ISA Transactions xxx (xxxx) xxx

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Practice article

Modeling, simulation and experimental realization of a new nonlinear

fuzzy PID controller using Center of Gravity defuzzification
∗
Debdoot Sain , B.M. Mohan
Department of Electrical Engineering, Indian Institute of Technology, Kharagpur, 721302, India

article info a b s t r a c t

Article history: Though Center of Gravity (CoG) defuzzification is a well-known and long-standing method in the
Received 7 August 2019 history of fuzzy systems, because of its computational complexity, its use in the field of modeling of
Received in revised form 11 October 2020 fuzzy controllers is almost nil. From literature, it appears that modeling of fuzzy Proportional Integral
Accepted 14 October 2020
Derivative (FPID) controllers is rarely attempted using CoG defuzzification. In fact, none of the FPID
Available online xxxx
controller models are obtained using both two-dimensional input space and CoG defuzzification. The
Keywords: available mathematical models of fuzzy Proportional Integral (FPI) and fuzzy Proportional Derivative
Fuzzy PID (FPD) controllers using two-dimensional input space and CoG defuzzification were due to Arun and
Nonlinear PID Mohan (2017). In this paper, the authors make an attempt to model and design an FPID controller using
Mathematical modeling two-dimensional input space and CoG defuzzification. The incremental control effort produced by the
Simulation
newly developed FPID controller is found by combining the individual control efforts produced by
Time-delay system
incremental FPI and incremental FPD controllers. The incremental FPI and incremental FPD controller
Real-time experimentation
Magnetic levitation structures are unveiled using two-dimensional input space, CoG defuzzification, Min t-norm, Max t-
conorm, and Larsen Product (LP) inference. The applicability and usefulness of the newly obtained FPID
controller are depicted with simulation and real-time experimentation.
© 2020 ISA. Published by Elsevier Ltd. All rights reserved.

1. Introduction equations or difference equations. Otherwise, the system is said

to be ill-defined. For ill-defined plants, fuzzy controllers are ex-
tremely useful. Though fuzzy controllers have been employed to
Background: Since inception, PID controllers have become one of
tackle various control situations, their input–output relationship
the favorite choices of control engineers, and over the years, they
was not known for long. The activity on the modeling of fuzzy
have been employed successfully to resolve numerous problems
controllers was started in the beginning of 1990s, and it is still
in the field of control. Simple and straightforward structure, ease
going on. Initially, modeling of fuzzy controllers was done for the
of implementation, and low cost of maintenance are the primary
simplest FPI and the simplest FPD controllers. As both transient
reasons behind their widespread popularity. However, it is ob-
and steady-state performances cannot be guaranteed by FPI and
served that PID controllers do not perform convincingly for linear
FPD controllers individually, researchers started to model FPID
systems of higher-order, nonlinear systems, and systems with
controllers.
time-delay. Moreover, it has also been noticed that when applied
to an unstable system, undesirable zeros are introduced by the Literature survey: Few important works related to the develop-
conventional PID controllers in the closed loop (CL) transfer func- ments of FPID controllers are discussed now. Realization of FPID
tion. These inappropriate zeros are responsible for high overshoot controllers using product-sum-gravity and simple fuzzy reason-
and reduced loop robustness. As a remedy to these problems, ing method was addressed by Mizumoto [1]. Analysis and design
different concepts have been adopted and clubbed together to of an FPID controller were done by Misir et al. [2] where the
make PID controllers more reliable and powerful over time. FPID control law was obtained by algebraically summing FPI law
While several control techniques are available for the control and fuzzy D law. The stability analysis of FPID controller was also
of well-defined systems, only a few actually take care of both addressed using Small-Gain theorem. Homaifar et al. [3] provided
well-defined and ill-defined systems. A system is said to be well- a systematic and hierarchical hybrid FPID controller design ap-
defined if it is possible to represent its dynamics with differential proach by applying a learning based algorithm. By representing
the fuzzy rules in various forms, Mann et al. [4] investigated the
∗ Corresponding author. structures of various FPID controllers. Hayashi et al. [5] showed
E-mail addresses: [email protected] (D. Sain), [email protected] that nonlinear FPID controller can easily be realized based on
(B.M. Mohan). conventional PID controller by employing simple fuzzy inference

https://doi.org/10.1016/j.isatra.2020.10.048
0019-0578/© 2020 ISA. Published by Elsevier Ltd. All rights reserved.

Please cite this article as: D. Sain and B.M. Mohan, Modeling, simulation and experimental realization of a new nonlinear fuzzy PID controller using Center of Gravity
defuzzification. ISA Transactions (2020), https://doi.org/10.1016/j.isatra.2020.10.048.
D. Sain and B.M. Mohan ISA Transactions xxx (xxxx) xxx

mechanism. Using incremental FPI and incremental FPD con- to the use of three-dimensional input space and CoG defuzzifi-
trollers, an FPID controller was constructed by Kim and Oh [6] cation, overall modeling and controller implementation become
for the control of nonlinear and uncertain systems. A PID-like extremely difficult.
fuzzy controller with reduced complexity was proposed by Tao
Motivation: From the literature, it seems to the authors that
and Taur [7] where the authors formed the input to fuzzy block by
modeling of FPID controllers is rarely attempted using CoG de-
combining the input variables in a linear fashion. However, such
fuzzification. In fact, none of the FPID controller models are
controller may not perform satisfactorily due to the reduction
obtained using both two-dimensional input space and CoG de-
of Degree of Freedom (DoF) while taking the combination of in-
fuzzification. The available mathematical models of FPI and FPD
puts. Different types of decomposed FPID controllers were tested
controllers using two-dimensional input space and CoG defuzzifi-
and compared by Golob [8]. These controllers, in their simplest
cation were due to Arun and Mohan [27] where the authors have
form, use three single input single output inferences with three
shown that CoG method is more accurate defuzzification strategy
different rule bases. One important feature of these controllers
as compared to the CoS method. As these controllers cannot sat-
is their simple structure and ease of implementation. Theory,
isfy the transient and steady-state specifications simultaneously,
design and simulation of predictive FPID controllers were studied
there is a need to develop an FPID controller which is accurate
by Lu et al. [9]. The advantage of such FPID controllers is their
and whose modeling is simple as compared to the modeling of
capability to control uncertain and complex linear and nonlinear
existing FPID controller [26] developed using three-dimensional
time-delay systems. A function dependent assessment strategy
input space. Without further ado, let us now state the objective
was addressed by Hu et al. [10] for the methodical analysis of FPID
of the present paper.
controllers. Using input fuzzy sets with nonlinear membership
functions (MFs), structural analysis of nonlinear FPID controllers Objective and brief overview: In this paper, the authors make
was performed by Haj-Ali and Ying [11]. A novel algorithm for an attempt to model and design an FPID controller using two-
providing the analytical solution of three-input FPID controllers dimensional input space and CoG defuzzification. The incremental
was proposed by Mann and Gosine [12] where the analysis was control effort produced by the proposed FPID controller is found
done with the help of Zadeh–Mamdani’s min–max-gravity fuzzy by combining the individual control signals generated by in-
reasoning. Modeling and stability analysis of few simplest FPID cremental FPI and incremental FPD controllers. The incremental
controllers were done by Mohan and Sinha [13] using Algebraic FPI and incremental FPD controller structures are unveiled using
Product (AP) AND operator, different combinations of OR op- Min t-norm, Max t-conorm, LP inference, CoG defuzzification,
erators and inference mechanisms, and Center of Sums (CoS) and two-dimensional input space. Note that such combination
defuzzification strategy. Analytical structures of few other FPID was not explored by Arun and Mohan [27] for the modeling of
controllers were derived by Mohan and Sinha [14], where they FPI and FPD controllers. The applicability and usefulness of the
showed that these controllers are unsuitable for the purpose of newly obtained FPID controller are depicted with one simulation
control. Karasakal et al. [15] developed a new strategy for on- example and one real-time experimentation. An unstable linear
line adjustment of fuzzy rule weights of the FPID controllers time-delay system is controlled in simulation, whereas an un-
and the efficiency of the method was depicted with simulation stable and nonlinear magnetic levitation system is considered
and real-time studies on linear and nonlinear plants. A fractional during experimental realization. The effectiveness of the newly
order FPID controller was proposed by Das et al. [16] where developed FPID controller is depicted by comparing its perfor-
the error and its fractional order derivative work as the inputs mance with that of the other controllers. Disturbance rejection
to the fuzzy logic controller and a fractional order integrator ability of the newly obtained FPID controller is demonstrated
is present in the output of the fuzzy controller. For nonlinear by applying disturbances. Moreover, time-delay associated with
uncertain chaotic processes, an interval Type-2 (IT2) optimal FPID the simulation example is increased to check how sensitive the
controller was obtained by Khooban et al. [17]. Online tuning, control system is to change in delay and how effectively the
simulation and experimental realization of FPID controllers via proposed FPID controller handles such a situation. Note that the
rule weighing depending on normalized acceleration were per- unstable linear time-delay system and magnetic levitation system
formed by Karasakal et al. [18]. Bounded Input Bounded Output were also considered by Sain and Mohan [28] but the systems
(BIBO) stability analysis of fuzzy P plus I plus D controller was were controlled using FPI and FPD controllers, respectively, not
done by Kumar et al. [19]. A novel multivariable predictive FPID with FPID controller. The FPI and FPD controllers used by the
controller was developed by Savran [20] where the PID and authors [28] were developed using height defuzzification which
fuzzy control strategies were incorporated in the framework of is least accurate as compared to the CoS and CoG methods.
predictive control. For nonlinear and uncertain systems, a fuzzy Moreover, the problem of steady-state error was noticed with
model dependent adaptive PID control scheme was developed the FPD controller designed by Sain and Mohan [28]. But in the
by Savran and Kahraman [21]. Mathematical modeling and BIBO present study such a case does not arise. Though same examples
stability analysis of IT2 FPID controller were performed by El- are considered in the present paper, but the underlying concepts
Nagar and El-Bardini [22] where the IT2 FPID controller structure used in both the papers are completely different.
was formed by combining one IT2 FPI controller and another
IT2 FPD controller in parallel. For enhancing the performance, Organization of the paper: The present paper is organized into
Fereidouni et al. [23] proposed an adaptive configuration for the seven sections. Sections 2 and 3 deal with the details of con-
FPID controller where output scaling factors are automatically ventional PID and FPID controllers. Mathematical modeling of
updated during controller operation. For multivariable control the proposed FPID controller is depicted in Section 4. Simulation
systems, a novel model based framework for the tuning of scal- and experimental realization of the newly obtained FPID con-
ing factors of FPID controller was proposed by Gil et al. [24] troller are shown respectively in Sections 5 and 6. In Section 7,
where a nonlinear optimization problem was solved by simulta- conclusion of the paper is presented.
neously satisfying a set of constraints. Using CoS defuzzification,
analytical structures of two nonlinear FPID controllers were ob- 2. Conventional PID controller
tained by Arun and Mohan [25]. Recently, Sain and Mohan [26]
obtained the mathematical model of an FPID controller using A typical CL computer control system is depicted in Fig. 1,
three-dimensional input space and CoG defuzzification. But due where the plant is Continuous Time (CT) in nature and T , di (t),
2
D. Sain and B.M. Mohan ISA Transactions xxx (xxxx) xxx

Fig. 1. A typical CL computer control system.

Fig. 2. FPID controller block diagram.

and do (t) denote the sampling time, input disturbance and output dt
= {KP1 ∆e(k) + KIdt e(k)} + {KP2 dt
∆e(k) + KDdt ∆2 e(k)}
disturbance, respectively.
= {∆
ˆ es (k) + ês (k)} + {∆
˜ es (k) + ∆˜ 2 es (k)}
The input–output relation of a CT linear PID controller, in
terms of error e(t) and control effort uPID (t), is given by = ∆uPI (k) + ∆uPD (k) (9)
∫ t
de(t) where KPdt = dt
KP1+ dt
KP2 ,
ês (k) = KIdt e(k), ∆es (k)
ˆ = dt
KP1 ∆e(k),
uPID (t) = KPct e(t) + KIct e(τ )dτ + KDct (1)
0 dt ∆˜ es (k) = KP2
dt
∆e(k) and ∆˜ 2 es (k) = KDdt ∆2 e(k). KP1
dt dt
and KP2 rep-
where KPct , KIct and KDct denote the gains associated with the linear resent the proportional gain of DT linear incremental PI and in-
P, I, and D controllers. Differentiating Eq. (1) with respect to t cremental PD controllers, respectively, ês (k) and ∆ ˆ es (k) represent
gives the scaled version of error and change of error of the DT linear
incremental PI controller, respectively, and ∆ ˜ es (k) and ∆ ˜ 2 es (k)
duPID (t) de(t) d2 e(t) represent the scaled version of change of error and double change
= KPct + KIct e(t) + KDct (2)
dt dt dt 2 of error of the DT linear incremental PD controller, respectively.
Henceforth for brevity we drop T in kT and (k − 1)T . At t = kT , From Eq. (9), it can be observed that the incremental control
we have signal generated by DT PID controller can be found by adding the
duPID (t) individual incremental control signals produced by DT PI and PD
= T −1 {uPID (k) − uPID (k − 1)} = T −1 ∆uPID (k)
∼ (3) controllers.
dt
de(t)
= T −1 {e(k) − e(k − 1)} = T −1 ∆e(k)
∼ (4) 3. Fuzzy PID controller
dt
e(t) = e(k) and (5) The structure and principal components of FPID controller are
d2 e(t) discussed in this section. Like DT linear incremental PID con-
= T {∆e(k) − ∆e(k − 1)} = T ∆ e(k)
∼ −1 −2 2
(6)
troller, the DT incremental FPID controller structure is achieved
dt 2
Using Eqs. (3)–(6), Eq. (2) can be expressed as by combining one incremental FPI controller and another in-
cremental FPD controller in parallel. The block diagram of the
∆uPID (k) = KPdt ∆e(k) + KIdt e(k) + KDdt ∆2 e(k) (7) proposed FPID controller is depicted in Fig. 2.
where ∆uPID (k) denotes the incremental control signal produced From Fig. 2, it can be observed that the scaled inputs of
by Discrete Time (DT) linear PID controller and KPdt (= KPct ), KIdt (= incremental FPI and incremental FPD controllers are esPI (k) =
K ct SePI e(k), ∆esPI (k) = S∆ePI ∆e(k), ∆esPD (k) = S∆ePD ∆e(k), and
KIct T ), and KDdt (= TD ) respectively represent the gains associated ∆2 esPD (k) = S∆2 ePD ∆2 e(k). The scaled versions of outputs of
with DT linear P, I, and D controllers.
the incremental FPI and incremental FPD controllers are given
The overall control effort produced by the DT linear PID con-
by ∆usPI (k) and ∆usPD (k), respectively. The outputs of fuzzy
troller at kth sampling moment is as follows:
controllers are usually nonlinear functions of inputs and the type
uPID (k) = uPID (k − 1) + ∆uPID (k) (8) of nonlinearity depends on many factors such as choice of MFs,
t-norms, t-conorms, inference methods, defuzzification strategies
Now Eq. (7) can be written as
etc. In Fig. 2, SePI , S∆ePI , S∆ePD , S∆2 ePD , SPI
−1 −1
and SPD denote
∆uPID (k) = (KP1
dt dt
+ KP2 )∆e(k) + KIdt e(k) + KDdt ∆2 e(k) the scaling factors of the incremental FPI and incremental FPD
3
D. Sain and B.M. Mohan ISA Transactions xxx (xxxx) xxx

Table 1
Incremental FPI controller Rule base.
Rule # IF
esPI (k) AND ∆esPI (k) THEN ∆usPI (k)
R1 −1
EPI ∆EPI
−1
∆UPI−1
R2a −1
EPI ∆EPI+1 ∆UPI0
R2b +1
EPI ∆EPI−1 ∆UPI0
R3 +1
EPI ∆EPI+1 ∆UPI+1

Fig. 3. MFs on input variables. 1. Response is above the reference and goes away from it
2. Response is above the reference and approaches it
3. Response is below the reference and approaches it
4. Response is below the reference and goes away from it
Each of which will be taken care of by one of the rules described
in the above rule base [29]. From the rule base it can be observed
that the outcomes of Rules R2a and R2b are the same. Instead of
having two separate rules, these two rules are merged into one
using fuzzy OR operator and the modified rule is called R2 .
The above rule base also works for the incremental FPD con-
troller provided esPI (k), ∆esPI (k), EPI
−1
, ∆EPI
−1 +1
, EPI , ∆EPI
+1
, ∆UPI−1 ,
Fig. 4. MFs on output variables.
∆UPI0 and ∆UPI+1 are replaced with ∆esPD (k), ∆2 esPD (k), ∆EPD −1
,
∆ EPD , ∆EPD , ∆ EPD , ∆UPD , ∆UPD and ∆UPD , respectively.
2 −1 +1 2 +1 −1 0 +1

Considering all possible combinations of input variables, the

scaled input planes esPI (k) − ∆esPI (k) (for incremental FPI con-
controllers. The input and output scaling factors are responsible
troller) and ∆esPD (k) − ∆2 esPD (k) (for incremental FPD controller)
for input normalization and output denormalization, respectively.
are divided into twenty regions and the regions are depicted in
From Fig. 2, it can be noticed that the components of incremental
Fig. 5.
FPI and incremental FPD controllers are the same and their details
From now onward the regions 13–20 are called the inner
are discussed in the following subsections.
regions. By utilizing the rules R1 to R3 , the control effort in
each region of the scaled input plane is obtained. The resultant
3.1. Fuzzification
outcomes (µ−1PI , µ0PI and µ+1PI ) of the antecedent parts of all
the three rules of the incremental FPI controller are provided in
The purpose of using fuzzification module is to convert scaled Table 2.
input variables into fuzzy variables. In this paper, L type and Γ
type of MFs are chosen on scaled input variables esPI (k), ∆esPI (k), 3.3. Inference mechanism
∆esPD (k), and ∆2 esPD (k) and L type, Π type and Γ type of MFs are
chosen on scaled output variables ∆usPI (k) and ∆usPD (k). The MFs Resultant outcomes of antecedent parts of control rules of the
used in this study are depicted in Figs. 3 and 4, respectively. incremental FPI and incremental FPD controllers are utilized to
−1 +1
The mathematical description of input MFs EPI and EPI is as modify the output MFs using a particular type of inference mech-
follows: anism. In this study, LP inference has been utilized for finding
⎧
⎨1
⎪ if esPI (k) ≤ −hePI the analytical structures of the incremental FPI and incremental
µE −1 (esPI (k)) =
−esPI (k)+hePI
if −hePI ≤ esPI (k) ≤ hePI FPD controllers. The modified MFs of output fuzzy sets using
2hePI
PI ⎪
⎩0 LP inference are depicted with dashed lines in Fig. 6 where the
if hePI ≤ esPI (k) values of µ−1PI , µ0PI , µ+1PI , µ−1PD , µ0PD and µ+1PD are provided
in Table 2.
⎧
⎨0
⎪ if esPI (k) ≤ −hePI 3.4. Defuzzification
esPI (k)+hePI
µE +1 (esPI (k)) = 2hePI
if −hePI ≤ esPI (k) ≤ hePI
PI
Defuzzification transforms aggregated fuzzy output to a crisp
⎪
⎩1 if hePI ≤ esPI (k)
value. Among different defuzzification methods, from the per-
Note that in a similar way the mathematical description of spective of control systems engineering, CoS strategy is the most
other input MFs can be obtained. The parameter API or APD of used one. For various defuzzification methods, readers may refer
the fuzzy set on the scaled output variable can take any value to [30]. Note that the Ref. [30] only covers the computational as-
in [0, BPI or BPD ). When API or APD = 0, the input–output pects of CoG defuzzification, not the analytical structures of fuzzy
relationship of the incremental FPI or incremental FPD controller logic controllers. Recently Arun and Mohan [27] have shown that
can be found with triangular MFs. CoG method is more accurate as compared to the CoS method.
In CoS method the overlapping region between two adjacent
3.2. Rule base MFs is considered twice, while it is considered only once in
CoG method. Because of computational complexity, CoG defuzzi-
Table 1 describes the rule base for the incremental FPI con- fication method remained almost dormant in modeling of fuzzy
troller. controllers. In this paper, the authors make an attempt to design
For most commonly encountered control problems these four a new nonlinear FPID controller by merging one incremental FPI
rules should be sufficient as there are only four different scenar- controller with another incremental FPD controller in parallel
ios: where the modeling of incremental FPI and incremental FPD
4
D. Sain and B.M. Mohan ISA Transactions xxx (xxxx) xxx

Fig. 5. Regions of scaled input planes of incremental FPI and incremental FPD controllers.

Fig. 6. Reference and inferred (via LP inference) MFs of fuzzy sets on scaled
output variables.

Fig. 7. Control effort variation of incremental FPI/FPD controller.

controllers is done using CoG defuzzification. From the inferred

MFs on output variables, the crisp values of the control efforts
produced by the incremental FPI and incremental FPD controllers
using CoG defuzzification can be obtained as
∫ (API +BPI )
−(API +BPI ) PI
µ ∆u (k) d∆u (k)
sPI sPI
∆usPI (k) =
∗
∫ (API +BPI ) (10)
−(API +BPI )
µPI d∆usPI (k)
∫ (APD +BPD )
µ ∆usPD (k) d∆usPD (k)
−(APD +BPD ) PD
∆u∗sPD (k) = ∫ (APD +BPD ) (11)
−(APD +BPD )
µPD d∆usPD (k)
Now the crisp value of the incremental control effort produced
by the nonlinear FPID controller can be obtained as

∆u∗sPID (k) = ∆u∗sPI (k) + ∆u∗sPD (k) (12) Fig. 8. Gain variation of incremental FPI/FPD controller in regions 13–20.

4. Mathematical models of the incremental FPI and FPD con-

1. The incremental FPI and incremental FPD controllers us-
trollers
ing Min t −norm, Max t −conorm, LP inference and CoG
defuzzification are nonlinear and variable structure con-
Using the information available in Table 2 and utilizing Eqs. trollers.
(10) and (11), the mathematical models of the incremental FPI 2. The controllers are model free controllers i.e. their struc-
and incremental FPD controllers are obtained. The analytical tures do not depend on the plant model.
structure of the incremental FPI controller is provided in Table 3.
Therefore, the proposed FPID controller is also a nonlinear,
For brevity the k term has been dropped from esPI (k), ∆esPI (k),
variable structure and model free controller.
∆esPD (k) and ∆2 esPD (k) in the analytical structures of the con-
trollers. In Table 3, by replacing esPI , ∆esPI , API , BPI , hePI and 5. Simulation study
h∆ePI with ∆esPD , ∆2 esPD , APD , BPD , h∆ePD and h∆2 ePD , respectively,
the mathematical model of the incremental FPD controller can be An unstable linear time-delay system is considered in simula-
found directly. Considering [−2, 2] as the UoD of the scaled input tion to demonstrate the applicability of the newly developed FPID
variables, the control surface of incremental FPI/FPD controller controller.
is provided in Fig. 7 where the controller parameter values are The controller parameters are obtained by minimizing the
following objective function
considered as: API = APD = 1, BPI = BPD = 2, hePI =
h∆ePI = h∆ePD = h∆2 ePD = 1. With the same parameters, in inner L−1
1∑
regions, the gain variation of the incremental FPI/FPD controller Jcost = {e2 + u2 } (13)
L
is depicted in Fig. 8. k=0

From the obtained models of FPI and FPD controllers it can be where e is system error and u is control signal at kth sam-
observed that pling moment, Tf is total time of simulation or experimentation
5
D. Sain and B.M. Mohan ISA Transactions xxx (xxxx) xxx

Table 2
Outcomes of antecedent parts of rules of incremental FPI controller.
Region 1, 3 2 4 5, 6 7, 8 9, 10 11, 12 13, 14 15, 16 17, 18 19, 20
µ−1PI 0 1 0 µE −1 µ∆E −1 0 0 µE − 1 µ∆E −1 µ∆E −1 µE −1
PI PI PI PI PI PI
µ0PI 1 0 0 µE +1 µ∆E +1 µE −1 µ∆E −1 µE + 1 µ∆E +1 µE − 1 µ∆E −1
PI PI PI PI PI PI PI PI
µ+1PI 0 0 1 0 0 µE +1 µ∆E +1 µ∆E +1 µE + 1 µE + 1 µ∆E +1
PI PI PI PI PI PI

Note: The resultant outcomes (µ−1PD , µ0PD and µ+1PD ) of the antecedent parts of all the three rules
of the incremental FPD controller can be obtained by replacing µE −1 , µE +1 , µ∆E −1 and µ∆E +1 with
PI PI PI PI
µ∆E −1 , µ∆E +1 , µ∆2 E −1 and µ∆2 E +1 , respectively.
PD PD PD PD

Table 3
Analytical structures of the incremental FPI controller.
Region/point ∆usPI
1, 3 0

2(A2PI + B2PI ) + 5API BPI

2,4 ∓
3(API + BPI )

{h2ePI (11A2PI + 20API BPI + 5B2PI ) ± 2esPI hePI (A2PI + API BPI − 2B2PI )
−e2sPI (BPI − API )2 }(∓hePI + esPI )
5, 6, 9, 10
6hePI {h2ePI (7API + 5BPI ) ± 2esPI hePI (BPI + API ) + e2sPI (BPI − API )}

{h2∆ePI (11A2PI + 20API BPI + 5B2PI ) ± 2∆esPI h∆ePI (A2PI + API BPI
−2B2PI ) − ∆e2sPI (BPI − API )2 }(∓h∆ePI + ∆esPI )
7, 8, 11, 12
6h∆ePI {h2∆ePI (7API + 5BPI ) ± 2∆esPI h∆ePI (BPI + API ) + ∆e2sPI
(BPI − API )}

[4∆e2sPI h4ePI (2A2PI + 5API BPI + 2B2PI ) − e2sPI h∆ePI {{esPI (∆esPI hePI +
esPI h∆ePI )(BPI − API )2 ± 3∆esPI h2ePI (B2PI − A2PI )} + hePI h∆ePI { ± esPI
(A2PI − 8API BPI + 7B2PI ) − hePI (17A2PI + 26API BPI − 7B2PI )}} + h3ePI
h∆ePI {∆esPI {esPI (25A2PI + 34API BPI + 13B2PI ) ± hePI (37A2PI + 76API
BPI + 31B2PI )} ± h∆ePI {esPI (49A2PI + 76API BPI + 19B2PI ) ± hePI (40A2PI
+76API BPI + 28B2PI )}}](∆esPI hePI + esPI h∆ePI )
13, 14, 17, 18 except points P and Q
±6h2ePI {2∆e3sPI h3ePI (BPI + API ) ± h∆ePI {∆esPI hePI {∆esPI hePI { ± 4esPI
(BPI + 2API ) + hePI (13BPI + 19API )} ± 2esPI h∆ePI {2esPI (2BPI + API )
±hePI (11BPI + 21API )}} ± e2sPI h2∆ePI {2esPI (3BPI − API ) ± hePI (17BPI +
15API )}}} + {6e2sPI hePI h∆ePI {e2sPI h2∆ePI + ∆esPI hePI (∆esPI hePI + 2esPI
h∆ePI )}}(BPI − API ) ± {12h4ePI h2∆ePI {5(∆esPI hePI + esPI h∆ePI ) ± 4hePI
h∆ePI }}(3BPI + 5API )

[4e2sPI h4∆ePI (2A2PI + 5API BPI + 2B2PI ) − ∆e2sPI hePI {{∆esPI (∆esPI hePI +
esPI h∆ePI )(BPI − API )2 ± 3esPI h2∆ePI (B2PI − A2PI )} + hePI h∆ePI { ± ∆esPI
(A2PI − 8API BPI + 7B2PI ) − h∆ePI (17A2PI + 26API BPI − 7B2PI )}} + hePI
h3∆ePI {esPI {∆esPI (25A2PI + 34API BPI + 13B2PI ) ± h∆ePI (37A2PI + 76API
BPI + 31B2PI )} ± hePI {∆esPI (49A2PI + 76API BPI + 19B2PI ) ± h∆ePI (40A2PI
+76API BPI + 28B2PI )}}](∆esPI hePI + esPI h∆ePI )
15, 16, 19, 20 except points P and Q
±6h2∆ePI {2e3sPI h3∆ePI (BPI + API ) ± hePI {esPI h∆ePI {esPI h∆ePI { ± 4∆esPI
(BPI + 2API ) + h∆ePI (13BPI + 19API )} ± 2∆esPI hePI {2∆esPI (2BPI + API )
±h∆ePI (11BPI + 21API )}} ± ∆e2sPI h2ePI {2∆esPI (3BPI − API ) ± h∆ePI (17BPI
+15API )}}} + {6∆e2sPI hePI h∆ePI {∆e2sPI h2ePI + esPI h∆ePI (esPI h∆ePI +
2∆esPI hePI )}}(BPI − API ) ± {12h2ePI h4∆ePI {5(∆esPI hePI + esPI h∆ePI )
±4hePI h∆ePI }}(3BPI + 5API )
2(A2PI + B2PI ) + 5API BPI
P, Q (∆esPI hePI + esPI h∆ePI )
6(API + BPI )hePI h∆ePI
Note: Upper and lower signs are used for regions 2, 5–8, and 13–16 and 4, 9–12, and 17–20, respectively. At point
P, esPI = −hePI and ∆esPI = −h∆ePI and at point Q , esPI = hePI and ∆esPI = h∆ePI .

T
and L = Tf . Note that for the proposed FPID controller any controller parameter values. Note that any other optimization
sampling time cannot be considered at the time of controller algorithm can also be employed but as the aim is to show the
implementation. Sampling time should be chosen depending on application of the proposed FPID controller, other algorithms are
the calculation time (Tc ) of a controller expression. For efficient not explored in this study. For detail explanation about GA, read-
control, Tc ≤ T . ers may refer to [31]. To find the proposed controller parameters,
The value of T is chosen as 0.001s. In this study, Genetic Algo- in the beginning a large search space is considered and in subse-
rithm (GA) is used as the optimization tool for finding the FPID quent iterations the search space is gradually narrowed down. To
6
D. Sain and B.M. Mohan ISA Transactions xxx (xxxx) xxx

Table 4
Result of performance comparison (Example 1).
Control scheme Mp (%) tr (s) ts (s) ue Jcost
PI [34] 111.09 27.6 844.1 223.2 0.1912
H∞ PI [34] 90.48 20.82 325.05 214.9 0.1492
PID [33] 77.24 13.26 162.15 3123 1.588
Fuzzy PID [6] 13.01 31.48 185.33 158.4 0.09939
Fuzzy PID [19] 12.61 57.09 162.11 146 0.09702
Proposed fuzzy PID 4.82 45.595 135.5 147.7 0.09283

Table 5
Result of performance comparison (Example 1, Delay = 24 s).
Control scheme Mp (%) tr (s) ts (s) ue Jcost
PI [34] 132.43 26.91 1085.25 283.3 0.2509
Fig. 9. Responses of the CL plant (Example 1). H∞ PI [34] 116.64 20.64 341.36 269 0.1893
PID [33] 110.72 13.255 247.6 3079 1.578
Fuzzy PID [6] 31.32 28.5 320.72 193.7 0.1176
Fuzzy PID [19] 22.46 51.41 302.2 176.7 0.114
Proposed fuzzy PID 16.35 40.02 247.33 177.5 0.1101

delay. The results obtained in presence of additional plant time-

delay are provided in Fig. 10 and Table 5. Note that for all the
controllers zero steady-state error is observed.
∑L−1 2
Note that in tables Mp , tr , ts , ue (= T k=0 u (k)), and Jcost
specify peak overshoot, rise time, settling time, control signal
energy and objective function value, respectively.
It can be noticed From Tables 4 and 5 that newly developed
FPID controller outperforms other controllers in terms of maxi-
mum overshoot, settling time and cost function. It is also noticed
Fig. 10. Responses of the CL plant (Example 1, Delay = 24 s). that in case of the proposed FPID controller rise time is more.
Though low rise time is obtained with PID controller in [33], large
maximum overshoot and high control energy are the problems
keep the optimization procedure simple, some assumptions are associated with this controller. Low control energy is observed
−1 −1
made which are as follows: API = APD = 0 and SPI = SPD = 1. with FPID controller in [19], but in comparison to the proposed
FPID controller, comparatively large maximum overshoot, rise
Example 1. In this example, the problem of concentration control time, settling time and cost function are the problems associated
in an unstable chemical reactor is discussed where Cinput (t) and with this controller. Based on the above discussion and consid-
Coutput (t) respectively represent the concentration at plant input ering the time-domain attributes in Tables 4 and 5, it can be
and plant output. A time-delay of 20s is considered as concen- concluded that the newly developed FPID controller controls the
tration transducer requires 20s to produce the process output system in Eq. (14) in a more desirable way in comparison to other
variable Coutput (t). The transfer function [32] of the plant is as controllers.
follows
6. Experimental validation
Coutput (s) 3.433e−20s
Gp1 (s) = = (14)
Cinput (s) 103.1s − 1 A magnetic levitation system (manufacturer: Feedback Instru-
ments, model no. 33–210) is considered in this section to experi-
From the above transfer function, it can be observed that there
mentally validate the FPID controller developed in this study. The
is a pole in the right half of s-plane which causes instability.
control of this system is a onerous task for the control practi-
Moreover, due to the presence of significant amount of delay,
tioners due to its nonlinear unstable behavior. Various control
system performance further degrades. To obtain satisfactory sys- methods such as 1 and 2-DoF PID control, fractional order PID
tem performance, the proposed FPID controller is employed. By control etc. have been applied over the years for the control
optimizing the cost function in Eq. (13) using GA, FPID con- of magnetic levitation system [35,36]. The differential equation
troller parameter values are found and summarized as follows: which governs the dynamics of the magnetic levitation system is
BPI = 2.9560, hePI = 0.7545, h∆ePI = 1, SePI = 0.0030 × given by
10−3 , S∆ePI = 1.5598, BPD = 0.8976, h∆ePD = 4.9700, h∆2 ePD =
2
4.9659, S∆ePD = 0.0103 and S∆2 ePD = 0.5369. mb ẍb = mb g − kc i2c x−
b (15)
The simulation is carried out for 2000s and the step responses where mb and xb denote the mass and position of the ball, respec-
of the system due to the proposed FPID and existing PI and PID tively, g is gravitational acceleration, ic is the coil current and kc
controllers are shown in Fig. 9. Moreover, during simulation, a is a constant which depends on coil parameters.
disturbance signal (do (t) = 0 for t < 1000s and do (t) = 0.2 mol/l The schematic layout and experimental set up of the magnetic
for 1000s ≤ t ≤ 2000s) is applied at the plant output to assess levitation system are provided in Figs. 11 and 12, respectively.
the disturbance rejection ability of the newly developed FPID con- Note that details of magnetic levitation system parameters can
troller. System performance in presence of different controllers be found in [35].
are compared and the result of comparison is provided in Table 4. Similarly like simulation example, the value of T is chosen as
Next, the plant time-delay is increased by 20% (i.e from 20s to 0.001s and the optimization is performed using GA. After mini-
24s) to check how sensitive the control system is to the change in mizing the objective function (Eq. (13)), the controller parameters
7
D. Sain and B.M. Mohan ISA Transactions xxx (xxxx) xxx

Table 6
Result of performance comparison (real-time).
Control scheme IAE ITAE ISE ITSE ue Jcost
Fuzzy PID 0.2756 0.3852 0.08207 0.01756 1.406 0.1488
Linear PID 0.6046 1.138 0.1531 0.08225 4.397 0.4551

Fig. 11. Schematic layout of the magnetic levitation system.

Fig. 15. Responses of CL magnetic levitation system with input disturbance

(real-time).
Fig. 12. Hardware set up of the magnetic levitation system.

Fig. 16. Responses of CL magnetic levitation system with output disturbance

(real-time).
Fig. 13. Responses of CL magnetic levitation system (simulation).

The performances of the linear and fuzzy PID controllers are

compared and the result of comparison is recorded in Table 6.

By observing Figs. 13–14 and Table 6, it can be concluded

that the newly obtained FPID controller has the ability to provide
better system performance with less control effort as compared to
the linear PID controller. The responses of the magnetic levitation
system are also obtained with input and output disturbances
and the corresponding results are provided in Figs. 15–16. Note
that input disturbance di (t) = 0.31V for 0 ≤ t ≤ 10s and
output disturbance is applied with hand between 16s and 17s.
At the time of checking the output disturbance rejection ability
Fig. 14. Responses of CL magnetic levitation system (real-time). of the newly obtained controller, the hardware experimentation
is continued for 24s and the system response is shown from 14s
to 24s only as before applying disturbance the system performs
are obtained as BPI = 1.0484, hePI = 0.1207, h∆ePI = 0.1205, normally. It can further be observed from Figs. 15–16 that also in
SePI = 0.0050, S∆ePI = 1.0943, BPD = 1.0911, h∆ePD = 0.1240, presence of disturbance, the proposed FPID controller performs
h∆2 ePD = 0.1375, S∆ePD = 1.0342, and S∆2 ePD = 11.4448. With the in a better way in comparison to the linear PID controller.
intention of comparison, DT linear PID controller is also designed
whose parameters are KPdt = 3.6988, KIdt = 2.5 × 10−3 and
7. Conclusion
KDdt = 65.8.
A sinusoidal reference is considered in this section whose
details are as follows: amplitude=0.3 V, frequency=0.4 Hz and In this paper, the exact model of a nonlinear FPID controller
bias=-1.5 V. The simulation and hardware experimentation are is unveiled using one incremental FPI controller and another
carried out for Tf = 10s and the corresponding responses are incremental FPD controller in parallel. The incremental FPI and
provided in Figs. 13 and 14. incremental FPD controller structures are unveiled using Min
As the ball position is converted into an equivalent voltage, t-norm, Max t-conorm, LP inference, CoG defuzzification, and
in graphs the position of the ball is shown in terms of volts. two-dimensional input space. The applicability and usefulness of
8
D. Sain and B.M. Mohan ISA Transactions xxx (xxxx) xxx

the newly obtained FPID controller are depicted with one simu- [15] Karasakal O, Guzelkaya M, Eksin I, Yesil E. An error-based on-line rule
lation example and one real-time experimentation. An unstable weight adjustment method for fuzzy PID controllers. Expert Syst Appl
2011;38(8):10124–32.
linear time-delay system is controlled in simulation, whereas an
[16] Das S, Pan I, Das S, Gupta A. A novel fractional order fuzzy PID controller
unstable and nonlinear magnetic levitation system is considered and its optimal time domain tuning based on integral performance indices.
during experimental realization. For better understanding, perfor- Eng Appl Artif Intell 2012;25(2):430–42.
mances of the proposed and other controllers are compared. It [17] Khooban MH, Alfi A, Abadi DNM. Control of a class of non-linear uncer-
is hoped that for accurate trajectory tracking and enhancing ex- tain chaotic systems via an optimal Type-2 fuzzy proportional integral
derivative controller. IET Sci Meas Technol 2013;7(1):50–8.
isting system performance, this newly developed FPID controller [18] Karasakal O, Guzelkaya M, Eksin I, Yesil E, Kumbasar T. Online tuning of
will be helpful in future. fuzzy PID controllers via rule weighing based on normalized acceleration.
Eng Appl Artif Intell 2013;26(1):184–97.
Declaration of competing interest [19] Kumar V, Mittal AP, Singh R. Stability analysis of parallel fuzzy P + fuzzy
I + fuzzy D control systems. Int J Autom Comput 2013;10(2):91–8.
[20] Savran A. A multivariable predictive fuzzy PID control system. Appl Soft
The authors declare that they have no known competing finan- Comput 2013;13(5):2658–67.
cial interests or personal relationships that could have appeared [21] Savran A, Kahraman G. A fuzzy model based adaptive PID controller design
to influence the work reported in this paper. for nonlinear and uncertain processes. ISA Trans 2014;53(2):280–8.
[22] El-Nagar AM, El-Bardini M. Derivation and stability analysis of the analyti-
cal structures of the interval type-2 fuzzy PID controller. Appl Soft Comput
References 2014;24:704–16.
[23] Fereidouni A, Masoum MAS, Moghbel M. A new adaptive configuration of
[1] Mizumoto M. Realization of PID controls by fuzzy control methods. Fuzzy PID type fuzzy logic controller. ISA Trans 2015;56:222–40.
Sets and Systems 1995;70(2–3):171–82. [24] Gil P, Lucena C, Cardoso A, Palma LB. Gain tuning of fuzzy PID controllers
[2] Misir D, Malki HA, Chen G. Design and analysis of a fuzzy for MIMO systems: a performance-driven approach. IEEE Trans Fuzzy Syst
proportional–integral–derivative controller. Fuzzy Sets and Systems 2015;23(4):757–68.
1996;79(3):297–314. [25] Arun NK, Mohan BM. Modelling, stability analysis and computational
[3] Homaifar A, Sayyarrodsari B, Nagle J, Bikdash M. Hierarchical learning- aspects of nonlinear fuzzy PID controllers using Mamdani minimum
based design of a hybrid fuzzy PID controller. Intell Autom Soft Comput inference. Int J Autom Control 2018;12(1):153–74.
1997;3(1):23–36. [26] Sain D, Mohan BM. Modelling of a nonlinear fuzzy three-input PID
[4] Mann GKI, Hu BG, Gosine RG. Analysis of direct action fuzzy PID controller controller and its simulation and experimental realization. IETE Tech Rev
structures. IEEE Trans Syst Man Cybern B 1999;29(3):371–88. 2020. http://dx.doi.org/10.1080/02564602.2020.1773326.
[5] Hayashi K, Otsubo A, Murakami S, Maeda M. Realization of nonlinear and [27] Arun NK, Mohan BM. Modeling, stability analysis, and computational
linear PID controls using simplified indirect fuzzy inference method. Fuzzy aspects of some simplest nonlinear fuzzy two-term controllers derived via
Sets and Systems 1999;105(3):409–14. center of area/gravity defuzzification. ISA Trans 2017;70:16–29.
[6] Kim JH, Oh SJ. A fuzzy PID controller for nonlinear and uncertain systems. [28] Sain D, Mohan BM. Simulation and real-time implementation of a
Soft Comput 2000;4(2):123–9. nonlinear fuzzy PI/PD controller. IFAC-PapersOnLine 2020;53(1):673–8.
[7] Tao CW, Taur JS. Flexible complexity reduced PID-like fuzzy controllers. [29] Ying H. Fuzzy control and modeling: analytical foundations and
IEEE Trans Syst Man Cybern B 2000;30(4):510–6. applications. New York: IEEE Press; 2000.
[8] Golob M. Decomposed fuzzy proportional–integral–derivative controllers. [30] Driankov D, Hellendoorn H, Reinfrank M. An introduction to fuzzy control.
Appl Soft Comput 2001;1(3):201–14. New Delhi: Narosa Publishing House; 1993.
[9] Lu J, Chen G, Ying H. Predictive fuzzy PID control: theory, design and [31] Goldberg DE. Genetic algorithms in search, optimization and machine
simulation. Inform Sci 2001;137(1–4):157–87. learning. India: Pearson Education; 1989.
[10] Hu BG, Mann GKI, Gosine RG. A systematic study of fuzzy PID [32] Rico JEN, Camacho EF. Unified approach for robust dead-time compensator
controllers-function-based evaluation approach. IEEE Trans Fuzzy Syst design. J Process Control 2009;19(1):38–47.
2001;9(5):699–712. [33] Onat C. A new design method for PI–PD control of unstable processes with
[11] Haj-Ali A, Ying H. Structural analysis of fuzzy controllers with nonlinear dead time. ISA Trans 2019;84:69–81.
input fuzzy sets in relation to nonlinear PID control with variable gains. [34] Shariati A, Taghirad HD, Fatehi A. A neutral system approach to H∞ PD/PI
Automatica 2004;40(9):1551–9. controller design of processes with uncertain input delay. J Process Control
[12] Mann GKI, Gosine RG. Three-dimensional min–max-gravity based 2014;24(3):144–57.
fuzzy PID inference analysis and tuning. Fuzzy Sets and Systems [35] Ghosh A, Krishnan TR, Tejaswy P, Mandal A, Pradhan JK, Ranasingh S.
2005;156(2):300–23. Design and implementation of a 2-DOF PID compensation for magnetic
[13] Mohan BM, Sinha A. The simplest fuzzy PID controllers: mathematical levitation systems. ISA Trans 2014;53(4):1216–22.
models and stability analysis. Soft Comput 2006;10(10):961–75. [36] Sain D, Swain SK, Kumar T, Mishra SK. Robust 2-DOF FOPID controller
[14] Mohan BM, sinha A. Analytical structures for fuzzy PID controllers? IEEE design for maglev system using Jaya algorithm. IETE J Res 2018. http:
Trans Fuzzy Syst 2008;16(1):52–60. //dx.doi.org/10.1080/03772063.2018.1496800.