You are on page 1of 27

Applied Intelligence manuscript No.

(will be inserted by the editor)

Designing an interval type-2 fuzzy disturbance


observer for a class of nonlinear systems based on
modified particle swarm optimization

Shokoufeh Naderi · Behrooz Rezaie⋆ ·


Mostafa Faramin

Received: date / Accepted: date

Abstract This paper presents a new interval type-2 fuzzy disturbance observer
design for a class of nonlinear systems using modified particle swarm optimiza-
tion. The design procedure has two main parts, including the selection of the
initial structure of the type-2 fuzzy disturbance observer and the optimization of
the observer parameters using a modified particle swarm optimization algorithm.
The modified particle swarm optimization algorithm has a better performance in
terms of the accuracy and convergence rate compared with the standard particle
swarm optimization and many other evolutionary algorithms. In this algorithm,
the upper and lower bound of the search space is defined for the parameters of each
particle based on their value, and weaker particles are substituted with new par-
ticles, resulting in searching in a reasonable space. To accentuate the outstanding
performance of the modified particle swarm optimization for the considered task,
its performance has been compared with five famous meta-heuristic optimization
algorithms. In addition, utilizing interval type-2 fuzzy systems in the proposed
observer has provided more robustness compared with type-1 fuzzy systems. The
effectiveness of the proposed fuzzy disturbance observer has been shown through
computer simulation and experimental results for the ball and beam system, while
the system was subjected to sinusoid and square disturbances, and a compari-
son has been drawn to indicate the superiority of the proposed fuzzy disturbance
observer over the other observers.

⋆ Corresponding author

Shokoufeh Naderi
E-mail: shuku naderi@yahoo.com
Behrooz Rezaie
Faculty of Electrical and Computer Engineering, Babol Noshirvani University of Technology,
Babol, Iran
Tel.: +123-45-678910
Fax: +123-45-678910
E-mail: brezaie@nit.ac.ir
Mostafa Faramin
E-mail: mostafa.faramin@stu.nit.ac.ir
2 Shokoufeh Naderi et al.

Keywords Fuzzy disturbance observer · Interval type-2 fuzzy system · Modified


particle swarm optimization · Ball and beam system

1 Introduction

Unknown disturbances, unmodeled dynamics, parameter uncertainties, and


nonlinearities are unfavorable factors that may exist in almost all practical sys-
tems. These effects, along with the existence of some noises in sensor measurement,
the fault in actuators, and other structural vibrations, can severely degrade the
performance of a control system and may even lead to instability in some cases.
Therefore, disturbance attenuation in dynamical systems is one of the challenging
subjects in the modern control theory and has been one of the most active research
areas in the last few decades (Yao and Guo 2013; Liu et al. 2012).
In dealing with systems affected by disturbance inputs, robust methods are
usually applied to reduce the effects of disturbance. To reject these influences,
usually, it is required to have some information about the disturbance. Therefore,
estimating the disturbance behavior and then taking control action has been sug-
gested as an attractive idea to improve the robustness of the system. To address
this, some disturbance observer-based control methods such as robust integral
feedback control (Faramin et al. 2019), adaptive output feedback control (Chen
and Ge 2015), and sliding mode control (Huang et al. 2019) have been proposed
for nonlinear systems.
Among the presented methods, fuzzy-based approaches have attracted much
attention, due to being model-free and related to the knowledge of the experts
and designers. The fuzzy logic, known as type-1 (T1) fuzzy logic, has been in-
troduced firstly by Zadeh (1965), and has a broad range of applications, such as
fuzzy group decision-making (Capuano et al. 2019; Meng et al. 2019; Lang et al.
2019), mathematics (Rudas et al. 2009), and engineering (Shaocheng et al. 2000;
Tseng et al. 2001). Moreover, fuzzy disturbance observers (FDO), which estimate
disturbances and uncertainties and can cope with parameter variations in the con-
trol process, have been investigated in many studies. For instance, in Kim (2002),
an FDO-based control has been improved as an alternative fuzzy control method.
Afterward, a discrete-time FDO has been proposed for nonlinear sampled systems
(Kim and Park 2004), and then an improved performance has been achieved in
Yoo et al. (2010) by updating the parameter vector. However, the fixed member-
ship function (MF) design of T1 fuzzy systems (T1FSs) comes with some inherent
difficulties of rule-based control design when the system has uncertainties.
Employing the extended version of T1FS, which is known as the type-2 fuzzy
system (T2FS), can overcome the limitations of T1FS. It can also improve the per-
formance of the control process, especially when dealing with uncertainties such as
unknown data and changing environments. Compared with T1FS, T2FS presents
more robustness and provides satisfactory results due to using the footprint of
uncertainty (FOU) of the MFs. In fact, in the type-2 (T2) fuzzy set, the member-
ship value is a T1 fuzzy number, which belongs to the [0, 1] interval. T2FSs have
been successfully employed in many applications (Castillo et al. 2016; Castillo and
Amador-Angulo 2018).
The relatively more simplified version of the T2FS is the interval T2 (IT2) fuzzy
set, where secondary membership values are set to be either 0 or 1. Since interval
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 3

T2FSs (IT2FSs) have verified their ability in real-world problems, most of the
T2FS research studies have only focused on this category of T2FSs. Some examples
of this research can be found in vectorization-optimization for classification of noisy
data (Wu and Huang 2013), deriving the analytical structure of a broad class of
IT2 Mamdani fuzzy controllers (Zhou and Ying 2013), sliding controller for wing
rock systems (Tao et al. 2012), and fuzzy neural networks (Lin et al. 2014).
Fuzzy logic systems are based on rules and linguistic terms whose design relies
on experts’ experiences. Thus, obtaining the parameters of the fuzzy system is a
complex task, especially for complex systems. Determining the correct values for
these parameters is vitally important because it can intensely affect the operation
of the system. These parameters can be tuned by trial and error, which is not
an easy task and also consumes a considerable amount of time. Hence, numer-
ous design methods have been proposed to optimize those elements and construct
IT2FSs, such as Kalman filter algorithms (Juang and Tsao 2008) and SVD-QR
method (Liang and Mendel 2000a), each having different properties and charac-
teristics. Some of these design methods emphasize on computational simplicity,
some involve no tuning of the MF parameters, and others involve lots of tuning
optimization of those parameters (Mendel 2017).
On the other hand, many researchers have proposed meta-heuristic and evo-
lutionary algorithms for optimizing fuzzy systems such as genetic algorithm (Oh
et al. 2011), ant colony optimization (Castillo et al. 2015), and particle swarm
optimization (PSO) (Maldonado et al. 2013). The advantage of evolutionary al-
gorithms over the derivative-based methods is the high rate of convergence, high
optimization accuracy, not getting trapped at a local extremum (Shi and Eber-
hart 1999), and also avoiding complex derivative formulas. These superiorities
have led to the development of many evolutionary algorithms in recent decades
(Storn and Price 1997; Leung et al. 2003; Eusuff and Lansey 2003; Karaboga 2005;
Atashpaz-Gargari and Lucas 2007; Karaboga and Basturk 2008; Cuevas et al.
2015; Salahshour et al. 2018). All of these algorithms contain similar steps. Locust
swarms algorithm (LSA), for instance, starts with an initial population. In this
algorithm, individuals mimic a group of locusts, which interact with each other
based on the biological laws of the cooperative swarm. The algorithm terminates
after reaching a certain number of iterations. (Karaboga 2005). Honey bee swarm
(HBS) is also a population-based algorithm, which can be used for solving uni-
modal and multi-modal numerical optimization problems. (Cuevas et al. 2015).
One of the important evolutionary algorithms is PSO, which is inspired by the
social behavior of bird flocks or fish schools (Dorigo et al. 1996).
The PSO was first introduced by Kennedy and Eberhart (1995). After that,
many researchers have improved the performance of the algorithm by investigating
some criteria such as the rate of convergence, efficiency, speed, and accuracy (Duan
and Gao 2011; Zhang et al. 2015). However, for large scale problems, in which the
search space of the parameters is not clear, these criteria are not usually satis-
fied. Modifications have been made in (Salahshour et al. 2018) to consider these
problems and adjust the convergence rate and accuracy of the algorithm. Simu-
lation results in (Salahshour et al. 2018) have proven the priority of the modified
algorithm in terms of speed, accuracy, and elapsed time over several evolutionary
algorithms, including artificial bee colony and improved genetic algorithm.
The objective of the present research is to propose a new IT2 FDO based on
the modified PSO (MPSO). Thanks to the advantages of the MPSO algorithm
4 Shokoufeh Naderi et al.

proposed in (Salahshour et al. 2018), this algorithm is considered to be suitable,


so it is employed in the design process. First, based on the expert knowledge or
performed experiments, the overall structure and inputs of the observer are deter-
mined. Then, the parameters of the MFs are optimized using the MPSO algorithm
with a specified initialization procedure. The optimized IT2 FDO is then applied to
the ball and beam system through computer simulation and experimental results,
while the system is subjected to sinusoid and squared disturbances. In summary,
this paper mainly focuses on:
• Introducing an IT2 FDO design method.
• Optimizing the FDO parameters with the aid of the MPSO algorithm.
• Introducing a certain initialization procedure for optimizing the IT2 FDO.
• Applying the proposed FDO to a practical ball and beam setup.
Unlike the previous studies, in this method, there is no limitation on the dis-
turbance function, and independence from the system parameters has enabled the
proposed FDO to handle any parameter uncertainties. Additionally, the proposed
observer can be applied to systems with any controller without changing it.
The rest of the paper has been organized as follows: Sect. 2 describes the
basic concepts of IT2FS and MPSO as well as the problem statement. Sect. 3
presents the proposed model. Sect. 4 shows the effectiveness of the proposed design
method by performing simulations and experimental results on a ball and beam
system. A comparison between the performance of the MPSO and some other
evolutionary algorithms is provided in Sect. 5, and finally, Sect. 6 offers some
conclusion remarks.

2 Backgrounds and problem formulation

The purpose of this section is to briefly review some preliminaries to IT2FSs


and the MPSO algorithm. The problem formulation is also given in this section.

2.1 IT2FS

The principles of all types of fuzzy systems are similar. A higher type of fuzzy
system only indicates a higher degree of fuzziness. Since a higher type alters the
nature of MFs, the MF-dependent operations change. However, the basic principles
of a fuzzy system, which are independent of MFs, do not change (Mendel 2017).
Apart from the nature of MFs, the main difference between T1 and T2 fuzzy
systems, is related to the output processing, wherein an IT2FS, the type of fuzzy
output sets is first reduced to a T1 fuzzy set and then is followed by a defuzzifier
block to obtain a crisp output. Fig. 1 shows the structure of an IT2FS consisting of
three stages of fuzzification, fuzzy inference engine, and output processing (Mendel
2004).
An IT2 MF can be defined based on two T1 MFs named the upper MF (UMF)
and lower MF (LMF). These MFs and the area between them are known as the
footprint of uncertainty (FOU) and are shown in Fig. 2 (Ontiveros et al. 2018).
This figure shows the IT2 membership grade of x denoted by µX̃ (x) which is an
interval set denoted by µX̃ (x) = [µX̃ (x), µX̃ (x)], where µX̃ (x) and µX̃ (x) are the
LMF and UMF respectively.
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 5

Fig. 1 IT2 fuzzy logic system.

2.1.1 Fuzzification

In the fuzzification stage, crisp input values are mapped into IT2 fuzzy sets.
With the given numeric inputs x = (x1 , . . . , xp )T , singleton fuzzifier maps a spe-
cific value of xj (j = 1, . . . , p), namely x′j , as follows:

(
1/1 xj = x′j
µX̃j (xj ) = (1)
1/0 xj ̸= x′j ,

where µX̃j (xj ) denotes the membership grade of xj in the IT2 fuzzy set X̃j .
There are two other kinds of fuzzifier for an IT2FS named the T1 non-singleton
and IT2 non-singleton, whose definition can be found in (Mendel 2017).

Fig. 2 IT2 MF.


6 Shokoufeh Naderi et al.

2.1.2 Fuzzy inference engine

The structure of IT2FS rules is similar to the T1 case, but some or all of the
involved sets are IT2.

Ri : IF x1 is F̃1i and . . . and xp is F̃pi , THEN y is G̃i , i = 1, . . . , M, (2)

where M is the number of rules, xp is the input p, F̃pi , is an MF on rule y and


input p, and y is the output on MF G̃ip . F̃pi and G̃ip are both in the form of µF̃
and µG̃ , respectively.
Fuzzy inference engine works in the same way as for T1FS, except that the
antecedents’ fuzzy sets and the consequent are IT2 fuzzy sets. The process consists
of combining the rules and mapping input IT2 fuzzy sets into output IT2 fuzzy
sets. For singleton fuzzification, the MF for a fired-rule output set, B̃ i , is as follow
(Mendel 2017):
 h i
µB̃ i (y) = 1/ f i (x′ ), f i (x′ ) ⋆µG̃i (y), (3)

where µB̃ i is the resulting MF in the consequent ith rule’s inference, ⋆ plays the
role of t-norm, and points f i and f i define an interval T1 (IT1) set named firing
interval, that makes use of both the LMF and UMF of the antecedent FOUs at
x = x′ and is expressed as follow:
h i
F i (x′ ) = f i (x′ ), f i (x′ ) , (4)

where
p
f i (x′ ) = µF̃ i (x′j ),
Y
(5)
j
j=1
p
i
f (x′ ) = µF̃ i (x′j ) .
Y
(6)
j
j=1

plays the role of t-norm. µF̃ i (x′j ) and µF̃ i (x′j ) denote the LMF
Q
In (5) and (6),
j j
and UMF of the IT2 MF, which have the minimum and maximum membership
grade of the FOU respectively (Fig. 2).

2.1.3 Type-reduction and defuzzification

One of the major steps in the rule-based fuzzy system is the movement from
a fuzzy set to a certain numeric value. A T2 rule-based fuzzy system first maps a
T2FS into a T1FS (type-reduction) and then maps that set into a certain numeric
value (defuzzification).
Type-reduction can be implemented in many ways. Ycos (x) in (7) is the center-
of-sets type-reduction, which is an interval set defined by two points yl and yr
(Liang and Mendel 2000b):
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 7

M
X
Z Z Z Z , f i yi
i=1
Ycos (x) = [yl , yr ] = ··· ··· 1 M
,1
1 M
y 1 ∈[y 1 y 1 ] y M ∈[y M y M ] f 1 ∈[f 1 f ] f M ∈[f M f ] X
fi
i=1
(7)
where yl and yr are the maximum and minimum values of y respectively. There
is no direct theoretical solution for (7), and calculation of yl and yr requires an
iterative procedure. Many algorithms have been proposed for computing these
points such as Karnik-Mendel (KM) (Karnik and Mendel 2001), enhanced KM
(EKM) (Wu and Mendel 2009), iterative algorithm with stopping condition (IASC)
(Melgarejo 2007; Duran et al. 2008), and enhanced IASC (EIASC) (Wu and Nie
2011).
Finally, once yl and yr have been specified, they can be used for computing
the numeric value of output based on the centroid defuzzification in the following
form:
yl + yr
y= . (8)
2

2.2 MPSO algorithm

PSO is a population-based stochastic optimization technique inspired by the


social behavior of bird flocking or fish schooling, and it was first introduced in
(Kennedy and Eberhart 1995). The algorithm begins with a swarm that is initial-
ized at random positions and velocities. The learning process depends on both the
experience of each particle itself and the experience of the most successful particle.
Particles use shared information to move towards their best neighbors, while they
constantly check their movement. The fitness value of each particle is calculated
by a cost function, which has an inverse relationship with the distance between
that particle and the target. Each particle’s position is set according to the best
solution that it has obtained so far, and the best overall solution that any particle
in the population has achieved until that moment. These parameters are denoted
by pbest (particle best) and gbest (global best), respectively. In every iteration, the
best solution of each particle is stored, and all particles modify their trajectories
based on these best solutions. Therefore, after each iteration, necessary modifica-
tions are made to the position and velocity of particles, and as a result, particles
gradually move towards the gbest until reaching the optimal solution. These mod-
ifications are performed depending on the following equations (Shi and Eberhart
1998):

vi (t + 1) = wvi (t) + c1 r1 (pbest,i (t) − xi (t)) + c2 r2 (gbest,i (t) − xi (t)), (9)


1 A T1 fuzzy set A is expressed as follow:
R
A = x∈X µA (X)/x
R
where X denotes a universe, µA (x) denotes a membership value, and denotes union over all
admissible x ∈ X with the associated MF µA (x) (Juang and Hsu 2009).
8 Shokoufeh Naderi et al.

xi (t + 1) = xi (t) + vi (t + 1), (10)

In (9), r1 and r2 are uniform random numbers distributed within the [0, 1] interval.
w is called the inertia weight and has great importance on the speed and the
convergence rate of the algorithm. The local and global searches are balanced by
this parameter. Choosing a large w value helps the global search while selecting
small values for this parameter facilitates the local search. In order to have the most
effectiveness, this value should be large at the beginning to perform an extensive
search over the whole search space, and it should then gradually decrease to find
the exact solution. c1 and c2 are positive acceleration constants, which are often
set to the same value in order to keep the balance between the global and local
search. These learning factors represent the relative influence of the stochastic
acceleration terms, which pulls particles toward pbest and gbest .
As can be seen in (9), the velocity of particles is affected by three factors.
The first one is the effect of each particle’s previous velocity and is controlled by
inertia weight. The second factor relies on the best experience of the particle. This
part makes the particle keep its information, so parameter c1 is referred to as the
cognitive acceleration factor. The third factor depends on the best experience of
the swarm, which means the particle’s movement is a result of other particles’
experience in the swarm. Therefore, parameter c2 is called the social acceleration
factor.
Implementation of the PSO algorithm is easy, and there are few parameters to
adjust. However, it has some disadvantages. If there are many local extremums, the
algorithm is more susceptible to get trapped in the local extremums. Furthermore,
in some problems, it is not possible to specify the search space precisely.
In (Salahshour et al. 2018), some modifications have been made to the standard
PSO to overcome its shortcomings. In this modified algorithm, the search space for
each particle is adjusted based on that particle, and new particles are substituted
for the weaker ones. As a result, the particles examine new trajectories to find the
optimal solution, and falling into local extremums will be avoided. A flowchart for
the MPSO algorithm is shown in Fig. 3. In the flowchart, the saturation value is
an arbitrary percentage of the maximum or minimum of the search space. ep and
et are determined based on the iteration and percentage of the initial swarm size,
respectively.

2.3 Problem formulation

Consider a class of uncertain nonlinear systems in the following form:



ẋi = xi+1 i = 1, 2, . . . n − 1

ẋn = f (x, u) + ∆f (x) + d (11)

y = h(x),

where x = [x1 , . . . , xn ]T ∈ Rn is the state vector. u ∈ R and y ∈ R are the


control input and the system output, respectively. f (x, u) : Rn × R → R and
h(x) : Rn → R are nonlinear functions, d : R+ → R is the external disturbance,
and ∆f (x) : Rn → R represents the system parameters uncertainty and modeling
error.
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 9

Fig. 3 Flowchart of the MPSO algorithm.

Let
D = ∆f (x) + d, (12)

then (11) becomes:



ẋi = xi+1 i = 1, 2, . . . , n − 1

ẋn = f (x, u) + D (13)

y = h(x),

where D denotes the total disturbance, including internal uncertainties from un-
modelled dynamics or variations of the system parameters and external distur-
bance.
10 Shokoufeh Naderi et al.

The objective of this paper is to design an FDO to monitor the total distur-
bance and drive the system output to track the desired reference in the presence
of internal parameter uncertainty and external disturbance.

3 FDO‌ design using MPSO

In general, it is very difficult to represent the total disturbance in (13) using


mathematical equations. Therefore, in the current study, the IT2 FDO is employed
to approximate the total disturbance.
In order to define the initial structure of the FDO, the control input and the
output or measurable states of the system can be chosen as the inputs of the
observer. The choice of input variables depends on its application. In some cases,
it may be challenging, and it is best to search for the advice of experts or to create
more fuzzy systems to realize which choices lead to a better design. The selected
state should be affected by the disturbance more than others and should also be
able to approximately reveal the disturbance changes. For most systems, choosing
the control input and a state of the system achieves satisfactory design accuracy.
This choice leads to the simplest fuzzy system structure with two inputs.
After specifying the structure of the FDO, the parameters of the MFs are opti-
mized in an offline manner by means of MPSO. Fig. 4 shows the block diagram of
the observer parameter setting using MPSO. D represents the disturbance, which
deteriorates the tracking performance of the system. The role of the disturbance
observer is to estimate this disturbance as closely as possible. The estimated dis-
turbance D̂ is then subtracted from the control signal to eliminate or minimize
the effect of the disturbance.
In the MPSO algorithm, the initialization of the population is a very important
task and can prevent falling into local extremums. Inspired by Mendel (2017), the
following three-step procedure is used to initialize the particles of the algorithm
in tuning the parameters of the IT2 FDO:
1. Design a non-singleton T1 FDO by optimizing its parameters using MPSO in
which the parameters of all particles are randomly initialized.

Fig. 4 Block diagram of setting the observer parameters using MPSO.


Designing an IT2 FDO for a class of nonlinear systems based on MPSO 11

2. Design a singleton IT2 FDO by optimizing its parameters using MPSO, in


which one particle is associated with the just designed non-singleton T1 FDO,
and where the parameters of all the remaining particles are randomly initial-
ized. If the system performance is acceptable, stop the algorithm, otherwise,
go to the next step.
3. Design a non-singleton IT2 FDO by optimizing its parameters using MPSO, in
which one particle is associated with the just designed singleton IT2 FDO, and
where the parameters of all the remaining particles are randomly initialized.
The mean-square error (MSE) is selected as the cost function to utilize MPSO
for finding optimal observer parameters. As mentioned, FDO parameters are op-
timized in an offline manner. The known disturbance, d, is applied to the system
during the optimization. Thus, during the optimization, the MPSO is engaged to
minimize the MSE. It is calculated as the difference between the real values of
disturbance and the values estimated by the FDO and is given by:

1 Xn 2
M SE = (di − dˆi ) , (14)
n i=1

where dˆi denotes the estimation value of the ith instance data. The applied dis-
turbance, d, at this stage, should contain different values of possible disturbances,
so the optimized FDO can accurately estimate any disturbance type and value.
The advantage of the proposed FDO is the simplicity of applying it to sys-
tems with PID controllers, which are the most commonly used controllers in the
industry because they are simple and adaptable in practical applications (Karimi
et al. 2018). Furthermore, in this method, uncertainties do not affect the FDO’s
performance because it is independent of the system parameters. Moreover, in this
method, there is no need for information about the uncertainties or disturbances.
It should be noticed that the structure and inputs of the IT2 FDO vary in
different applications. In the next section, a more detailed explanation of the design
procedure, including the FDO structure and the MPSO setting will be presented.

4 Results and discussions

To show the effectiveness of the proposed scheme in the previous section, the
simulation and experimental results for a well-known ball and beam system are
given in this section.

4.1 System and control structure

Fig. 5 shows the schematic of the ball and beam system consisting of a beam
on which the metal ball is free to roll. The beam is fitted with a linear transducer
to measure the position of the ball, which means it outputs a voltage signal pro-
portional to the position of the ball. One side of the beam is attached to a lever
arm that can be coupled to the load gear of the servo-motor unit. By controlling
the position of the servo, the beam angle can be adjusted to balance the ball to the
desired position. The dynamical model of the ball and beam system is as follows:
12 Shokoufeh Naderi et al.

Fig. 5 Schematic of the ball and beam system.


ẋ1 = x2


m grarm rb2 sin(α(t))
ẋ2 = Lb beam (mb rb2 +Jb )
+ ∆f (x) + d (15)


y = x .
1

Considering o in the middle of the beam, as the reference origin, the state vector
of the ball and beam system is defined as x = [x1 x2 ], where x1 = r is the
ball position with regards to o, and x2 is the velocity of the ball. y and d are
the output of the system and the external disturbance, respectively, and ∆f (x)
represents the uncertainty of the parameters. The definition of other parameters
and their simulation values are given in Table 1. The variation of α is in the direct
relationship with the variation of the servo-motor angle θ:

θrarm = αLbeam . (16)

Table 1 Simulation parameters.

Symbol Description ‌ Value

mb mass of the ball 0.064 kg


rb radius of the ball 0.0127 m
rarm lever arm offset 0.0254 m
Lbeam length of the beam 0.4255 m
Jb ball inertia 4.129 × 10−6 kg m2
g acceleration of gravity 9.8 m/s2
τ time constant of the motor 0.0285 s
K DC motor gain 1.76 rad/s V
V motor input voltage
θ servo load gear angle
α beam pitch
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 13

Fig. 6 Block diagram of the control scheme for the ball and beam system.

The servo-motor dynamic is modeled as a first-order system with an integrator,


and is represented by the following transfer function (Oh et al. 2011):

θ(s) K
= . (17)
V (s) s(τ s + 1)

Fig. 6 depicts the block diagram of the cascade control and the disturbance
observer that is used for rejecting the disturbance. Based on the measured ball
position, r, the ball and beam controller in the outer loop computes the servo load
angle, θd ,, which is required for attaining the desired ball position, rd . The inner
loop is a servo position control system. Thus, the servo controller calculates the
required motor voltage for the load angle to track a given desired load angle. The
controllers of the inner and outer loops are both proportional-derivative (PD).
The proportional and derivative gain values of PD controllers are chosen as
13.71 and 8 × 10−5 for the inner loop controller, and 6.58 and 4.69 for the outer
loop controller, respectively.

Fig. 7 Type of MFs.


14 Shokoufeh Naderi et al.

Table 2 Upper and lower MFs parameters of the optimized IT2 FDO.

Input1 Input2

UMF LMF UMF LMF

c11 -0.8837 -0.3686 -0.2682 -0.3008


c12 -0.5017 -0.1775 -0.0161 0.1199
c21 -1.1908 0.2290 -0.0136 0.0572
c22 -0.5437 0.0904 0.1428 0.0834
c31 0.0605 0.0030 0.1895 0.1531
c32 0.2407 0.0197 0.2977 0.2026
c41 0.4911 0.4095 0.3657 0.2879
c42 0.9614 0.9438 0.5005 0.4973
σ11 0.2299 0.0010 0.0004 0.0004
σ12 0.7338 0.1244 0.0222 0.0004
σ21 0.7426 0.0576 0.1943 0.3014
σ22 0.4207 0.4384 0.0004 0.0306
σ31 0.0010 0.0024 0.1569 0.1656
σ32 0.0010 0.0221 0.1340 0.0943
σ41 0.3668 0.0010 0.0689 0.0030
σ42 0.4588 0.1072 0.0004 0.0004

4.2 FDO structure

The first step in designing the FDO, is to determine the inputs and output of
the observer. The ball and beam system has only two state variables, so choosing
the inputs of the observer is not a difficult task. As it was explained in Sect. 3,
one of the inputs of the observer is the control signal. Both states of the system
are influenced by the disturbance. Hence, they both can be chosen as the second
input. The simulations have shown that choosing the first state (the output of the
system) results in an FDO with better performance. Therefore, the control signal
and the output of the system are selected as the input variables, and the estimated
disturbance is the output variable. Gaussian MFs pictured in Fig. 7 are chosen for
all variables. There are 4 MFs for each input and output, including negative big
(NB), negative (N), positive (P), and positive big (PB). Table 2 presents the final
values of the optimized IT2 inputs’ parameters.
The inference engine contains 16 fuzzy rules that are presented in Table 3 and
have the following form:
Rule i : if u is G1 and r is G2 , then D̂ is G3 , i = 1, . . . , 16,
where G1 , G2 and G3 are the fuzzy set associated with the inputs and output
variables.
The crisp output is obtained using a weighted average defuzzification and KM
method in the type-reduction stage. Therefore, the fuzzy output of FDO can be
obtained by (8).

4.3 The MPSO setting

The inertia weight, w, is considered to be equal to 1 at the beginning, but in


every iteration, it linearly decreases to 0.98 of its value. The acceleration constants
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 15

are both set to 2. In the elimination phase of the algorithm, the elimination percent,
et , and the elimination period, ep , are set to 75 and 40, respectively, and the
population size is selected to be 40.
As stated in Subsect. 2.2, MPSO modifies the search space for each particle
based on itself. Table 4 shows the maximum and minimum values in the search
range given to the algorithm and the adjusted values by the algorithm itself. The
minimum and maximum values of the search space for the second input, r, are
parametrically determined in terms of the desired ball position, rd . Thus, the
observer is able to perform well for any reference signal.
Setting the cost function is an important issue in the optimization procedure.
Computing the MSE from the early seconds, when the system is not stable yet,
negatively affects the estimated disturbance after the stability. On the contrary,
neglecting the first seconds in computing the MSE, decreases the estimation er-
ror after the system achieves stability. In fact, there is a trade-off between the
estimated disturbance errors before and after the stability of the system. Fig. 8
shows the estimated disturbance in the case of computing the MSE from the third
second (case 1) and the first second (case 2) for the T1 and IT2 FDOs, in which
sinusoid disturbance is applied to the system in the time interval of 20 to 35 sec-
onds. In case 1, tracking the disturbance after the stability of the system in both
T1 and IT2 is almost the same, but peaking the estimated disturbance up in the
first seconds in T1, leads to an increase in the overshoot of the state response. In

Table 3 Fuzzy rules of the FDO.

u NB N P PB

NB P PB PB PB
N N P P PB
P NB N N P
PB NB NB NB N

0.1 Actual Disturbance Actual Disturbance


0.02
T1 FDO T1 FDO
IT2 FDO IT2 FDO
0.01
Disturbance Estimation

Disturbance Estimation

0.05

0 10 -3
10 -3 -0.01 5
5

-0.05 -0.02 0
0

-0.03
-5 -5
-0.1
32 34 36 32 34 36 38

0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time(sec) Time(sec)

(a) case 1 (b) case 2

Fig. 8 Estimated disturbance: (a) computing the MSE from the third second in the MPSO,
(b) computing the MSE from the first second in the MPSO.
16 Shokoufeh Naderi et al.

Table 4 Range of parameters of the optimized IT2 FDO.

Input1 Input2

given values adjusted values given values adjusted values

cmin -0.98 -1.37 -6 rd -7.2 rd


cmax 0.98 0.98 10 rd 12 rd
σmin 0.001 0.001 0.0004 0.0004
σmax 0.75 0.75 0.300 0.306

Table 5 Cost function values.

RMSE

case 1 case 2

T1 FDO 0.0140 0.0047


IT2 FDO 0.0072 0.0032

case 2, the superiority of the IT2 FDO over T1 FDO is quite clear. Table 5 shows
the values of RMSE in cases 1 and 2 for the T1 and IT2 FDOs, which indicates
a lower value of RMSE, and thus, a better performance in case 2 and for the IT2
FDO.

4.4 Simulation results

In order to exhibit the superiority of the proposed method, simulations draw


a comparison between the presented IT2 FDO, T1 FDO, and the proposed FDO
in (Kim 2002), which is still the most used FDO (Dou et al. 2017; Wang et al.
2016). The structure of the rules in T1 FDO is the same structure used in the IT2
FDO. The type and the number of MFs for both T1 and IT2 FDOs are the same,
and the parameters of the T1 FDO are also optimized using the MPSO. For the
proposed FDO in Kim (2002), results have been obtained without considering any
parameter uncertainty.
There are two scenarios concerning the type of the applied disturbance to the
system in the simulations:
1. Sinusoid disturbance with the amplitude of ±0.005 (10% of the reference) and
the frequency of 0.4π in the time interval of 15 to 25 seconds and with the
amplitude of ±0.003 and the frequency of 0.8π in the time interval of 35 to 50
seconds.
2. Square disturbance with the amplitude of ±0.005 in the time interval of 15 to
25 seconds and with the amplitude of ±0.003 in the time interval of 35 to 50
seconds.
Fig. 9 shows the simulation results for the first scenario. The disturbance es-
timation error of T1 FDO in the first 8 seconds (before the convergence of the
system) shows more fluctuations compared with the IT2 FDO, which results in
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 17

0.06
Actual Disturbance
T1 FDO
0.2
IT2 FDO 0.05
Fuzzy Adaptive DO
0.0504
Disturbance Estimation

-3
0.15 10

Ball Position (m)


0.04
5 0.05
0.1 0 0.03
0.0496
-5
15 20 25 30
0.05 0.02
15 20 25
Reference
No Disturbance Observer
0 0.01 T1 FDO
IT2 FDO
Fuzzy Adaptive DO
-0.05 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time(sec) Time(sec)

(a) (b)

0.35
T1 FDO
0.3 IT2 FDO
0 Fuzzy Adaptive DO
Disturbance Estimation Error

0.25 10 -3
10 -3
Control Signal (radian)

-0.05 0.2 1
2
0.15 0
0
-0.1 -1
0.1
-2
0.05 15 20 25 30
-0.15
15 20 25 30
0

T1 FDO -0.05
-0.2
IT2 FDO
Fuzzy Adaptive DO
-0.1

0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time(sec) Time(sec)

(c) (d)

Fig. 9 Simulation results in the first scenario: (a) disturbance and it’s estimation, (b) tracking
position, (c) estimation error, (d), control signal.

a higher maximum overshoot in the ball position response for T1 FDO. Conse-
quently, the control signal, θ, for T1 FDO fluctuates more than the IT2 FDO.
After the convergence of the system, the disturbance estimation for both T1 and
IT2 FDOs is almost the same, but the IT2 FDO achieves higher accuracy and
lower steady-state error. For the proposed FDO in Kim (2002), the estimated dis-
turbance rises in the first second and reaches a peak of 0.22, which leads to an
increase in the settling time of the system. This FDO cannot estimate the distur-
bance until the fifteenth second when the convergence of the system is achieved.
Furthermore, the estimated disturbance error after the fifteenth second is slightly
greater than T1 and IT2 FDOs.
The results of the second scenario are shown in Fig. 10. The performance of
the FDOs in the first eight seconds is similar to the first scenario. Compared
to the sinusoid disturbance, the square disturbance creates more fluctuations in
the estimated disturbance for T1 FDO. Again, these oscillations lead to more
fluctuation in the control effort for T1 FDO. The results for IT2 FDO do not
fluctuate, but the steady error for the estimation of negative square disturbance
is a little higher than it is for T1 FDO. The plot for the proposed FDO in Kim
(2002) illustrates a significant estimated disturbance error in comparison to the
T1 and IT2 FDOs.
18 Shokoufeh Naderi et al.

0.06
Actual Disturbance
T1 FDO
0.2
IT2 FDO 0.05
Fuzzy Adaptive DO
Disturbance Estimation

-3 0.0505
0.15 10

Ball Position (m)


0.04
5 0.05

0.1 0 0.0495
0.03
-5
-10 15 20 25 30
0.05 0.02
15 20 25 30
Reference
No Disturbance Observer
0 0.01 T1 FDO
IT2 FDO
Fuzzy Adaptive DO
-0.05 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time(sec) Time(sec)

(a) (b)

0.35
T1 FDO
0.3 IT2 FDO
0 Fuzzy Adaptive DO
Disturbance Estimation Error

0.25 10 -3
10 -3
Control Signal (radian)

0.2 3
-0.05
5
0.15 0
0
-0.1 -3
0.1
-5
0.05 15 20 25 30
-0.15
15 20 25 30
0

T1 FDO -0.05
-0.2
IT2 FDO
Fuzzy Adaptive DO
-0.1

0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time(sec) Time(sec)

(c) (d)

Fig. 10 Simulation results in the second scenario: (a) disturbance and it’s estimation, (b)
tracking position, (c) estimation error, (d) control signal.

In general, the obtained trajectories for the T1 and IT2 FDOs are very close;
however, it is evident that the IT2 FDO works a little better, especially, there is
a noticeable difference in decreasing the state overshoot.
The performance of the proposed FDO in Kim (2002), T1 and IT2 FDOs are
compared in Table 6 based on RMSE and integral time absolute error (ITAE)
criteria. In both scenarios, these values indicate a smaller error and a better per-
formance of the IT2 FDO compared with the T1 FDO. Moreover, values for the
T1 and IT2 FDOs are remarkably lower than those of the proposed FDO in Kim
(2002).
Although no parameter uncertainty has been applied in the simulations of the
proposed FDO in Kim (2002), poor performance can be observed for the considered
task. Thus, it cannot be applied to systems with a PD controller. Meanwhile, the
proposed FDO can operate at any level of uncertainty because the inputs of the
observer do not depend on the parameters of the system at all.
Overall, the simulation results prove that the IT2 FDO is superior compared
to the T1 FDO, and these two outperform the proposed FDO in Kim (2002). How-
ever, for the complex systems in which the FDO requires more inputs, defining the
rules of the observer could be a challenging and time-consuming task. In contrast,
for systems with less complexity, this method can be very effective and appropri-
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 19

ate, because applying the FDO does not change the controller, and uncertainty
cannot affect its performance. Furthermore, it can be used for systems subjected
to all types of disturbances, and in this method there is not any restriction on the
disturbance.
Regarding the computational complexity and required time, the IT2 FDO re-
quires more calculations and time, especially when its parameters are chosen us-
ing MPSO. The reason behind this is that IT2 FSs use another algorithm in the
type-reduction stage. However, they show a better performance in the results.
Therefore, there is no need to perform many experiments to find excellent values
of the parameters. Moreover, the parameters of the observer are optimized in an
offline manner, and once they are found, the observer can work for any reference
in the system. In other words, performing the MPSO in each run is not necessary.
Besides, the results in Sect. 5 will show that the MPSO is fast in convergence in
comparison to other optimization algorithms.
To provide a further comparison, another recent method (Cuevas et al. 2018)
is used to be compared with the proposed method. The differences between this
work and the proposed method are as follows:

• In Cuevas et al. (2018), a nonlinear system has been identified based on the
Hammerstein model. In other words, the nonlinear system is modeled through
the adaptation of an ANFIS system, taking advantage of the similarity between
the ANFIS and Hammerstein model, while in the proposed model in this paper,
the disturbance is modeled by an IT2 FS.
• In Cuevas et al. (2018), the gravitational search algorithm (GSA) has been
used to determine the parameters of the ANFIS model to avoid obtaining
suboptimal solutions by hybrid learning algorithm of the ANFIS. But in the
proposed model in this paper, parameters of the ITS FS are optimized using
the MPSO algorithm.
• In Cuevas et al. (2018), inputs of the ANFIS model have been the system input
and its delayed version, while the inputs in this paper are the system output
and control signal.

A disturbance observer is designed for the ball and beam system that was
described in this section according to the method proposed in Cuevas et al. (2018),
then it is compared with the IT2 FDO. The results of this comparison are shown in
Fig. 11. Before the stability of the system, the disturbance estimation error of the
IT2 FDO is significantly less than that of Cuevas et al. (2018) method. Besides,
there is a steady error in the disturbance estimation of Cuevas et al. (2018) method,
and in the presence of a negative square disturbance, the disturbance estimation

Table 6 Comparison between the performance of FDOs (simulation results).

RMSE ITAE

scenario 1 scenario 2 scenario 1 scenario 2

T1 FDO 0.0047 0.0048 0.9751 1.2199


IT2 FDO 0.0033 0.0033 0.3885 0.6430
FDO proposed in (Kim 2002) 0.0357 0.0358 2.5052 3.9196
20 Shokoufeh Naderi et al.

0.12
0.02
Cuevas (2018) method
T2 FDO
0.1
0

Disturbance Estimation Error


Disturbance Estimation

0.08 10 -3
-0.02 10 -3
5
5
0.06
-0.04 0 0
-5 0.04
-0.06 -5
-10
15 20 25 30 0.02
-0.08 15 20 25 30

Actual Disturbance 0
-0.1
Cuevas (2018) method
T2 FDO
-0.02
-0.12
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time(sec) Time(sec)

(a) case 1 (b) case 2

Fig. 11 Comparison between the IT2 FDO proposed in this paper and the method proposed
in Cuevas et al. (2018): (a) disturbance and its estimation,(b) estimation error.

of this method has more fluctuations. Table 7 shows the values of RMSE for these
two methods, which indicates a lower estimation error for the proposed IT2 FDO
in this paper.

Table 7 RMSE values for the IT2 FDO proposed in this paper and the method proposed in
Cuevas et al. (2018).

Cuevas et al. (2018) IT2 FDO

RMSE 0.0479 0.0032

Compared to the presented method in Cuevas et al. (2018), the number of


adjustable parameters in the IT2 FDO is greater, which is a disadvantage. Never-
theless, the IT2 FDO performs better, because T2 FSs provide more robustness,
taking advantage of the FOU of MFs. Moreover, the MPSO algorithm avoids con-
vergence to sub-optimal solutions due to its higher accuracy and convergence rate,
which was explained in Sect. 5 and its ability to define the upper and lower search
bound of parameters based on their values.

4.5 Experimental results

The experimental setup of the ball and beam system is presented in Fig. 12. The
sensor outputs are amplified before transmitting to the digital ports of the data
acquisition card, and the analog output of this card is connected to the amplifier to
carry the attenuated motor voltage control signal. The amplifier is then connected
to the motor to transmit the amplified voltage, which is applied to the motor.
Observer rules and system parameters in experimental results are similar to those
in the simulation section. The amplitude of the applied disturbance to the system
is more than the simulation section to make the results clearer.
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 21

Fig. 12 Experimental setup of the ball and beam system.

The PD controller of the inner loop cannot be used to control the ball position
in the actual ball and beam system directly. The position of the ball is measured
through an analog sensor, which is inherently noisy. Taking the derivative of this
noisy signal causes the output to be in an amplified high-frequency signal, which
is finally fed back into the motor and makes a grinding noise. In the practical
implementation, a high-pass filter is used to prevent this.
Fig. 13 illustrates the disturbance estimation, tracking position, and control
signal of the T1 and IT2 FDOs in the second scenario. According to Fig. 13(a)
and Fig. 13(c), the estimated disturbance of the IT2 FDO in the early seconds has
less error and fluctuation than the T1 FDO. The difference between settling times
and overshoot percentages of tracking position for the T1 and IT2 FDOs shown
in Fig. 13(b) is considerable, and Fig. 13(d) shows less fluctuation of the control
signal for the IT2 FDO. Table 8 contains the RMSE and ITAE values of estimation
errors. The lower value of these errors for the IT2 FDO verifies its superiority in
disturbance tracking.

Table 8 Comparison between the performance of T1 and IT2 FDOs (experimental results).

RMSE ITAE

T1 FDO 0.0156 1.5296


IT2 FDO 0.0095 0.9370

To provide a statistic analysis of experimental results, the Wilcoxon test (Garcı́a


et al. 2009) is used. Results are shown in Table 9, from which we can discover
whether there is a statistically significant difference between the T1 and IT2 FDOs.
22 Shokoufeh Naderi et al.

0.04

0.02
Disturbance Estimation

-0.02

-0.04

No Disturbance Observer
T1 FDO
-0.06
IT2 FDO

0 5 10 15 20 25
Time(sec)

(a) (b)

0.05
50 No Disturbance Observer
T1 FDO
0.04 IT2 FDO 0
Disturbance Estimation Error

0.03 -50
Control Signal (radian)

0 5 10 15 20 25
0.02
50 T1 FDO
0.01
0
0
-50
-0.01 0 5 10 15 20 25

-0.02 50 IT2 FDO

0
-0.03
-50
-0.04
0 5 10 15 20 25 0 5 10 15 20 25
Time(sec) Time(sec)

(c) (d)

Fig. 13 Experimental results in the second scenario: (a) disturbance and it’s estimation, (b)
tracking position, (c) estimation error, (d) control signal.

The Wilcoxon signed-rank test shows that the null hypothesis has been rejected,
according to the p − value < 0.001. Indeed, based on the Z test (-5.542) which
is significant at the error level of less than 0.01, it can be said that statistically
with a confidence level of 99%, there is a highly significant difference between the
performance of the T1 and IT2 FDOs.

Table 9 Results of the Wilcoxon test.

Z p-value

T1 vs IT2 FDO -5.542 0.000

A comparison between the simulation and experimental results shows that the
settling time of the system in practical results is more than its value in simulation
results. The reason for this inequality is the difference between the initial values
of the system states. In the simulation, the initial condition is chosen to be zero,
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 23

Table 10 Comparison between the performance of evolutionary algorithms in 500 iterations.

Evolutionary algorithm Cost function value Elapsed time (hour)

MPSO 25.44 4.58


IGA 26.93 7.38
ICA 27.69 5.58
DE 28.44 6.93
ABC 29.81 6.25
SFLA 31.62 12.26

while in the experiment, the initial value of the first state is equal to half the beam
length (−20) because the ball is first situated at the end of the beam.

5 Comparison between the MPSO and other evolutionary algorithms

In this section, a comparison between the performance of some well-known


evolutionary algorithms is provided to prove the desirable performance of the
MPSO algorithm in optimizing IT2FSs.
The considered algorithms for the comparison are as follows:

1. MPSO (Salahshour et al. 2018)


2. Improved genetic algorithm (IGA) (Leung et al. 2003)
3. Artificial bee colony (ABC) (Karaboga and Basturk 2008)
4. Differential evolution (DE) (Storn and Price 1997)
5. Imperialist competitive algorithm (ICA) (Atashpaz-Gargari and Lucas 2007)
6. Shuffled frog-leaping algorithm (SFLA) (Eusuff and Lansey 2003)

All the algorithms have been used to optimize parameters of the IT2 FDO
presented in Sect. 4 with the same inputs and rules. In order to have a valid and
fair comparison, initial conditions, population size, search space, and cost function
have been considered the same for all algorithms. They all have been run five
times, and the best result of each algorithm is illustrated in Table 10, in which
the performance of the algorithms in 500 iterations is shown. In the table, it can
be clearly observed that among the examined algorithms the MPSO results in
the lowest elapsed time and cost function value. In other words, it has the least
computational complexity and outperforms the others in terms of the elapsed
time. Fig. 14 shows the convergence profile of cost functions in 500 iterations. The
plot for the MPSO continuously decreases over the whole 500 iterations, while the
cost function value of the other algorithms levels off and remains unchanged after
reaching a specific iteration.
Considering the time-based comparison, all algorithms have been run for five
hours, and results are shown in Table 11. This table provides a comparison in terms
of the iteration and convergence rate. The number of iterations in five hours is
higher for the MPSO compared to others by a marked difference of approximately
100 iterations, which shows the higher speed of this algorithm. Moreover, it has
the lowest number of cost function values among the six, verifying its accuracy
and faster convergence rate.
24 Shokoufeh Naderi et al.

Table 11 Comparison between the performance of evolutionary algorithms for 5 hours of run.

Evolutionary algorithm Cost function value Iteration number

MPSO 25.43 545


IGA 26.93 328
ICA 27.69 443
DE 28.44 352
ABC 29.81 392
SFLA 31.62 186

The main characteristic of the MPSO, which helps to improve the accuracy
and convergence rate is that the algorithm avoids falling into local extremums,
by substituting the weaker particles with the highest cost function values by new
ones. It takes a long time for other algorithms to go out of local extremums and
improve their solutions due to their random search. Moreover, the MPSO can
change its search space and modify it when the optimized solution is close to the
lower or upper bound of the search space. These features have made the algorithm
to show a satisfactory result and a fast convergence rate without a large number
of iterations to be performed.
Since an IT2 MF is defined based on two T1 MFs, the number of the parameters
of an IT2FS which are needed to be optimized are about twice as many as this
number for a T1FS. The large number of these parameters and the fact that the
precise bound of the search space is not known, make the MPSO a suitable choice
for optimizing IT2FSs.

75
MPSO
70 31.62 SFLA
DE
65 IGA
29.81 ICA
Cost Function Value

60 ABC

55 28.44
27.69
50 26.93
45
25.45
40
440 460 480 500
35

30

25
0 100 200 300 400 500
Iteration

Fig. 14 Cost function values in 500 iteration.


Designing an IT2 FDO for a class of nonlinear systems based on MPSO 25

6 Conclusion

In this study, the design procedure of an IT2 FDO, which is not affected by
uncertainties, has been proposed for nonlinear systems. The MPSO algorithm has
been applied to optimize parameters of the inputs and output MFs in an offline
manner, after defining the initial structure of the FDO. One of the main contribu-
tions of this work is through the use of IT2 fuzzy systems as being the extended
version of T1 fuzzy systems to provide more robustness by taking advantage of
the FOU of the MF, and the other is through the use of the MPSO algorithm to
improve the accuracy and convergence rate in optimization. Comparisons between
the MPSO and other evolutionary algorithms have shown the higher ability of the
MPSO to avoid the local extremums. By the means of the FDO, the uncertain
factors, including disturbance, have been estimated without requiring any prior
information about them and adding extra complexity to the controller. To evalu-
ate the performance of the presented FDO, it has been tested on a ball and beam
system. Simulation and experimental results have indicated the desirable perfor-
mance of the IT2 FDO and its superiority over other FDOs in terms of reducing
the settling time, overshoot, and fluctuation of the state response.
Due to the satisfactory performance of the proposed FDO, it can be widely
used in practical applications. Applying this observer to other systems affected
by disturbance in the real world and designing generalized T2 FDO as being an
improvement over the IT2 fuzzy logic system can be considered in future research.

References

Atashpaz-Gargari E, Lucas C (2007) Imperialist competitive algorithm: an algorithm for op-


timization inspired by imperialistic competition. In: 2007 IEEE congress on evolutionary
computation, IEEE, pp 4661–4667
Capuano N, Chiclana F, Herrera-Viedma E, Fujita H, Loia V (2019) Fuzzy group decision
making for influence-aware recommendations. Computers in Human Behavior 101:371–
379
Castillo O, Amador-Angulo L (2018) A generalized type-2 fuzzy logic approach for dynamic
parameter adaptation in bee colony optimization applied to fuzzy controller design. Infor-
mation Sciences 460:476–496
Castillo O, Neyoy H, Soria J, Melin P, Valdez F (2015) A new approach for dynamic fuzzy
logic parameter tuning in ant colony optimization and its application in fuzzy control of a
mobile robot. Applied soft computing 28:150–159
Castillo O, Cervantes L, Soria J, Sanchez M, Castro JR (2016) A generalized type-2 fuzzy
granular approach with applications to aerospace. Information Sciences 354:165–177
Chen M, Ge SS (2015) Adaptive neural output feedback control of uncertain nonlinear systems
with unknown hysteresis using disturbance observer. IEEE Transactions on Industrial
Electronics 62(12):7706–7716
Cuevas E, González A, Zaldı́var D, Pérez-Cisneros M (2015) An optimisation algorithm based
on the behaviour of locust swarms. International Journal of Bio-Inspired Computation
7(6):402–407
Cuevas E, Dı́az P, Avalos O, Zaldı́var D, Pérez-Cisneros M (2018) Nonlinear system identi-
fication based on anfis-hammerstein model using gravitational search algorithm. Applied
Intelligence 48(1):182–203
Dorigo M, Maniezzo V, Colorni A (1996) Ant system: optimization by a colony of cooperat-
ing agents. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics)
26(1):29–41
Dou L, Su P, Zong Q, Ding Z (2017) Fuzzy disturbance observer-based dynamic surface control
for air-breathing hypersonic vehicle with variable geometry inlets. IET Control Theory &
Applications 12(1):10–19
26 Shokoufeh Naderi et al.

Duan Y, Gao Y (2011) A new improved particle swarm optimization algorithm. In: Computa-
tional Intelligence and Security (CIS), 2011 Seventh International Conference on, IEEE,
pp 42–46
Duran K, Bernal H, Melgarejo M (2008) Improved iterative algorithm for computing the gen-
eralized centroid of an interval type-2 fuzzy set. In: Fuzzy Information Processing Society,
2008. NAFIPS 2008. Annual Meeting of the North American, IEEE, pp 1–5
Eusuff MM, Lansey KE (2003) Optimization of water distribution network design using the
shuffled frog leaping algorithm. Journal of Water Resources planning and management
129(3):210–225
Faramin M, Rezaie B, Rahmani Z (2019) Robust integral feedback control based on interval
observer for stabilising parameter-varying systems. IET Control Theory & Applications
13(15):2455–2464
Garcı́a S, Molina D, Lozano M, Herrera F (2009) A study on the use of non-parametric tests
for analyzing the evolutionary algorithms’ behaviour: a case study on the cec’2005 special
session on real parameter optimization. Journal of Heuristics 15(6):617
Huang J, Zhang M, Ri S, Xiong C, Li Z, Kang Y (2019) High-order disturbance-observer-based
sliding mode control for mobile wheeled inverted pendulum systems. IEEE Transactions
on Industrial Electronics 67(3):2030–2041
Juang CF, Hsu CH (2009) Reinforcement interval type-2 fuzzy controller design by online rule
generation and q-value-aided ant colony optimization. IEEE Transactions on Systems,
Man, and Cybernetics, Part B (Cybernetics) 39(6):1528–1542
Juang CF, Tsao YW (2008) A self-evolving interval type-2 fuzzy neural network with online
structure and parameter learning. IEEE Transactions on Fuzzy Systems 16(6):1411–1424
Karaboga D (2005) An idea based on honey bee swarm for numerical optimization. Tech. rep.,
Technical report-tr06, Erciyes university, engineering faculty, computer . . .
Karaboga D, Basturk B (2008) On the performance of artificial bee colony (abc) algorithm.
Applied soft computing 8(1):687–697
Karimi Z, Batmani Y, Khosrowjerdi MJ (2018) Multiobjective fault-tolerant fixed-order/pid
control of multivariable discrete-time linear systems with unmeasured disturbances. Opti-
mal Control Applications and Methods 39(5):1648–1662
Karnik NN, Mendel JM (2001) Centroid of a type-2 fuzzy set. Information Sciences 132(1-
4):195–220
Kennedy J, Eberhart R (1995) Particle swarm optimization. proceeding of IEEE International
Conference on Neural Network pp 1942–1948
Kim E (2002) A fuzzy disturbance observer and its application to control. IEEE Transactions
on Fuzzy Systems 10(1):77–84
Kim E, Park C (2004) Fuzzy disturbance observer approach to robust tracking control of non-
linear sampled systems with the guaranteed suboptimal h∞ performance. IEEE Transac-
tions on Systems, Man, and Cybernetics, Part B (Cybernetics) 34(3):1574–1581
Lang G, Miao D, Fujita H (2019) Three-way group conflict analysis based on pythagorean
fuzzy set theory. IEEE Transactions on Fuzzy Systems
Leung FHF, Lam HK, Ling SH, Tam PKS (2003) Tuning of the structure and parameters
of a neural network using an improved genetic algorithm. IEEE Transactions on Neural
networks 14(1):79–88
Liang Q, Mendel JM (2000a) Designing interval type-2 fuzzy logic systems using an svd-qr
method: Rule reduction. International Journal of Intelligent Systems 15(10):939–957
Liang Q, Mendel JM (2000b) Interval type-2 fuzzy logic systems: theory and design. IEEE
Transactions on Fuzzy systems 8(5):535–550
Lin YY, Liao SH, Chang JY, Lin CT, et al. (2014) Simplified interval type-2 fuzzy neural
networks. IEEE Trans Neural Netw Learning Syst 25(5):959–969
Liu H, Guo L, Zhang Y (2012) An anti-disturbance pd control scheme for attitude control and
stabilization of flexible spacecrafts. Nonlinear Dynamics 67(3):2081–2088
Maldonado Y, Castillo O, Melin P (2013) Particle swarm optimization of interval type-2 fuzzy
systems for fpga applications. Applied Soft Computing 13(1):496–508
Melgarejo M (2007) A fast recursive method to compute the generalized centroid of an interval
type-2 fuzzy set. In: Fuzzy Information Processing Society, 2007. NAFIPS’07. Annual
Meeting of the North American, IEEE, pp 190–194
Mendel JM (2004) Computing derivatives in interval type-2 fuzzy logic systems. IEEE Trans-
actions on Fuzzy Systems 12(1):84–98
Mendel JM (2017) Uncertain rule-based fuzzy systems. Springer
Designing an IT2 FDO for a class of nonlinear systems based on MPSO 27

Meng FY, Tang J, Fujita H (2019) Consistency-based algorithms for decision-making with
interval fuzzy preference relations. IEEE Transactions on Fuzzy Systems 27(10):2052–2066
Oh SK, Jang HJ, Pedrycz W (2011) A comparative experimental study of type-1/type-2 fuzzy
cascade controller based on genetic algorithms and particle swarm optimization. Expert
Systems with Applications 38(9):11217–11229
Ontiveros E, Melin P, Castillo O (2018) High order α-planes integration: a new approach to
computational cost reduction of general type-2 fuzzy systems. Engineering Applications of
Artificial Intelligence 74:186–197
Rudas IJ, Batyrshin IZ, Zavala AH, Nieto OC, Vargas LV (2009) Digital fuzzy parametric
conjunctions for hardware implementation of fuzzy systems. In: 2009 IEEE International
Conference on Computational Cybernetics (ICCC), IEEE, pp 157–166
Salahshour E, Malekzadeh M, Gholipour R, Khorashadizadeh S (2018) Designing multi-layer
quantum neural network controller for chaos control of rod-type plasma torch system using
improved particle swarm optimization. Evolving Systems pp 1–15
Shaocheng T, Jiantao T, Tao W (2000) Fuzzy adaptive control of multivariable nonlinear
systems1. Fuzzy sets and systems 111(2):153–167
Shi Y, Eberhart R (1998) A modified particle swarm optimizer. In: Evolutionary Computation
Proceedings, 1998. IEEE World Congress on Computational Intelligence., The 1998 IEEE
International Conference on, IEEE, pp 69–73
Shi Y, Eberhart RC (1999) Empirical study of particle swarm optimization. In: Evolutionary
computation, 1999. CEC 99. Proceedings of the 1999 congress on, IEEE, vol 3, pp 1945–
1950
Storn R, Price K (1997) Differential evolution–a simple and efficient heuristic for global opti-
mization over continuous spaces. Journal of global optimization 11(4):341–359
Tao CW, Taur JS, Chang CW, Chang YH (2012) Simplified type-2 fuzzy sliding controller for
wing rock system. Fuzzy sets and systems 207:111–129
Tseng CS, Chen BS, Uang HJ (2001) Fuzzy tracking control design for nonlinear dynamic
systems via ts fuzzy model. IEEE Transactions on fuzzy systems 9(3):381–392
Wang S, Ren X, Na J (2016) Adaptive dynamic surface control based on fuzzy disturbance ob-
server for drive system with elastic coupling. Journal of the franklin institute 353(8):1899–
1919
Wu D, Mendel JM (2009) Enhanced karnik–mendel algorithms. IEEE Transactions on Fuzzy
Systems 17(4):923–934
Wu D, Nie M (2011) Comparison and practical implementation of type-reduction algorithms
for type-2 fuzzy sets and systems. In: FUZZ-IEEE, pp 2131–2138
Wu GD, Huang PH (2013) A vectorization-optimization-method-based type-2 fuzzy neural
network for noisy data classification. IEEE Trans Fuzzy Systems 21(1):1–15
Yao X, Guo L (2013) Composite anti-disturbance control for markovian jump nonlinear systems
via disturbance observer. Automatica 49(8):2538–2545
Yoo W, Ji D, Won S (2010) Synchronization of two different non-autonomous chaotic systems
using fuzzy disturbance observer. Physics Letters A 374(11-12):1354–1361
Zadeh LA (1965) Fuzzy sets. Information and control 8(3):338–353
Zhang L, Tang Y, Hua C, Guan X (2015) A new particle swarm optimization algorithm with
adaptive inertia weight based on bayesian techniques. Applied Soft Computing 28:138–149
Zhou H, Ying H (2013) A method for deriving the analytical structure of a broad class of
typical interval type-2 mamdani fuzzy controllers. IEEE Transactions on Fuzzy Systems
21(3):447–458

You might also like