applied

sciences
Article
Methods of Pre-Identification of TITO Systems
Milan Saga 1 , Karel Perutka 2 , Ivan Kuric 1 , Ivan Zajačko 1, * , Vladimír Bulej 1 , Vladimír Tlach 1
and Martin Bezák 3

1 Faculty of Mechanical Engineering, University of Zilina, Univerzitna 8215/1, 01026 Zilina, Slovakia;
[email protected] (M.S.); [email protected] (I.K.); [email protected] (V.B.);
[email protected] (V.T.)
2 Faculty of Applied Informatics, Tomas Bata University in Zlin, Nad Stráněmi 4511,
760 05 Zlin, Czech Republic; [email protected]
3 VIPO, a. s., Gen. Svobodu 1069/4, 95801 Partizanske, Slovakia; [email protected]
* Correspondence: [email protected]; Tel.: +421-910956861

Abstract: The content of this article is the presentation of methods used to identify systems before
actual control, namely decentralized control of systems with Two Inputs, Two Outputs (TITO)
and with two interactions. First, theoretical assumptions and reasons for using these methods are
given. Subsequently, two methods for systems identification are described. At the end of this
article, these specific methods are presented as the pre-identification of the chosen example. The
Introduction part of the paper deals with the description of decentralized control, adaptive control,
decentralized control in robotics and problem formulation (fixing the identification time at the
existing decentralized self-tuning controller at the beginning of control and at the beginning of
any set-point change) with the goal of a new method of identification. The Materials and methods



section describes the used decentralized control method, recursive identification using approximation


polynomials and least-squares with directional forgetting, recursive instrumental variable, self-tuning
Citation: Saga, M.; Perutka, K.; controller and suboptimal quadratic tracking controller, so all methods described in the section are
Kuric, I.; Zajačko, I.; Bulej, V.; Tlach, those ones that already exist. Another section, named Assumptions, newly formulates the necessary
V.; Bezák, M. Methods of
background information, such as decentralized controllability and the system model, for the new
Pre-Identification of TITO Systems.
identification method formulated in Pre-identification section. This section is followed by a section
Appl. Sci. 2021, 11, 6954. https://
showing the results obtained by simulations and in real-time on a Coupled Drives model in the
doi.org/10.3390/app11156954
laboratory.
Academic Editors: Yuri Nikitin and
Anton Civit
Keywords: pre-identification; least squares method; instrumental variable method; robotics; sensor

Received: 21 April 2021

Accepted: 23 July 2021
Published: 28 July 2021 1. Introduction
Most processes in practice are processes that have multiple inputs and multiple out-
Publisher’s Note: MDPI stays neutral puts, and these are influenced by interactions. These systems can be controlled by a
with regard to jurisdictional claims in centralized or decentralized controller. The main advantages of decentralization include
published maps and institutional affil- simplifying the overall task by dividing it into a set of sub-tasks. Decentralized control
iations. is very often used in practice. Decentralized charge control of electric vehicles is a nice
application. There was introduced a fully decentralized and participatory learning mecha-
nism for privacy-preserving coordinated charging control of electric vehicles that regulates
three Smart Grid socio-technical aspects: (i) reliability, (ii) discomfort and (iii) fairness [1].
Copyright: © 2021 by the authors. Another good application of decentralized control is a quadrocopter, namely outdoor
Licensee MDPI, Basel, Switzerland. flocking of quadcopter drones with decentralized model predictive control [2]. From the
This article is an open access article theoretical point of view, there was proposed a decentralized explicit (closed-form) itera-
distributed under the terms and tive formula that solves convex programming problems with linear equality constraints
conditions of the Creative Commons and interval bounds on the decision variables [3], or decentralized control problem for
Attribution (CC BY) license (https://
non-affine large-scale systems with nonaffine functions possibly being discontinuous [4], or
creativecommons.org/licenses/by/
decentralized adaptive tracking control strategy consisting of a steady-state controller and
4.0/).

Appl. Sci. 2021, 11, 6954. https://doi.org/10.3390/app11156954 https://www.mdpi.com/journal/applsci

Appl. Sci. 2021, 11, 6954 2 of 15

modified optimal feedback controller. Design parameters-dependent feasibility conditions

were formulated by using Lyapunov theory to guarantee the existence of the proposed
decentralized control scheme [5]. Decentralized voltage control is another example of a
decentralized strategy. It includes network partitioning strategy for the optimal voltage
control of Active Distribution Networks actuated by means of a limited number of Dis-
tributed Energy Storage Systems [6]. Another paper is concerned with the problem of
decentralized event-triggered dynamic output feedback control for large-scale systems
with unknown interconnections. By using a modified cyclic-small gain condition and
introducing a free-matrix-based integral inequality, a sufficient condition was derived to
ensure that the overall closed-loop system is asymptotically stable with a prescribed H∞
performance [7]. There was also implemented fuzzy decentralized control, for example
an adaptive fuzzy decentralized control approach for a class of nonlinear systems with
unknown control direction and different performance constraints. In the control design, the
different performance constraints, that were the prescribed performance error constraints
for some subsystems and the asymmetric time-varying output constraints for the others,
could be unified as one form by selecting proper performance functions [8]. Adaptive
control is another area that has expanded the use of decentralized control, and here are
at least a few such examples. A minimal-neural-networks-based design approach was
presented for the decentralized output feedback tracking of uncertain interconnected strict-
feedback nonlinear systems with unknown time varying delayed interactions unmatched
in control inputs [9]. The decentralized output-feedback adaptive backstepping control
scheme was also proposed for a class of interconnected nonlinear systems with unknown
actuator failures by introducing a kind of high gain K-filters [10], or decentralized output-
feedback adaptive control scheme for a class of interconnected nonlinear systems with
input quantization. Both logarithmic quantizers and improved hysteretic quantizers were
studied, and a linear time-varying model was introduced to handle the difficulty caused
by quantization [11]. A decentralized adaptive backstepping control scheme was also
proposed for a class of interconnected systems with nonlinear multisource disturbances
and actuator faults. The nonlinear multisource disturbances comprised two parts: one
was the time-varying parameterized uncertainty; the other was the dynamic unexpected
signal formulated by a nonlinear exogenous system [12]. Additionally, the problem of
decentralized adaptive backstepping control for a class of large-scale stochastic nonlinear
time-delay systems with asymmetric saturation actuators and output constraints was also
solved [13]. It can be mentioned that a backstepping-based robust decentralized adaptive
neural H ∞ tracking control method was addressed for a class of large-scale strict feedback
nonlinear systems with uncertain disturbances [14]. Decentralized control was imple-
mented in robotics. The example where a discrete-time decentralized neural identification
and control for large-scale uncertain nonlinear systems at a two degree of freedom planar
robot was implemented can be mentioned [15], or the work that investigated the use of a
decentralized control system for suppressing the vibration of a multi-link flexible robotic
manipulator using embedded smart piezoelectric transducers [16]. Decentralized motion
coordination algorithms were proposed for the robots so that they collectively moved in a
rectangular lattice pattern from any initial position [17]. Mobile robot formations differ in
accordance to the mission, environment and robot abilities. In the case of decentralized
control, the ability to achieve the shapes of these formations has to be built in the con-
trollers of each autonomous robot [18]. A decentralized control algorithm for the robots to
accomplish the sweep coverage was also proposed. The sweep coverage was achieved by
coordinating the robots to move along a given path that was unknown to the vehicles a
priori [19].
During the simulation experiments in the real-time laboratory in recent years, we
reveal the fact that some time is needed to get the appropriate behaviour of control when
the self-tuning controller is used. This time depends on the type of the controlled system.
This is the known problem of self-tuning control because the controller needs the adequate
model of the system. This is one problem that we solve by a new approach described in
Appl.
Appl. Sci. 2021, 11,
Sci. 2021, 11, 6954
x FOR PEER REVIEW 33 of
of 15
14

the fact
this thatby
paper, wethe
used the decentralized
method controllers for theAnother
named as pre-identification. control of multivariable
problem systems.
comes from the
If one
fact set-point
that we used changes its value, it influences
the decentralized controllersall
forother main subsystems
the control by interactions
of multivariable systems.
and
If therefore
one set-pointthe model its
changes of subsystems changes.
value, it influences allThis
other could
mainbesubsystems
also fixed by
byainteractions
self-tuning
and therefore
controller butthe model
some timeofissubsystems
needed to changes.
obtain theThis could
stable be alsoTherefore,
model. fixed by aby
self-tuning
the new
controller but some time is needed to obtain the stable model. Therefore, by
method described in this paper, named as pre-identification, we also solved this problem. the new
method described in this paper, named as pre-identification, we also solved this problem.
2. Materials and Methods
2. Materials and Methods
2.1. Decentralized Control
2.1. Decentralized Control
Using the decentralized approach, the control is divided into a set of sub-tasks that
Using the decentralized approach, the control is divided into a set of sub-tasks that are
are matched by simple controllers. These partial tasks will then give us the overall course
matched by simple controllers. These partial tasks will then give us the overall course of
of control. The main advantages of decentralized control are primarily that a more com-
control. The main advantages of decentralized control are primarily that a more complex
plex system is divided into a set of simple tasks and the resulting controller is more flexi-
system is divided into a set of simple tasks and the resulting controller is more flexible [20].
ble [20].
A special example of multidimensional systems is a system with two inputs and two
A special example of multidimensional systems is a system with two inputs and two
outputs. This can be realized by the so-called P structure, see Figure 1. In this case, the
outputs. This can be realized by the so-called P structure, see Figure 1. In this case, the
inputs to the systems describing the interactions are the values of the action signals of the
inputs to the systems describing the interactions are the values of the action signals of the
SISO controllers and their outputs are added to the opposite outputs of the main diagonal
SISO controllers
systems. and
From this their outputs
figure, aretransfer
we get the added to the opposite
function outputs
equations of model
of the the main
in diagonal
the form
systems. From this figure, we get the transfer function equations of the model in the form
Y GS21 GS12 GR22
GS1𝐺= =1 𝑌= 1 GS11 − 𝐺𝑆21 𝐺𝑆12 𝐺𝑅22
= 𝐺𝑆11 −1 + GS22 GR22 (1)
(1)
𝑆1 U1
𝑈1 1 + 𝐺𝑆22 𝐺𝑅22
Y2 𝑌2 G𝐺S21 𝐺S12
𝑆21G 𝐺R11
𝑆12G 𝑅11
𝑆2 = = =
GS2𝐺= 𝐺𝑆22−−
GS22 (2)
U2𝑈2 11++G𝐺S11 G𝐺R11
𝑆11 𝑅11

Figure 1. Decentralized system control with two inputs and two outputs, the so-called
so-called PP structure.
structure.

2.2. Recursive Identification

2.2. Recursive Identification Using
Using Approximation
Approximation Polynomials
Polynomials
A prerequisite for good control is the most accurate
A prerequisite for good control is the most accurate description
description of regulated
of the the regulated
sys-
system. Identification is the procedure by which the mathematical model of a system is
tem. Identification is the procedure by which the mathematical model of a system is ob-
obtained. The beginnings of identification based on continuous models date back to the
tained. The beginnings of identification based on continuous models date back to the mid-
middle of the 20th century. For continuous-time identification, the identified model is
dle of the 20th century. For continuous-time identification, the identified model is in the
in the form of the differential equations. Differential equations contain expressions with
form of the differential equations. Differential equations contain expressions with deriva-
derivatives over time that are not measurable. It is possible to replace the segment by an
tives over time that are not measurable. It is possible to replace the segment by an approx-
approximation polynomial whose derivatives can be calculated analytically in advance
imation polynomial whose derivatives can be calculated analytically in advance and then
and then calculated numerically, see Figure 2. This approach was for example used by
calculated numerically, see Figure 2. This approach was for example used by Perutka [20].
Perutka [20].
Appl.
Appl.Sci. 2021,
Sci. 2021,11,
11,x 6954
FOR PEER REVIEW 4 of 1415
4 of

Identification
Figure2.2.Identification
Figure scheme
scheme forcontinuous-time
for continuous-timesystems.
systems.

2.3. Least Squares Method with Exponential Forgetting

2.3. Least Squares Method with Exponential Forgetting
The estimation of model parameters is computed as
The estimation of model parameters is computed as
𝜃̂θ̂(𝑘) (k −
= 𝜃θ̂̂ (𝑘
(k) = − 11)) +
+K (k)ê((k𝑘)
𝐾(𝑘)𝑒̂ ) (3)
(3)
The
Thegain
gainvector
vectorisiscalculated
calculatedasas
𝐶(𝑘 − 1)𝜙(𝑘)
𝐾(𝑘) = C ( k − 1) φ ( k ) (4)
K (k) = 1 + 𝜙 𝑇T(𝑘)𝐶(𝑘 − 1)𝜙(𝑘) (4)
1 + φ ( k ) C ( k − 1) φ ( k )
and covariance matrix
and covariance matrix
𝐶(𝑘 − 1)𝜙(𝑘)𝜙 𝑇 (𝑘)𝐶(𝑘 − 1)
𝐶(𝑘) = 𝐶(𝑘 − 1) − C (k − 1𝑇)φ(k)φ T (k)C (k − 1) (5)
C (k) = C (k − 1) − 1 + 𝜙 T(𝑘)𝐶(𝑘 − 1)𝜙(𝑘) (5)
1 + φ ( k ) C ( k − 1) φ ( k )
The following applies to the calculation of the prediction error
The following applies to the calculation of the prediction error
𝑒̂ (𝑘) = 𝑦(𝑘) − 𝜙 𝑇 (𝑘)𝜃̂(𝑘 − 1) (6)
T
In the case of exponential forgetting, ) −criterion
ê(k) = y(kthe −identification
φ (k)θ̂ (kof 1) is (6)
𝑘
In the case of exponential forgetting, the 𝑘−𝑖
criterion of identification is
𝐽 = ∑ (𝜑 𝑒(𝑖))2 (7)
𝑖=𝑘
k0
2
J = ∑ ( ϕk−i e(i )) (7)
where the exponential forgetting factor isi= chosen
k0
in the range of 0 to 1, the most common
near 1.
If the exponential forgetting factor is chosen in the range of 0 to 1, the most common
where
near 1.
𝜙 𝑇 (𝑘)𝐶(𝑘 − 1)𝜙(𝑘) > 0 (8)
If
T
Then φ ( k ) C ( k − 1) φ ( k ) > 0 (8)
Appl. Sci. 2021, 11, x FOR PEER REVIEW 5 of 14

Appl. Sci. 2021, 11, 6954 5 of 15

𝐶(𝑘 − 1)𝜙(𝑘)𝜙 𝑇 (𝑘)𝐶(𝑘 − 1)

C(𝑘) = 𝐶(𝑘 − 1) − (9)
Then 𝜂 −1 + 𝜙 𝑇 (𝑘)𝐶(𝑘 − 1)𝜙(𝑘)
C ( k − 1) φ ( k ) φ T ( k ) C ( k − 1)
where C ( k ) = C ( k − 1 ) − −1 (9)
η + φ T ( k ) C ( k − 1) φ ( k )
1 − 𝜑(𝑘)
where 𝜂(𝑘) = 𝜑(𝑘) − (10)
1−ϕ 𝜉(𝑘)
(k)
η (k) = ϕ(k) − (10)
If ξ (k)
If
𝜙 𝑇 (𝑘)𝐶(𝑘 − 1)𝜙(𝑘) = 0 (11)
φ T ( k ) C ( k − 1) φ ( k ) = 0 (11)
Then
Then
C (𝐶(𝑘)
k) = =
C (𝐶(𝑘
k −−1)1) (12)
(12)
Furthermore
Furthermore

𝜑(𝑘) = {1 + (1 + 𝜌)[ln(1 + 𝜉(𝑘 − 1))]  −1

( v ( k − 1) + 1) η ( k − 1) ξ ( k − 1)
 
ϕ(k) = 1 + (1 + ρ)[ln(1 + ξ (k − 1))] + −1 (13)
(13)
(𝑣(𝑘 −11)++ k − 1) +
ξ (1)𝜂(𝑘 −η1)(k − 1) 𝜉(𝑘 1+
− ξ1)(k − 1−1
)
+[ − 1] }
1 + 𝜉(𝑘 − 1) 2+ 𝜂(𝑘 − 1) 1 + 𝜉(𝑘 − 1)
e (k)
η (k) = 2 (14)
λ(𝑒k)(𝑘)
𝜂(𝑘) = (14)
𝜆(𝑘)
v(k ) = ϕ(k)[v(k − 1) + 1] (15)
𝑣(𝑘)= 𝜑(𝑘)[𝑣(𝑘 − 1) (15)
e2+(k1]

)
λ ( k ) = ϕ ( k ) λ ( k − 1) + (16)
1 + ξ (𝑒k2−(𝑘)1)
𝜆(𝑘) = 𝜑(𝑘) [𝜆(𝑘 − 1) + ] (16)
T 1 + 𝜉(𝑘 − 1)
ξ ( k ) = φ ( k ) C ( k − 1) φ ( k ) (17)
𝜉(𝑘) =is𝜙in𝑇 (𝑘)𝐶(𝑘
The parameters estimation vector the form − 1)𝜙(𝑘) (17)
The parameters estimation  vector is in the form 
θ̂ T (k) = â 0 , â 1 , . . . , â deg ( a ) , b̂0 , b̂1 , . . . , b̂deg ( b ) , d (18)
𝜃̂ 𝑇 (𝑘) = (𝑎̂0 , 𝑎̂1 , … , 𝑎̂deg(𝑎) , 𝑏̂0 , 𝑏̂1 , … , 𝑏̂deg(𝑏) , 𝑑) (18)
and
and regressor
regressor
𝑇 (n−(𝑛−1) (𝑡𝑘 ), 𝑢(𝑡𝑘 ), … , 𝑢((𝑚)
 
tk ), . 𝑘. .),, …
1) m) (𝑡 ), 1) (19)
φ T (k)𝜙=(𝑘)−=y((−𝑦(𝑡 −y, −𝑦
L 𝐿 ( tk ), u ( tk ), . . . , u 𝐿L ( tk ), 1
𝑘 (19)

2.4.
2.4. Self-Tuning
Self-TuningController
Controller
The
The main
main reason
reason for
for using
using adaptive
adaptive control
control isis that
that the
the systems
systems change
change over
over time
time or
or
the
the characteristics
characteristics of
of the
the controlled
controlled system
system areareunknown.
unknown. The The basic
basic principle
principleof
ofadaptive
adaptive
systems
systems is to change
change thethecharacteristics
characteristicsofofthe
thecontroller
controller based
based onon
thethe characteristics
characteristics of
of the
the controlled process [21]. The general scheme of the self-tuning controller
controlled process [21]. The general scheme of the self-tuning controller is shown in Figure is shown
in
3. Figure 3.

Figure 3. The general scheme of the self-tuning controller.

 Appl. Sci. 2021, 11, x FOR PEER REVIEW 6 of 14

Appl. Sci. 2021, 11, 6954 6 of 15

Figure 3. The general scheme of the self-tuning controller.

2.5.
2.5.Suboptimal
SuboptimalLinear
LinearQuadratic
QuadraticTracking
TrackingController
Controller
The
The method was introduced by Dostál[22].
method was introduced by Dostál [22].IfIfthe
thesystem
systemofofFigure
Figure4 4isisconsidered
considered

Figure4.4.System
Figure Systemwith
withfeedback
feedbackcontroller.
controller.

Letus
Let usminimize
minimizeaaquadratic
quadraticfunctional
functionalwith
withtwo
twopenalty
penaltyconstants
constants
Z ∞ n∞
2 2 (𝑡) + 𝜑𝑢 2
o
J = = ∫ µe
𝐽 {𝜇𝑒
(t) + ϕue2 (̃t)(𝑡)}𝑑𝑡
dt (20)
(20)
0 0

TheLaplace
The Laplaceimage
imageofofthe
theset
setpoint
pointholds
holds
ℎ𝑤 (𝑠)
𝑤(𝑠) =
hw (s) (21)
w(s) = 𝑠𝑓𝑤 (𝑠) (21)
s f w (s)
It holds for degrees of polynomials
It holds for degrees of polynomials
deg(ℎ𝑤 ) ≤ deg(𝑓𝑤 ) , 𝑓𝑤 (0) ≠ 0 (22)
deg ( hw ) ≤
We calculate stable polynomials deg(nf was
g and f w (0) 6 =
), results of 0spectral factorizations (22)

(𝑎𝑠) ∗ 𝜑𝑎𝑠g+and
We calculate stable polynomials n as=results
𝑏 ∗ 𝜇𝑏 𝑔 ∗ 𝑔, 𝑛of∗spectral
𝑛 = 𝑎 ∗ 𝑎factorizations (23)
We solve the following
( as) ∗diophantine
ϕas + b ∗ µbequation
= g ∗ g, n ∗ n = a ∗ a (23)
𝑎𝑠𝑝 + 𝑏𝑞 = 𝑔𝑛 (24)
We solve the following diophantine equation
Considering the transfer function of the system
asp + bq = gn (24)
𝑏0
𝐺(𝑠) = 2 (25)
𝑠 + 𝑎1 𝑠 + 𝑎0
Considering the transfer function of the system
then the controller is
b0
G (s) = 2 2 (25)
s + 2a𝑠1 s +
𝑞 + a𝑞01 𝑠 + 𝑞0
𝐹(𝑠)𝑄(𝑠) = 2
(26)
𝑠(𝑝2 𝑠 + 𝑝1 𝑠 + 𝑝0 )
then the controller is
2
In this case, the polynomials have theqform 2 s + q1 s + q0
F (s) Q(s) = (26)
s( p2 s2 + 2p1 s + p0 )
𝑔(𝑠) = 𝑔3 𝑠 3 + 𝑔2 𝑠 + 𝑔1 𝑠 + 𝑔0 (27)
In this case, the polynomials have the form
2 (28)
𝑛(𝑠) = 𝑠 + 𝑛1 𝑠 + 𝑛0
and to calculate their coefficients = g3 s3 + by
g(s) obtained 2
+ g1 s +factorization
g2 sspectral g0 (27)

n(s) = s2 + n1 s +2n0 (28)

𝑔0 = √𝜇𝑏0 (29)
and to calculate their coefficients obtained by spectral factorization
q 𝑔 + 𝜑𝑎2
𝑔1 = √2𝑔 (30)
2 0 0
g0 = µb02 (29)
q 𝑔 + 𝜑(𝑎2 − 2𝑎 )
𝑔2 = √2𝑔 (31)
3 1 1 0
g1 = 2g2 g0 + ϕa20 (30)
q 𝑔3 = √𝜑 (32)
g2 = 2g3 g1 + ϕ( a21 − 2a0 ) (31)
√ 0 (33)
g3 𝑛=0 = ϕ√𝑎2 (32)
q
n0 = a02 (33)
Appl. Sci. 2021, 11, 6954 7 of 15

q
n1 = 2n0 − a21 − 2a0 (34)

2.6. Calculation of Derivatives Using Approximation Functions

To calculate the derivatives, we approximate the closest neighborhood for a given
time by the approximation function. For example, we will use the Lagrange polynomial in
the form
( x − b)( x − c) ( x − a)( x − c) ( x − b)( x − b)
P2 ( x ) = f ( a) + f (b) + f ( ac) (35)
( a − b)( a − c) (b − a)(b − c) (c − a)(c − b)

whose first derivative is

2x −(b+c) 2x −( a+c)
f 0 (x) ∼
0
= P2 ( x ) = f a + (b− a)(b−c)
( a−b)( a−c) ( )
f (b)
−( a+b) (36)
+ (2x
c− a)(c−b)
f (c)

and second derivative is

2 f ( a) 2 f (b) 2 f (c)
f // ( x ) ∼
= P2// ( x ) = + + (37)
( a − b)( a − c) (b − a)(b − c) (c − a)(c − b)

2.7. Recursive Instrumental Variable Method

The instrumental variable method is a modification of the least squares method.
The least squares method uses the quadratic criterion and the existence of one global
minimum. However, a prerequisite for successful least-squares modelling is that the fault is
represented by white noise with zero mean value. The instrumental variable method does
not make it possible to determine the noise properties, but is based on weaker assumptions
than the least squares method. The identification proceeds according to number 5. As with
the least squares method, the method of the instrumental variable can also be formulated
recursively [22–25]. The parameter estimation vector has the form
 
θ̂ T (k) = â0 , â1 , . . . , âdeg(a) , b̂0 , b̂1 , . . . , b̂deg(b) , d (38)

and data vector

( n −1) (m)
θ̂ T (k) = (−y(tk ), . . . − y L ( t k ), u ( t k ), . . . , u L ( t k ), 1 (39)

The gain vector is calculated by relation

C ( k − 1) z ( k )
L(k) = (40)
1 + φ T ( k ) C ( k − 1) z ( k − 1)

In addition to the data vector, it is necessary to know the covariance matrix

C ( k − 1) z ( k ) φ T ( k ) C ( k − 1)
C ( k ) = C ( k − 1) − (41)
1 + φ T ( k ) C ( k − 1) z ( k )

and instrument vector

z ( k ) = ( u ( t k ), u ( t k −1 ), . . . , u ( t k − n − m ), (42)

which we choose as a set of delayed inputs. The prediction error is calculated by

ê(k) = y(k) − φ T (k)θ̂ (k − 1) (43)

and estimating the parameters according to

θ̂ (k) = θ̂ (k − 1) + L(k )ê(k ) (44)

Appl. Sci. 2021, 11, 6954 8 of 15

3. Assumptions
3.1. Decentralized Controllability
Assume the existence of a stable minimum phase Linear Time Invariant (LTI) in time of
a continuous multidimensional system of the dimension N × N. Its Laplace transformation
S(s) : S(t) > S(s), which we call the transformed system is in the form

· · · S1N (s)
 
S11 (s)S12 (s)
S(s) = 
 .. .. .. 
(45)
. . . 
S N1 (s)S N2 (s) · · · S NN (s)

where Sij (s), i = 1, 2, . . . , N, j = 1, 2, . . . is Laplace transformation of the ij-th

subset of Sij (t) of the transformed system S(s). The transformed system S(s) je is decentrally
controllable only when its main diagonal is dominant.

3.2. System Model and Shape of Reference Signal

Suppose there exists a system S(t) and a transformed system S(s) as described above.
Then we formulate a model created for the purpose of decentralized control, which we call
M(t), and its Laplace transformation M(s). Consider M(s) as a stable minimum phase linear
time invariant multivariate diagonal matrix in the form

M(s) 0 ··· 0
 
 0 M(s) ··· 0 
M(s) =  .. .. .. (46)
 
.. 
 . . . . 
0 0 ··· M N (s)

where Mi (s), i = 1, 2, . . . , N is the Laplace transformation of the i-th submodule

Mi (t) of the model M(t)of the transformed system S(s). This assumes minimal impact of
extra-diagonal transmissions, which is important because of the deployment of a decentral-
ized controller. Simplification of the N-dimensional system to N-dimensional systems is
simplified.
Furthermore, suppose that the Laplace transformation of the reference signal vector
r(s) is always in the form

r (s) = ( R1 (s), R2 (s), . . . , R N (s)) (47)

where Ri (s), i = 1, 2..., N is the i-th Laplace reference signal of Laplace transformation of the
reference signal vector r(s) and has the form

hi
Ri ( s ) = (48)
s
where hi ∈ R, i = 1, 2, . . . , N, is the i-th part by constant function, i.e., reference signal,
which is only a combination of p-l step changes of the signal of its different constant values

ji1 f or t ∈ h0, ti1 )





 ji2 f or t ∈ hti1 , ti2 )
hi = .. , 0 < ti1 < ti2 < · · · < tip−1 < tip (49)


 .
jip f or t ∈ htip−1 , tip )


where jil ∈ R, i = 1, 2, . . . , N, l = 1, 2, . . . , p, the l-th constant function i-th in parts

by the constant function hi , t is the time til ∈ R+ , i = 1, 2, . . . , N, l = 1, 2, . . . , p,
the l-th moment of the i-th in portions of the constant function hi . This means that each
non-zero element of the matrix M(s) has exactly one non-zero element of the vector r(s), i.e.,
that each partial transmission of the overall system model has a reference signal defined
for it. As for the form of the reference signal, it is a constant function in parts. This function
Appl. Sci. 2021, 11, 6954 9 of 15

is approximated from an arbitrary but predetermined number of p segments of a different

but concise value, i.e., it varies over time.

4. Pre-Identification
Consider the validity of the assumption of decentralized controllability, system de-
scription and system model. Then, the transformed system S(s) can be viewed as a black
box model, and let it be identified by direct and/or indirect time-continuous algorithms. In
time, continuous model identification can be realized by following steps: The controller
is not connected in the closed circuit. The values of the vector of difference of output
quantities and reference signals E(t) are sent to the input of the system S(t). The values of
the reference signals are the same and at the same time as those that will be used during
regulation.
1. The controller is not connected in the closed circuit. The values of the vector of
difference of output quantities and reference signals E(t) are sent to the input of the
system S(t). The values of the reference signals are the same and at the same time as
those that will be used during regulation.
2. If switching control is considered, each time interval of the control of the system
S(s) at which all reference signals have a constant value is identified separately, in
so-called Identification Elements (IE).
3. Each identification element is identified several times, each time by a different identi-
fication algorithm, and the obtained model can be verified by comparison with the
measured data. The obtained model, which is most consistent with the measured
data, is then used for control. Let us call this method of Identification More Than One
Method (IMTOM).
The system S(s) is completely identified by the above procedure before the actual
regulation begins, therefore identification during the regulation is not necessary. This
procedure is suitable for processes where the same procedure is repeated many times.

5. Results
5.1. Simulation Results
The verified system is described as
!
3.7 0.4
GS (s) = s2 +5.2s+4.6 s2 +4.4s+3.8 (50)
0.6 8.7
s2 +10.6s+10.2 s2 +7.4s+8.2

Since it is the system with two inputs and two outputs, we obtained two suboptimal
linear quadratic controllers in the form that was described in the previous section. These
controllers had the following penalty constants

µ1 = 1, ϕ1 = 30, µ2 = 1, ϕ2 = 30 (51)

We used our algorithm, pre-identification, at this system and we obtained the follow-
ing results, see Figures 5–8. First, we performed system response on the given set-points
depicted in Figure 5. Together with this system response, we obtained the system pre-
identified parameters shown in Figure 6, for the first subsystem in the left one and for
the second subsystem in the right one. According to these pre-identified parameters we
performed the simulation of control which is shown in Figure 7. During the control, the con-
troller parameters were changing, and they are recorded in Figure 8, for the first controller
in left one and and for the second controller in the right one.
ppl. Sci. 2021, 11, x FOR PEER REVIEW

pl. Sci. 2021, 11, x FOR PEER REVIEW

Appl. Sci. 2021, 11, 6954 10 of 15

Appl. Sci. 2021, 11, x FOR PEER REVIEW

Figure 5. System response (green and cyan) on the given set-points (blue and

Figure
Figure
Figure 5. System
5. System
5. System response
response and(green
(green(green
response cyan)
and and
on cyan) cyan)
the given theon the
onset-points
given given
(blue set-points
and red).
set-points (blue and
(blue and red).

Figure
Figure 6. System
6. System pre-identification
pre-identification using recursive using recursive
intrumental intrumental
variable—1st variable—1s
subsystem (left), 2nd
Figure 6. System pre-identification using recursive intrumental variable—1st subsystem (
(right).
(right).
(right).
Figure 6. System pre-identification using recursive intrumental variable—1st
(right).

Figure 7. Output of control.

Figure
Figure 7. Output
7. Output of
of control. control.

Figure 7. Output of control.

Figure 8. Controllers parameters during simulation—1st (left), 2nd (right).

 Appl. Sci. 2021, 11, 6954 11 of 15

Figure 7. Output of control.

Figure
Figure8. Controllers parameters
8. Controllers during simulation—1st
parameters (left), 2nd (right).
during simulation—1st (left), 2nd (right).
Appl. Sci. 2021, 11, x FOR PEER REVIEW
5.2. Results in Real-Time at Laboratory Model
5.2.WeResults
verifiedinthe
Real-Time
presentedatmethod
Laboratory Modelcontrol using MATLAB at CE108 Cou-
in real-time
pled Drives Apparatus
We verified theModel [26] which
presented is shown
method in the Figure
in real-time 9. The using
control laboratory model at CE
MATLAB
CE108 allows solving practical problems of tensioning and speed of material movement in
pled Drives Apparatus Model [26] which is shown in the Figure 9. The laborator
continuous
two wheels processes. An example
are fixed, theiris the speedcan
speed and tension of the thread
betensioning
measured andwhen rewinding The th
regulated.
CE108
from one allows
spool to solving practical
another, which must problems
be controlled. of This situation and speed
is adapted forof material m
labora-
move
tory
(located
in continuous
measurements
on a
processes. movable
so that theAn example
elastic
arm
band isis
suspended
the speed
mounted
on a
and wheels.
on three
spring)
tensionThe and
of the
lower
simulates
thread
two when
along
ing
wheels arewith
from onetension
fixed, spool
their toand
speed speed
another,
can be which
measured measurements.
must
and Two
be controlled.
regulated. The thirdservomotors
This can
wheel move contro
situation is ada
(located on a movable
movementmeasurements
laboratory arm suspended
and tensioning on a
of the
so that spring)
thebelt. and
elasticThe simulates
bandspeed a workstation
is 0–3000
is mounted along
threewith
on rpm, which co
wheels. T
tension and speed measurements. Two servomotors control the speed of movement and
voltage of 0–10 V. Tension measurement is indirect through the angle of the
tensioning of the belt. The speed is 0–3000 rpm, which corresponds to a voltage of 0–10 V.
from measurement
Tension −10 ° to 10 is°,indirect
whichthrough
corresponds
the angle of tothe
a voltage
movable arm, from −10
from V◦to
−10 to 10
10 ◦V.
, Inpu
are located
which corresponds ontothe frontfrom
a voltage panel−10 V oftothe
10 V.device;
Inputs and it outputs
is the are control
located voltage
on the to t
front panel of the device; it is the control voltage to the servomotor amplifiers, which are
amplifiers, which are bidirectional, and which are inputs. There are four ou
bidirectional, and which are inputs. There are four outputs, the voltage corresponding to
age
the corresponding
speed to the speed
of the two servomotors and the of the two
voltage servomotors
corresponding and the
to the tension andvoltage
the co
the tension
speed of the belt.and the speed of the belt.

Figure
Figure 9. Photo
9. Photo of CE108
of CE108 Coupled Coupled Drives
Drives Apparatus Apparatus
model connected model
with PC connected with
with MATLAB. PC wi

Using the pre-identification method and fully implementing interactio

plants, we obtained the following mathematical model to be used at contro
0.78
2
0
𝐺𝑆 (𝑠) = (𝑠 + 2.66𝑠 + 1.33 )
4.16
 Appl. Sci. 2021, 11, 6954 12 of 15

Using the pre-identification method and fully implementing interactions in the main
plants, we obtained the following mathematical model to be used at control of speed
!
0.78
s 2 +2.66s +1.33 0
GS (s) = 4.16 (52)
0 s2 +11.66s+16.66

Since it is the system with two inputs and two outputs, we obtained two suboptimal
linear quadratic controllers in the form that was described in the previous section. These
controllers had the following penalty constants

µ1 = 1, ϕ1 = 1, µ2 = 1, ϕ2 = 0.85 (53)

We used our algorithm, pre-identification, at this system and we obtained the follow-
ing results, see Figures 10–13. First, we performed system response on the given set-points
depicted in Figure 10. Together with this system response, we obtained the system pre-
identified parameters shown in Figure 11, for the first subsystem in the left one and for
the second subsystem in the right one. According to these pre-identified parameters we
ppl. Sci. 2021, 11, x FOR PEER REVIEW performed the simulation of control which is shown in Figure 12. During the control,
the controller parameters were changing, and they are recorded in Figure 13, for the first
controller in left one and and for the second controller in the right one.
Appl. Sci. 2021, 11, x FOR PEER REVIEW

Figure
Figure
Figure 10.
10.10. System
System response
System response
(green(green
response (green
and cyan)
and and
oncyan)
the cyan)
given
on on(blue
set-points
the given the given
and set-points
red).(blue
set-points (blue
and red). an

Figure
Figure 11.11. System
System pre-identification
pre-identification usinginstrumental
using recursive recursivevariable—1st
instrumental variable—1st
subsystem (left), 2nd subsyst
(right).
2nd (right).
Figure 11. System pre-identification using recursive instrumental variable
2nd (right).
Appl. Sci. 2021, 11, 6954 13 of 15

Figure 11. System pre-identification using recursive instrumental variable—1st subsyst

2nd (right).

Figure
Figure 12.12. Output
Output of control.
of control.

Figure
Figure 13.13. Controllers
Controllers parameters
parameters during simulation—1st
during simulation—1st (left),
(left), 2nd (right). 2nd (right).
6. Conclusions
6. Conclusions
This paper presents the new method of identification named as pre-identification
on theThis paperlevel
theoretical presents the new method
and subsequently verified of identification
it by simulations andnamed
in the as pre-identific
real-time
experiments
the theoretical level and subsequently verified it by simulations and inthethe real-
at Coupled Drives model in the laboratory. The results confirm that
method can be successfully used and future work will focus on the verification of this
periments at Coupled Drives model in the laboratory. The results confirm that the
method on more examples both in simulation and in laboratory conditions. This new
can beenhances
method successfully used
the usage of aand future work
decentralized will focus
self-tuning on in
controller the
theverification of this me
way that it fixes
the time the adaptive controller needs to adapt its model.

Author Contributions: Conceptualization, K.P. and I.K.; methodology, K.P. and I.K.; software, I.Z.;
validation, V.B., K.P. and V.T.; formal analysis, M.B.; investigation, K.P. and V.T.; resources, K.P.; data
curation, K.P. and M.B.; writing—original draft preparation, K.P.; writing—review and editing, I.Z.;
visualization, K.P. and I.Z.; supervision, K.P. and M.S.; project administration, K.P. and I.K.; funding
acquisition, M.S. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by by the European Regional Development Fund under the
project CEBIA-Tech No. CZ.1.05/2.1.00/03.0089. This research was funded by the Ministry of
Education, Science, research and Sport of the Slovak Republic under the project STIMULY MATADOR
1247/2018. Project title: Research and development of modular reconfigurable production systems
using Smart Industry principles for automotive with pilot application in MoBearing Line industry.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Appl. Sci. 2021, 11, 6954 14 of 15

Acknowledgments: This work was supported by the European Regional Development Fund under
the project CEBIA-Tech No. CZ.1.05/2.1.00/03.0089. This article was made under support of project:
STIMULY MATADOR 1247/2018. Project title: Research and development of modular reconfig-
urable production systems using Smart Industry principles for automotive with pilot application in
MoBearing Line industry.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Pournas, E.; Jung, S.; Yadhunathan, S.; Zhang, H.; Fang, X. Socio-technical smart grid optimization via decentralized charge
control of electric vehicles. Appl. Soft Comput. J. 2019, 82, 10573.
2. Yuan, Q.; Zhan, J.; Li, X. Outdoor flocking of quadcopter drones with decentralized model predictive control. ISA Trans. 2017, 71,
84–92. [CrossRef] [PubMed]
3. Blanchini, F.; Casagrande, D.; Fabiani, F.; Giordano, G.; Pesenti, R. Network-decentralized optimization and control: An explicit
saturated solution. Automatica 2019, 103, 379–389. [CrossRef]
4. Wang, Y.; Hu, J.; Zheng, Y. Improved decentralized prescribed performance control for non-affine large-scale systems with
uncertain actuator nonlinearity. J. Frankl. Inst. 2019, 356, 7091–7111. [CrossRef]
5. Qu, Q.; Zhang, H.; Feng, T.; Jiang, H. Decentralized adaptive tracking control scheme for nonlinear large-scale interconnected
systems via adaptive dynamic programming. Neurocomputing 2017, 225, 1–10. [CrossRef]
6. Bahramipanah, M.; Cherkaoui, R.; Paolone, M. Decentralized voltage control of clustered active distribution network by means of
energy storage systems. Electr. Power Syst. Res. 2016, 136, 370–382. [CrossRef]
7. Liu, D.; Yang, G.H. Decentralized event-triggered output feedback control for a class of interconnected large-scale systems. ISA
Trans. 2019, 93, 156–164. [CrossRef] [PubMed]
8. Shi, W.; Yan, F.; Li, B. Adaptive fuzzy decentralized control for a class of nonlinear systems with different performance constraints.
Fuzzy Sets Syst. 2019, 374, 1–22. [CrossRef]
9. Choi, Y.H.; Yoo, S.J. Decentralized adaptive output-feedback control of interconnected nonlinear time-delay systems using
minimal neural networks. J. Frankl. Inst. 2018, 355, 81–105. [CrossRef]
10. Wang, C.; Wen, C.; Guo, L. Decentralized output-feedback adaptive control for a class of interconnected nonlinear systems with
unknown actuator failures. Automatica 2016, 71, 187–196. [CrossRef]
11. Wang, C.; Wen, C.; Lin, Y.; Wang, W. Decentralized adaptive tracking control for a class of interconnected nonlinear systems with
input quantization. Automatica 2017, 81, 359–368. [CrossRef]
12. Wang, Z.; Zhang, B.; Yuan, J. Decentralized adaptive fault tolerant control for a class of interconnected systems with nonlinear
multisource disturbances. J. Frankl. Inst. 2018, 355, 4493–4514. [CrossRef]
13. Si, W.; Dong, X.; Yang, F. Decentralized adaptive neural control for interconnected stochastic nonlinear delay-time systems with
asymmetric saturation actuators and output constraints. J. Frankl. Inst. 2018, 355, 54–80. [CrossRef]
14. Li, X.; Liu, X. Backstepping-based decentralized adaptive neural H∞ tracking control for a class of large-scale nonlinear
interconnected systems. J. Frankl. Inst. 2018, 355, 4533–4552. [CrossRef]
15. Tellez, F.O.; Loukianov, A.G.; Sanchez, E.N.; Corrochano, J.B. Decentralized neural identification and control or uncertain
nonlinear systems: Application to planar robot. J. Frankl. Inst. 2010, 347, 1015–1034. [CrossRef]
16. Halim, D.; Luo, X.; Trivailo, P.M. Decentralized vibration control of a multi-link flexible robotic manipulator using smart
piezoelectric transducers. Acta Astronaut. 2014, 104, 186–196. [CrossRef]
17. Cheng, T.M.; Savkin, A.V. Decentralized Control of Multi-robot Systems for Rectangular Aggregation. IFAC Proc. Vol. 2011, 44,
11574–11579. [CrossRef]
18. Necsulescu, D.; Pruner, E.; Kim, B.; Sasiadek, J. Decentralized Control of Autonomous Mobile Robots Formations using Velocity
Potentials. IFAC Proc. Vol. 2012, 45, 31–36. [CrossRef]
19. Cheng, T.M.; Savkin, A.; Javed, F. Decentralized Control of a group of mobile robots for deployment in sweep coverage. Robot.
Auton. Syst. 2011, 59, 497–507. [CrossRef]
20. Perutka, K. Adaptive LQ control with pre-identification of two tanks laboratory model. In Annals of DAAAM and Proceedings of
the International DAAAM Symposium; Danube Adria Association for Automation and Manufacturing: Wienna, Austria, 2009; pp.
439–440.
21. Bobal, V.; Böhm, J.; Fessl, J.; Machacek, J. Digital Self-Tuning Controllers; Springer: London, UK, 2009.
22. Dostal, P.; Bobal, V. The suboptimal tracking problem in linear systems. In Proceedings of the 7th Conference on Control and
Automation, Haifa, Israel, 28–30 June 1999; pp. 667–673.
23. Perutka, K. Decentralized Adaptive Control. Ph.D. Thesis, Tomas Bata University iz Zlin, Zlín, Czech Republic, 2007.
24. Velíšek, K.; Holubek, R.; Sobrino, D.R.; Ružarovský, R.; Vetríková, N. Design of a robotized workstation making use of the
integration of CAD models and Robotic Simulation software as way of pairing and comparing real and virtual environments. In
MATEC, 4th International Conference on Computing and Solutions in Manufacturing Engineering 2016—CoSME´16; EDP Sciences:
Youlis, France, 2017.
Appl. Sci. 2021, 11, 6954 15 of 15

25. Holubek, R.; Ružarovský, R.; Delgado Sobrino, D.R.; Košt’ál, P.; Švorc, A.; Velíšek, K. Novel trends in the assembly process as the
results of human—The industrial robot collaboration. In MATEC Web of Conferences; EDP Sciences: Youlis, France, 2017; Volume
137.
26. TQ Tecquioment Academia-CE108 Coupled Drives Apparatus. November 2018. Available online: https://www.tecquipment.
com/assets/documents/datasheets/CE108-Coupled-Drives-Datasheet.pdf (accessed on 28 July 2021).