This is achieved thanks to models and scales with dynamic constraints on which all BIT algorithms are built. The report
provides examples of the use of intelligent sensor networks for monitoring and control of power generation and water supply
systems.
96
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
The experience of such developments made it Q is the generalized algorithm for creating a typical
possible to form a typical peripheral sensor network sensor network module;
module that implements all the above functions for a Q1 is the creation of a conceptual model of an object
separate network section. with dynamic constraints (MDO) for this section of the
It consists of the following subsystems: sensor network;
integrated sensor sets that measure system Q2 is the Sensor system selection;
parameters; Q3 is the construction of scales with dynamic
intelligent controllers that implement the functions constraints for each sensor in the intelligent controller;
of intelligent processing of primary measurement Q5 is the implementation of measurements, obtaining
information, consisting in the integration of primary measurement information from sensors;
information from individual sensors in order to Q6 integration of individual measurements of various
obtain estimates, and the coordination of protocols sensors according to the modified Bayes formula;
for transmitting solutions received on this part of Q7 interpretation of integrated information and its
the network to the central module; presentation in the form of an assessment of the state
subsystems of interfaces and web services in of a network section;
monitoring mode; Q8 is the determination of dynamics and dynamic
databases of primary measurement information. models, trends in individual parameters and in general
The type of a typical sensor network module is according to the state of the system in this section of
illustrated in Fig. 1. A conceptual model for obtaining the network and trends in the development of
a solution in a peripheral sensor network module based situations;
on Bayesian intelligent technologies (BITS). Q_9 is the assessment of risks and potentials of the
The conceptual model of creating a solution based situation and display by methods of cognitive graphics;
on Bayesian intelligent technologies (BITS) is based Q10 is the interpretation of the current situation;
on the methodology and principles of creating models Q_11 is the generating recommendations to improve
with dynamic constraints based on the regularizing the situation.
Bayesian approach [1-3]. The stages of the complex Convolution of information (in formulas (1) and
solution of the problem of creating an intelligent big (3) this action is indicated by the symbol *) about the
data processing system can be represented by the value of the element x_ij, i=1,...I is performed
following model: according to the modified Bayes formula for discrete
laws of hypothesis distribution in the form of
Q=Q_1*Q_2*Q_3*Q_4*Q_5*Q_6*Q_7*Q_8*Q_9* membership functions (linguistic scale of BII for
Q_10*Q_11 (1), processing non-numeric, qualitative information) or
probability distributions (numerical scale of BII for
where: processing non-numeric, qualitative information). The
97
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
probability of a solution (hypotheses about the state of 2. Conceptual Model for Monitoring the State
an object, the value of a parameter, or the like) is of the Water Supply Network
calculated by the formula:
The conceptual model for monitoring the state of
𝜇 ℎ ℎ | 𝑋 |𝑄 the I-th section of the water supply network based on
| | ° ℎ , , 𝑥 , | 𝑋 |𝑄 Bayesian intelligent technologies (BIT) has the
2 , following form:
∑ , , | , , | | |
Qi=Q_1*Q_2*Q_3*Q_4*Q_5 (3),
where:
ℎ is the integral (for a set of information flows 𝑋 ))
where:
regularizing Bayesian estimation (RBO); Qi is the integral assessment of the condition of the
ℎ is the RBO for X_i data stream; water supply network section;
𝜇 ℎ ℎ | 𝑋 |𝑄 is the posteriori probability of Q_1 is the integral assessment of the condition of the
evaluation ℎ . well on the section of the water supply network;
The general scheme of the system being developed Q_2 is the integral assessment of the condition of a
consists of: a measurement storage server, sensors; a linear section of the water supply network ;
controller; measurement processing and transmission Q_3 is the integral assessment of the condition of the
to the server; a workstation; a neural network for external sections of the water supply network;
generating scripts; a web service; clients; IIS Server Q_4 is the integral assessment of the network status
Windows Infoanalytics; MS SQL; a mathematical based on information from users;
apparatus for data analysis and decision-making; Q_5 is the integral assessment of the network status
system management. based on information from experts.
A typical peripheral module of an intelligent sensor The conceptual model of monitoring the condition
network, as shown in Fig. 1, includes basic measuring of the water supply network as a whole can be written
instruments and an intelligent controller, as well as the as follows:
necessary modules for organizing web services.
Measuring instruments may include both sensors for Q = * i=1,I Qi (4)
measuring physical parameters and virtual measuring
instruments for measuring integral, non-physical and Such methods and technologies are used in the
non-quantitative parameters. This architecture of a implementation of real projects. In particular,
typical module allows you to combine, integrate and examples of applied solutions of an intelligent sensor
process measurement and observation data in network for assessing the state of the water supply
monitoring mode. The use of special scales with network of one of the cities of the Russian Federation
dynamic constraints [1, 2] allows measurements to be are given below.
made under conditions of uncertainty, significant Fig. 2 shows the interpretation of the situation
dynamism of the measured parameters and the active according to the integral factor "Information from
influence of the external environment. metering devices" on 12.11.2021.
98
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
99
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(027)
Summary: An important place in modern methodologies and technologies of personnel management at the enterprise is
occupied by the direction of innovative development, culture of innovation. For this direction, from the position of
intellectualization of innovative processes and, in particular, motivation and involvement of personnel, non-quantitative
indicators are characteristic, which, following the ISO series standards, should be measurable.
This led to the use of Bayesian intelligent technologies, in particular Bayesian intelligent measurements, for the purpose
of measuring indicators of the state of the company's personnel and their management. The aim of the work is to build a
systematic methodological basis and means of creating a digital environment for the implementation of the principles of the
culture of innovation based on fundamental and best practical psychological work by means of BIT and BII. The
implementation of a number of methodological constructions based on individual modules of the Infoanalytic computer system
is shown, which contributes to the digitalization of this activity, the dissemination of experience and knowledge for the training
and continuous training of specialists, including remotely.
100
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
conditions of information uncertainty. The that affect the organization of labor at the present time
effectiveness of the proposed approach is confirmed by and should be measured:
its methodological principles focused on information 1. The sixth technological order;
processing in conditions of uncertainty 2. Globalization and decentralization of work;
(incompleteness, fuzziness and inaccuracy of 3. New behavior;
psychological information) and the practice of its 4. The growth of creative professions
application to create intelligent systems implemented All these global trends will only intensify in the
for a wide range of tasks related to monitoring and future, but it is innovative companies that have already
management of complex objects, which include the learned how to use them in their work and benefit from
above innovative processes. these total changes.
In the course of this scientific work, the following Technoclade is a set of generally accepted and
results were obtained. applied production technologies and scientific and
1. The terminological basis of the theoretical technological progress
provisions of the culture of innovation and motivation The technological order is not just one change, but
is determined. a whole series of related and complementary changes.
2. A historical review of the existing concepts of The sixth technoclass is primarily associated with the
the implementation of psychological innovations and synthesis of sciences. It is associated with
motivation is carried out. breakthroughs in the field of artificial intelligence and
3. A critical review of methodological approaches robotics. The company's ability to adapt to new
and best practices for the implementation of innovation technological conditions will directly affect its
processes and motivation of individuals or groups of viability. The simplest example associated with the
subjects in accordance with the formulated emergence and spread of the Internet, we saw in the
requirements of methodological and practical fifth technoclass.
relevance. Our education and the structure of social relations
4. The conditions and factors that favor and hinder are still in the fourth, industrial technological order.
the implementation of innovative psychological They are focused on working in a hierarchy,
techniques and motivation of subjects are formulated consistency and efficiency, having an ideal sample and
5. The regularizing Bayesian approach and working according to a template. But the sixth techno-
Bayesian intellectual technologies based on it are layout allows using robotics, the Internet, things and
reasonably chosen for the implementation of the new analytical abilities of computers to replace not
principles of psychological development of subjects only manufacturing professions, but also some office
and communities, including labor. intellectual tasks. We are already seeing an actively
6. The methods of implementation of developing market of bots capable of recognizing and
psychological innovations and motivation based on the solving a client's problem. This means that the
conclusions and methodological decisions made in the company's personnel will have to learn not only to use
work are proposed. new technological capabilities, but also to move from
The practical significance of the work is: purely performing activities to creativity and task
1. Developed practical methods and setting. Innovative business has such experience. He
recommendations for the implementation of the develops creativity in his employees, supports the
principles of the culture of innovation, as well as, as model: "Need - training - development - need - training
part of it, the motivations of subjects that have been - development".
tested on a number of communities. including labor. All this means that companies are turning into
2. Conducted psychological projects and universities and colleges. The performing discipline
experiments, the results of which can be used as inherent in rigid industrial structures is being replaced
effective practices. by creative and intellectual work. Enterprises should
3. Implementation of a number of methodological become a platform for continuous development:
constructions based on individual modules of the develop mentoring competencies, implement training
Infoanalytic computer system, which contributes to the tools and various professional collaborations.
digitalization of this activity, the dissemination of It was possible to identify six integral indicators of
experience and knowledge for the training and the state of the enterprise's innovation sphere, which
continuous training of specialists, including remotely. were broadcast unambiguously and were found in the
The results obtained make it possible to optimize absolute majority of innovative organizations.
the personnel management system in the organization, They include: Transformation, Self-organization,
as well as to predict the nature of the impact of certain Integrity, Diversity, World-centricity, Creation
managerial influences on the level of satisfaction, (entrepreneurship). These indicators together form the
involvement and effectiveness of employees of the basis for the development and implementation of
organization. innovations. They interact and complement each other,
In this study, we turned to the experience of leading creating a catalyst effect for breakthrough solutions,
consulting companies, as well as experts, and based on discoveries, research and technical creativity.
this critical review, we identified four global changes
101
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
102
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Fig. 1. Implementation of the conceptual model of the innovation sphere in the "Infoanalyst" environment.
103
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
(028)
Summary: The article deals with the problem of searching for leaks from pipelines by receiving and processing acoustic
emission signals that occur when the liquid flows at the leakage site. The purpose of the work is to review existing solutions
for acoustic monitoring, determine the main characteristics of acoustic emission signals and assess the possibility of their
recognition in the monitoring process using artificial neural networks.
Keywords: Acoustic emission method, Acoustic leak detector, Correlation leak detector, Acoustic monitoring, Spectral
analysis, Correlation analysis, Artificial neural networks.
104
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
2. Problem Statement and Review of may be other acoustic vibrations - industrial, transport
Acoustic Monitoring Equipment and natural interference, as well as noise created by the
liquid when interacting with various inhomogeneities
As a result of exposure to adverse factors (flange joints, growths, etc.) inside the pipeline (see
(corrosion, overpressure, extraneous mechanical Fig. 1).
influences, manufacturing defects, etc.), holes and The most common acoustic search for leakage is by
cracks appear in the initially continuous wall of the the signal of the sensor moving on the surface of the
pipeline, through which the leakage of liquid under ground near the pipeline and connected to an acoustic
pressure begins. Leaks are routinely classified by flow receiver in the hands of the operator (sensor D3 and
rate per unit of time: receiver AP in Fig. 1). When the sensor is found
- "very weak" - up to 1 l / min; directly above the leakage site, the signal recorded by
- "weak" - from 1 to 5 l / min; the acoustic receiver will be the greatest. this method
- "average" - from 5 to 15 l / min; is a classic "maximum method", then its error is
- "strong" - more than 15 l / min. relatively large and is ~ 1 m at a depth of penetration
When the liquid jet flows out, acoustic vibrations of the pipe pipeline (1.5 ... 2) m. The appearance of
(acoustic emission) occur, which spread in the soil, common acoustic leak detectors is shown on the
liquid and pipeline walls and can be recorded by Fig. 2.
acoustic sensors. Also in the pipeline and near it there
A smaller error in determining the leakage site is instrumental error of correlation leak detectors is 0.1 m
provided by the correlation method, in which two or less. The appearance of common correlation leak
sensors are used, which are installed, as a rule, on the detectors is shown on Fig. 3.
pipeline on both sides of the presumed leakage Unfortunately, correlation leak detectors also have
(sensors D1 and D2 in Fig. 1). Signals from the sensors drawbacks that can significantly reduce the accuracy
are amplified and transmitted by the UP1 and UP2 of the search for a leakage site, including the
transmitters to the KP correlation receiver, which following:
calculates the correlation function of the two signals (a) Almost always, the correlation function
and determines the time shift between the signals to its calculated from signals from watchers has several
maximum. Knowing the distance between the sensors peaks, the largest of which does not necessarily
and the speed of the call in the walls of the pipeline, it correspond to a real leak. Acoustic interference from
is possible to determine the coordinates of the leakage mechanisms of pumping stations, elevators, valves,
site with high accuracy. The typical value of the etc. can spread in the pipeline. The operator has to
105
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
manually, by the method of subsequent approximation, depending on the material and diameter of the pipe. At
select the frequency band of signal analysis in order to the same time, the speed of sound depends on a number
build up from the crunching signals. of other parameters, especially on the thickness of the
b) The section of the pipeline between the sensors pipe walls. The thickness of the walls of worn pipes
may have bends and taps, due to which the distance can differ from the original several times, as a result of
between the sensors may differ from that entered into which the difference between the real significance of
the correlation receiver. Most regulatory documents the speed of sound and the given one can be up to 30%!
contain a strong recommendation to trace the surveyed To reduce this error, many correlation leak detectors
section of the pipeline using a tracer before starting implement a mode for measuring the speed of the call
work. by solving the inverse problem - determining the delay
c) The source of the greatest error can be the between signals for a source with pre-known
calculated value of the speed of sound in the walls of coordinates.
the pipeline, which, as a rule, is set in tabular form
Despite the disadvantages listed above, portable case for transportation and programming and software
acoustic and correlative leak detectors provide control, which allows you to determine the correlation
acceptable results in most cases and are widely used in between the signals of any pair of sensors (see Fig. 4a).
organizations operating heat and water supply The practicals of using Enigma are as follows: during
networks. Less common are more expensive multi- the day, the dutchiks are installed on the pipeline, the
position leak detectors, which include several sensors sensors are programmed to start at night (when the
at once, which can be placed on the pipeline according level of sidely noise is minimal).The next day, the
to a given Schema. The sensors are synchronized with sensors are removed, information is read from them,
each other and simultaneously record acoustic signals, then the sensors are moved to a new section of the
storing them in their own memory. The data can then pipeline. The internal power supply of the sensors
be read on the PC for further processing. ensures operation for 5 years. The sensors themselves
A certain fame in domestic organizations received are water-resistant, can be operated at a depth of up to
a multi-position leak detector Enigma company 10 meters.
Primayer (UK). It consists of 8 self-powered sensors, a
106
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
The Enigma software allows you to calculate the the road cost 45350 dollars. Across the U.S. as a whole,
correlation functions between the signals of any pair of it is estimated that U.S. utilities spend more than $900
sensors (a total of 28 possible combinations) and million annually to fix water leaks and repair pipe
evaluate the number or absence of leakage in the form ducts.
of the correlation function. Correlation functions with The capabilities of geographically distributed
a minimum background level and a distinct maximum monitoring systems make it possible to implement
are considered to correspond to a leak with a high more complex and efficient algorithms for determining
degree of probability and are marked with a red fill the facts and locations of leakage. Thus, [1] data on the
color (see Fig. 4b). if the background level of the use in Toronto (Canada) of a system of "smart" sensors
correlation function is higher, the fill color will turn and artificial intelligence to determine "anomalous"
yellow (the probability of leakage is average), with a noise in the pipeline, different from the norm and
very high background level, the fill color is blue corresponding to leakage, are given. The server of the
(probably leakage accuracy is small). system analyzes the data received from the sensors and
An even more expensive, but also more effective sends the result to the central control room, which
method of detecting leaks in the present time is the automatically forms a request for repair and sends the
continuous monitoring of communications using repair team.
territorially distributed systems, the sensors of which In general, the above acoustic monitoring tools
are constantly installed in wells, on hydrants, in make it possible to make an introduction that in this
thermal chambers, etc. Reading information from such area, as in many others, there is a transition from
sensors is carried out remotely, as a rule, using traditional "semi-intuitive" methods of work to the use
GPS/GPRS or LoRaWAN channels. The most of intelligent computing systems that can allow:
widespread such systems have received abroad, where - Reduce the routine burden on operators;
with their help they expect to significantly reduce - Reduce the time to find and clarify the place of
losses from water leaks. leakage;
In 2019, Primayer entered into a £40 million ($53 - Reduce the cost of repairing communications.
million) framework agreement with Anglian Water, a To implement intelligent processing algorithms, it
utility in east England. Water utilities in England and is necessary to study the defining characteristics of
Wales are encouraged to reduce losses by at least 15% leakage signals in the time and spectral domain. The
between 2020 and 2025 and halve them by 2050 [4]. obtained characteristics will make it possible to assess
As part of the agreement, about 3500 Enigma3hyQ the hardware characteristics of the complex of
intelligent sensors using multi-point noise correlation processing devices and the necessary methods and
technology will be installed on the Anglian Water algorithms for decision-making.
water supply network. To visualize the locations of the
leak, the location of each sensor is displayed on Google
Maps and Street View. An example of the installation 3. Experimental Studies and Analysis
of Primayer sensors is shown in Fig. 5. of the Characteristics of Acoustic
Emission Signals
107
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
108
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
If spectral analysis shows how much the signal is ANNs to determine the fact of the presence or absence
"similar" to a set of sinusoids, then correlation analysis of leakage.
allows you to assess the similarity of the signal with Based on the results obtained, it can be
itself, i.e. to what extent it and the physical process that preliminarily concluded that the marked features of the
generates it are random. The autocorrelation functions spectrum and the autocorrelation function of the
(ACFs) of the recorded acoustic emission signals acoustic emission signal make it possible to determine
computed in MATLAB 2012 are shown in Fig. 9. the fact of the presence or absence of leakage. When
The following features of the ACF obtained can be leakage occurs, the spectrum and ACF of the signal
noted: acquire a characteristic form, which allows us to talk
1) In the presence of leakage, ACF has a more about the feasibility of their recognition using neural
periodic structure, with a shorter period for weaker network algorithms.
leaks, and a longer period for stronger ones;
2) In the presence of a leak, the first side petal of References
the ACF (marked in the figure with a red circle) is
noticeably higher than that of the ACF signal without [1]. https://www.everest.ua/ru/v-kanade-potery-vody-yz-
leakage. Also, the level of the side lobe of the ACF gorodskyh-trub-obnaruzhyvayut-s-pomoshhyu-
changes slightly when the amount of leakage changes. yntellektualnyh-datchykov, 26. 02.2021.
[2]. https://files.stroyinf.ru/Data1/46/46843/
index.htm#i21807, View date 26. 02.2021.
4. Conclusion [3]. http://www.vstmag.ru/ru/archives-all/2010/2010-
4/263-obnaruzhenije-skrytyh, View date 26. 02.2021.
The article discusses the main varieties of [4]. https://www.waterworld.com/technologies/flow-
equipment for acoustic leakage - acoustic, correlation level-pressure-measurement/article/16190942/
and multi-position leak detectors and geographically revolutionizing-remote-leak-identification, View date
26.02.2021
distributed acoustic monitoring systems, presents the [5]. https://www.waterworld.com/home/article/
results of spectral and correlation analysis of various 14070815/how-to-nix-nonrevenue-water-with-
leakage signals and assesses the possibility of using acoustic-leak-detection, View date 2/26/2021.
109
2nd IFSA Winter Conference on Automation, Robotics & Communications for Industry 4.0 (ARCI’ 2022),
2-3 February 2022, Andorra la Vella, Andorra
Fig. 9. Autocorrelation functions of acoustic leakage emission signals recorded in the experiment.
110
View publication stats