Вы находитесь на странице: 1из 10

Analytical structure and stability analysis of a fuzzy PID controller

B.M. Mohan
*
, Arpita Sinha
Department of Electrical Engineering, Indian Institute of Technology, Kharagpur 721302, India
Received 14 July 2006; received in revised form 31 May 2007; accepted 10 June 2007
Available online 14 June 2007
Abstract
Analytical structure for a fuzzy PIDcontroller is introduced by employing two fuzzy sets for each of the three input variables and four fuzzy sets
for the output variable. This structure is derived via left and right trapezoidal membership functions for inputs, trapezoidal membership functions
for output, algebraic product triangular norm, bounded sum triangular co-norm, Mamdani minimum inference method, and center of sums (COS)
defuzzication method. Conditions for bounded-input bounded-output (BIBO) stability are derived using the Small Gain Theorem. Finally, two
numerical examples along with their simulation results are included to demonstrate the effectiveness of the simplest fuzzy PID controller.
# 2007 Elsevier B.V. All rights reserved.
Keywords: Fuzzy PID controllers; Analytical structure; BIBO stability
1. Introduction
Most of the industrial processes are still the conventional
PID controllers due to their simple and robust design,
affordable price, and effectiveness for linear systems [3,10].
So far two different congurations have been reported for PID
control as shown in Fig. 1. Due to their linear structure,
conventional PID controllers are usually not effective if the
processes involved are higher order and time delay systems,
nonlinear systems, complex and vague systems without precise
mathematical models, and systems with uncertainties [3,8,10].
Till 1990 most of the research activity in fuzzy control systems
area dealt with applications rather than theory. Around
1990 people began developing analytical structures for fuzzy
controllers and analyzing the same, with an objective of
building fuzzy control theory so that it can be applied in the
similar lines of conventional control theory [21]. It has been
observed that fuzzy PI [15,21] and fuzzy PD [11,14] controllers
can handle the above mentioned systems better than their
conventional counterparts. Fuzzy PI controllers are preferred
more to fuzzy PD controllers as fuzzy PD controllers are not
able to eliminate steady state errors [16]. However, fuzzy PI
controllers show poor performance during the transient phase
for higher order processes due to their internal integration
operation. To obtain overall improved performance, fuzzy PID
controllers are preferred [8,16].
A hybrid fuzzy logic proportional plus conventional
integral-derivative controller in incremental form has been
presented [10]. This controller is obtained by considering an
incremental fuzzy logic controller in place of the proportional
term in a conventional PID controller. Also a sufcient
condition for BIBO stability of this controller is derived using
the Small Gain Theorem. As a matter of fact, stability analysis
of fuzzy control systems has been extensively discussed in [1].
This book presents several techniques for stability analysis of
both Mamdani type and Takagi-Sugeno type fuzzy models. An
in-depth treatment on analysis and design of fuzzy control
systems via a linear matrix inequality approach has been given
in [17] to deal with systems described by Takagi-Sugeno type
fuzzy models.
Anewfuzzy PIDcontroller structure, based on conguration
1 [16] in Fig. 1, has been proposed. To tune the parameters of
the fuzzy controller on line, a parameter adaptive method via
peak observer has been presented. A two-level tuning strategy
has been introduced [9] which rst tries to set up the
relationship between fuzzy proportional/integral/derivative
gain and scaling gains at the high level, and then optimally
tunes the control resolution at low level. Based on the
conguration 1, a fuzzy PID controller has been suggested [7]
using minimum inference engine and center average defuzzi-
cation, which behaves approximately like a parameter varying
PID controller. In order to improve further the performance in
www.elsevier.com/locate/asoc
Applied Soft Computing 8 (2008) 749758
* Corresponding author. Fax: +91 3222 282262/255303.
E-mail address: mohan@ee.iitkgp.ernet.in (B.M. Mohan).
1568-4946/$ see front matter # 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2007.06.003
transient and steady states, an adaptive method via function
tuner has been developed [19] to tune the scaling factors of the
fuzzy controller on line. An adaptive method via relative rate
observer has been proposed [5] for tuning the scaling factors of
the fuzzy logic controller in an on-line manner.
Fuzzy PI and fuzzy PD controllers are combined to get a
fuzzy PID controller according to the conguration 2 [8] in
Fig. 1. Its knowledge base consists of two-dimensional rule
bases for PI and PD controls. A tuning method, based on gain
margin and phase margin specications, has been proposed [20]
for determining the parameters of the fuzzy PID controller.
Also sufcient conditions for BIBO stability have been
determined. Several forms of decomposed proportional-
integral-derivative fuzzy logic controllers (fuzzy P fuzzy
I fuzzy D form, fuzzy PD fuzzy I form, fuzzy
PI conventional D form, fuzzy P conventional ID form
and fuzzy PI fuzzy PD form) have been tested and compared
[4]. To obtain simple structures, the activities of the
proportional, integral and derivative parts of the fuzzy PID
controller are dened with simple rules in proportional rule
base, integral rule base and derivative rule base.
A function-based evaluation approach has been proposed
[6] for a systematic study of fuzzy PID controllers, and the
fuzzy controllers have been analyzed using ve simple
evaluation criteria (control-action composition, input cou-
pling, gain dependency, gain-role change, and rule/parameter
growth).
It has been already proved by Mizumoto [13] that PID
controllers can be obtained by using fuzzy control methods like
product-sum-gravity method (algebraic product triangular
norm, Larsen product inference method, bounded sum
triangular co-norm and center of gravity method for
defuzzication) and simplied fuzzy reasoning method (special
case of product-sum-gravity method when fuzzy sets in
consequent part are of same size). However, PID controller
cannot be constructed by min-max-gravity method as this
method gives a complicated inference result of nonlinear form
for a simple fuzzy reasoning form, see Fig. 5 in [13]. It has been
shown that extrapolative reasoning can also be accomplished by
product-sum-gravity method and simplied fuzzy reasoning
method by extending membership functions of antecedent parts
of fuzzy rules [13].
In this paper attempts are made to obtain analytical structure
of the fuzzy PID controller (conguration 3 in Fig. 1) by
employing algebraic product triangular norm, bounded sum
triangular co-norm, left (G-type) and right (L-type) trapezoidal
membership functions for inputs, trapezoidal membership
functions for output, nonlinear control rules, Mamdani
minimum inference method, and COS method of defuzzica-
tion. Conditions for BIBO stability of fuzzy PID control
systems are obtained. Finally, to demonstrate the superiority of
fuzzy PID controller over the conventional PID controller,
simulation results of two examples are included.
The paper is organized as follows. The next section deals
with the fundamental components of a typical fuzzy PID
controller. Section 3 presents the fuzzy PIDanalytical structure.
Section 4 is about BIBO stability analysis of fuzzy PID control
systems. Section 5 deals with design aspects of the simplest
fuzzy PID controller. Section 6 includes simulation results
while Section 7 consists of concluding remarks.
Fig. 1. PID controllers: (a) conguration 1; (b) conguration 2; (c) conguration 3.
B.M. Mohan, A. Sinha / Applied Soft Computing 8 (2008) 749758 750
2. Components of fuzzy PID controllers
The incremental control signal generated by a discrete-time
PID controller is given by
Du(kT) = u(kT) u[(k 1)T[
= K
d
P
v(kT) K
d
I
d(kT) K
d
D
a(kT) (1)
where K
d
P
, K
d
I
, and K
d
D
are, respectively the proportional,
integral and derivative constants of digital PID controller,
the velocity v(kT), displacement d(kT) and acceleration
a(kT) are given by
v(kT) =
d(kT) d[(k 1)T[
T
(2)
d(kT) = e(kT) (3)
a(kT) =
v(kT) v[(k 1)T[
T
; (4)
e(kT) is the error signal given by r(kT) y(kT), r(kT) the
reference signal, y(kT) the process output, and Tis the sampling
period. The time derivatives of displacement d(t) and velocity
v(t) at t = kT are obtained by using backward-difference rule
approximation of d(t) and v(t), as shown in Eqs. (2) and (4).
Eq. (1) is known as velocity algorithm and it is a widely used
form of digital PID control. The principal structure of a fuzzy
PID controller (see Fig. 2), based on conguration 3 of Fig. 1,
consists of the following components.
2.1. Scaling factors
An input variable to a fuzzy logic controller can be mapped
onto the interval [l; l[ by multiplying it by a normalization
factor. N
d
, N
v
, N
a
and N
Du
are the normalization factors for d, v,
a and Du, respectively. Denormalization maps the normalized
output value into its physical output domain. N
1
Du
is the
reciprocal of N
Du
, called denormalization factor. These scaling
factors play a role similar to that of the gain coefcients K
d
P
, K
d
I
and K
d
D
in a conventional PID controller.
2.2. Fuzzication
Fuzzication converts crisp values of controller inputs into
fuzzy sets that can be used by the inference engine (refer to
Section 2.4) to activate and apply the control rules. The fuzzy
PID controller employs three inputs: the error signal e(kT) (i.e.,
displacement d(kT)), the rst-order time derivative of e(kT)
(i.e., velocity v(kT)), and the second-order time derivative of
e(kT) (i.e., acceleration a(kT)). These inputs are fuzzied by a
combination of L-type and G-type membership functions as
illustrated in Fig. 3 where d
N
, v
N
and a
N
are the normalized
inputs. The mathematical description of L-type and G-type
membership functions is, respectively given by
m
nx
=
1 L _ x _ l
x l
2l
l _ x _ l
0 l _ x _ L
_

_
(5)
m
px
=
0 L _ x _ l
x l
2l
l _ x _ l
1 l _ x _ L
_

_
(6)
Note that
m
nx
m
px
= 1 (7)
Fig. 2. Fuzzy PID control system.
B.M. Mohan, A. Sinha / Applied Soft Computing 8 (2008) 749758 751
The fuzzy controller has a single output, called incremental
control output Du(kT). The membership functions for the
normalized output Du
N
are shown in Fig. 4. The constants l,
L, m and M are chosen by the designer.
2.3. Control rule base
The following control rules are considered [12] in terms of
the input and output fuzzy sets shown in Figs. 3 and 4,
respectively:
v R1: If d
N
= n d& v
N
= n v & a
N
= n a then Du
N
= O
2
v R2: If d
N
= p d & v
N
= n v & a
N
= n a then Du
N
= O
1
v R3: If d
N
= p d & v
N
= n v &a
N
= p a then Du
N
= O
1
v R4: If d
N
= n d & v
N
= n v & a
N
= p a then Du
N
= O
1
v R5: If d
N
= n d &v
N
= p v &a
N
= p a then Du
N
= O
1
v R6: If d
N
= n d & v
N
= p v & a
N
= n a then Du
N
= O
1
v R7: If d
N
= p d & v
N
= p v & a
N
= n a then Du
N
= O
1
v R8: If d
N
= p d &v
N
= p v &a
N
= p a then Du
N
= O
2
where the & symbol in the antecedent part represents the
fuzzy AND operation which is considered here as
algebraic product triangular norm, and is dened as
m(d
N
; v
N
; a
N
) = m
i
(d
N
) m
j
(v
N
) m
k
(a
N
) (8)
where mrepresents the combined membership function of the
rule-antecedent part and i, j and k are the ith, jth and kth fuzzy
sets on d
N
, v
N
and a
N
, respectively. Notice that the control
rules are nonlinear as the output fuzzy sets are not linearly
related to the input fuzzy sets.
2.4. Inference engine
Overall value of the incremental control output variable is
computed by the inference engine by considering the individual
contribution of each rule in the rule base. For this,
corresponding to each rule, rst the degree of match is found
by using the algebraic product triangular norm in Eq. (8). Then
the degree of match is used to determine the inferred output
fuzzy set using Mamdani minimum inference method which is
dened as
min ( m; m(Du
N
))
where m is the outcome of algebraic product triangular norm
operation and m(Du
N
) is the membership function of the output
fuzzy set Du
N
. The reference output fuzzy set (trapezoidal),
and the inferred output fuzzy set (shown with hatching) corre-
sponding to Mamdani minimum inference are shown in
Fig. 5. The area of inferred fuzzy set is given by
m(2 m mu)((2M 3m)=3), where u is the ratio of upper
side width (2m) and base width 2((2M 3m)=3) of the tra-
pezoid, and is given by
u =
3m
2M 3m
See Appendix A for the details. Here we consider 0 _ u <1 to
avoid overlap between upper sides of two adjacent output fuzzy
sets.
As there are three inputs to the fuzzy PID controller, it is
necessary to consider all possible combinations of these
variables in a 3D space. A point, say (x
1
, y
1
, z
1
), in a 3D space
can always be distinctly shown by taking its projection on the
xy-, yz-, and zx-planes. So, as shown in Fig. 6, 20 input
combinations are considered in each (d
N
v
N
, d
N
a
N
, and
v
N
a
N
) plane so that the state point (d
+
N
, v
+
N
, a
+
N
) can be
uniquely located in the 3D cell (subspace) represented by the
triplet (n
I
; n
II
; n
III
) where n
I
; n
II
; n
III
= 1; 2; . . . ; 20. For exam-
ple, the triplet (10, 11, 18) represents the 3Dcell with 10 from I,
11 from II, and 18 from III of Fig. 6.
The control rules (R1)(R8) of the fuzzy PID controller
are used to evaluate appropriate control law in each valid cell
(n
I
; n
II
; n
III
). By using the algebraic product triangular norm
the outcome of premise part of each rule is found for all valid
cells and is shown in Table 1. There are altogether 20
20 20 = 8000 cells in the 3D input space. Not all 8000
cells are valid cells; only a few of them are valid. A cell
(n
I
; n
II
; n
III
) is said to be valid if and only if the relations
between d
N
and v
N
, and d
N
and a
N
produce the relation
Fig. 3. Input membership functions.
Fig. 4. Output membership functions.
Fig. 5. Illustration of Mamdani minimum inference.
B.M. Mohan, A. Sinha / Applied Soft Computing 8 (2008) 749758 752
between v
N
and a
N
. For example, the cell (7, 2, 6) is a valid
cell because the relations d
N
_ v
N
and d
N
_a
N
produce the
relation v
N
_d
N
_a
N
which is satised by the relation
v
N
_a
N
.
It may be seen from the control rules that the output fuzzy
sets O
1
and O
1
are red three times each. In such a situation,
a fuzzy triangular co-norm is used [15] to evaluate combined
output fuzzy sets corresponding to the rule sets
(R2); (R4); (R6) and (R3); (R5); (R7). The bounded
sum triangular co-norm used here is dened as
min 1; m
A
(x) m
B
(y)
Since the fuzzy controller has three inputs and algebraic
product triangular norm is used, sum of all the outcomes
corresponding to either rule set is less than unity. Therefore the
combined memberships using bounded sum triangular co-norm
are given by
m(R2)m(R4)m(R6) <1 and m(R3)m(R5)m(R7) <1
2.5. Defuzzication
Defuzzication module converts fuzzy information into crisp
information. The most commonlyusedCOSmethodis employed
to defuzzify the incremental control output. This is expressed as
Du
N
(kT)
=
A(m(R1))(M) [A(m(R2)) A(m(R4))
A(m(R6))[(M=3) [A(m(R3)) A(m(R5))
A(m(R7))[(M=3) A(m(R8))(M)

8
i=1
A(m(Ri))
(9)
Fig. 6. Regions of the fuzzy PID controller input combinations.
Table 1
Outcomes of algebraic product operation of premise part of control rules (R1)(R8) for valid 3D cells
Cells (R1) (R2) (R3) (R4) (R5) (R6) (R7) (R8)
(1, 1, 1) to (8, 8, 8)
a
m
1
m
2
m
3
m
4
m
5
m
6
m
7
m
8
(9, 17, 9) to (16, 17, 16) 0 m
nv
0 0 0 0 m
pv
0
(9, 18, 12) to (16, 18, 13) 0 0 m
nv
0 0 0 0 m
pv
(10, 11, 18) to (11, 12, 18) 0 0 m
pd
m
nd
0 0 0 0
(10, 16, 17) to (11, 15, 17) m
nd
m
pd
0 0 0 0 0 0
(12, 19, 12) to (13, 19, 13) 0 0 0 m
nv
m
pv
0 0 0
(12, 20, 9) to (13, 20, 16) m
nv
0 0 0 0 m
pv
0 0
(14, 12, 19) to (15, 11, 19) 0 0 0 0 m
nd
0 0 m
pd
(14, 15, 20) to (15, 16, 20) 0 0 0 0 0 m
nd
m
pd
0
(17, 9, 10) to (17, 10, 11) 0 m
na
m
pa
0 0 0 0 0
(17, 17, 17) 0 1 0 0 0 0 0 0
(17, 18, 18) 0 0 1 0 0 0 0 0
(18, 13, 11) to (18, 14, 10) m
na
0 0 m
pa
0 0 0 0
(18, 19, 18) 0 0 0 1 0 0 0 0
(18, 20, 17) 1 0 0 0 0 0 0 0
(19, 13, 14) to (19, 14, 15) 0 0 0 0 m
pa
m
na
0 0
(19, 19, 19) 0 0 0 0 1 0 0 0
(19, 20, 20) 0 0 0 0 0 1 0 0
(20, 9, 15) to (20, 10, 14) 0 0 0 0 0 0 m
na
m
pa
(20, 17, 20) 0 0 0 0 0 0 1 0
(20, 18, 19) 0 0 0 0 0 0 0 1
m
1
= m
nd
m
nv
m
na
, m
2
= m
pd
m
nv
m
na
, m
3
= m
pd
m
nv
m
pa
, m
4
= m
nd
m
nv
m
pa
, m
5
= m
nd
m
pv
m
pa
, m
6
= m
nd
m
pv
m
na
, m
7
= m
pd
m
pv
m
na
, and m
8
= m
pd
m
pv
m
pa
.
a
Only valid cells in (1, 1, 1) through (8, 8, 8).
B.M. Mohan, A. Sinha / Applied Soft Computing 8 (2008) 749758 753
where A(m(Ri)) is the area of the inferred fuzzy set corre-
sponding to rule Ri.
3. Analytical structure
In this section, we obtain analytical structure of a fuzzy PID
controller using algebraic product triangular norm, bounded
sum triangular co-norm, Mamdani minimum inference method,
L-type and G-type input fuzzy sets and trapezoidal output fuzzy
sets. The analytical structures in the following three cases are
derived by substituting the inferred areas A(m()), correspond-
ing to all eight rules, into Eq. (9), grouping various terms
according to the normalized inputs d
N
, v
N
, and a
N
, and
simplifying. In the following the sampling time kT is shown
as k for simplicity. The analytical structure is:
Case (a). l _ d
N
(k), v
N
(k), a
N
(k) _ l
Du(k) =
_
2M
3N
Du
_
N
1
v
N
(k) N
2
d
N
(k) N
3
a
N
(k)
D
(10)
where
N
1
= (7 u)l
5
(1 u)[(a
2
N
(k) d
2
N
(k))l
3
a
2
N
(k)d
2
N
(k)l[
(11)
N
2
= (7 u)l
5
(1 u)[(v
2
N
(k) a
2
N
(k))l
3
v
2
N
(k)a
2
N
(k)l[
(12)
N
3
= (7 u)l
5
(1 u)[(d
2
N
(k) v
2
N
(k))l
3
d
2
N
(k)v
2
N
(k)l[
(13)
and
D = (15 u)l
6
(1 u)(d
2
N
(k) v
2
N
(k) a
2
N
(k))l
4
(d
2
N
(k)v
2
N
(k) v
2
N
(k)a
2
N
(k) a
2
N
(k)d
2
N
(k))l
2
d
2
N
(k)v
2
N
(k)a
2
N
(k) (14)
in which u is as dened in Section 2.4, and l is a parameter
shown in Fig. 3.
Case (b). One normalized input is in the interval [l; l[ and
the remaining two normalized inputs are not in the interval
[l; l[, see Fig. 3. The Du(k) in different cells is as follows:
Du(k) =
_
2lM
3N
Du
_
(1 u)x
(3 u)l
2
(1 u)x
2
for cells in Table 2 (15)
Du(k) =
_
2M
3N
Du
_
(3 u)l
2
(1 u)lx (1 u)x
2
(3 u)l
2
(1 u)x
2
for cells in Table 2 (16)
Du(k) =
_
2M
3N
Du
_
(3 u)l
2
(1 u)lx (1 u)x
2
(3 u)l
2
(1 u)x
2
for cells in Table 2 (17)
with x as dened in Table 2.
Case (c). All the three normalized inputs d
N
(k), v
N
(k) and
a
N
(k) are not in the interval [l; l[, see Fig. 3.
Du(k) =
_
M
3N
Du
_
for cells (17; 17; 17);
(18; 19; 18); (19; 20; 20) (18)
Du(k) =
_
M
3N
Du
_
for cells (17; 18; 18); (19; 19; 19);
(20; 17; 20) (19)
Du(k) =
_
M
N
Du
_
for cell
(18; 20; 17) (20)
Du(k) =
_
M
N
Du
_
for cell (20; 18; 19) (21)
It is to be noted here that the maximum and minimum values of
Du
N
(k) always remain at M and M, respectively for any
value of u in 0 _ u <1. This can be explained in simple terms in
the following manner: Eq. (20) is applicable for the case d
N
, v
N
,
a
N
[L; l[. As long as d
N
, v
N
, a
N
are in the interval [L; l[
the value of Du is always (M=N
Du
) only. In other words, the
value of Du
N
is (M) only. Similarly, Eq. (21) is applicable for
the case d
N
, v
N
, a
N
[l; L[. According to this equation, the value
of Du
N
is always (M). Interestingly, Eqs. (20) and (21) are
independent of u. As far as maxima and minima are concerned,
Eqs. (18) and (19) can be ruled out as the value of Du in Eq. (18)
is one-third of the value of Du in Eq. (20), and it is true even
with Eqs. (19) and (21). So, if M and Mare, respectively the
maximum and minimum values of Du
N
, we must be able to get
this same maxima and minima information even fromEqs. (10),
(15), (16) and (17) which are expressed in terms of u. By means
of rigorous simulation study, it is observed that for any value of
u in [0; 1) the maximum value (M) and the minimum value
(M) of Du
N
are found to occur only at (l; l; l) and (l; l; l),
respectively.
4. BIBO stability of fuzzy PID control systems
In this section BIBO stability analysis of the fuzzy PID
control system, shown in Fig. 7, is done using the Small Gain
Theorem [18].
Table 2
Attributes of x
x Eq. (15) with cells Eq. (16) with cells Eq. (17) with cells
d
N
(k) (10, 11, 18), (11, 12, 18), (14, 15, 20), (15, 16, 20) (14, 12, 19), (15, 11, 19) (10, 16, 17), (11, 15, 17)
v
N
(k) (9, 17, 9), (16, 17, 16), (12, 19, 12), (13, 19, 13) (9, 18, 12), (16, 18, 13) (12, 20, 9), (13, 20, 16)
a
N
(k) (17, 9, 10), (17, 10, 11), (19, 13, 14), (19, 14, 15) (20, 9, 15), (20, 10, 14) (18, 13, 11), (18, 14, 10)
B.M. Mohan, A. Sinha / Applied Soft Computing 8 (2008) 749758 754
Let the subsystems G
1
and G
2
represent fuzzy PID controller
and plant under control, respectively. The overall feedback
system is described by the equations
e
1
= u
1
y
2
; e
2
= u
2
y
1
; y
1
= G
1
e
1
; y
2
= G
2
e
2
Suppose that both subsystems G
1
and G
2
are causal and stable,
and let g
1
= g(G
1
), the gain of G
1
and g
2
= g(G
2
), the gain of
G
2
. Also suppose that there are constants b
1
, b
2
, g
1
_0 and
g
2
_0 so that
|y
1
| = |G
1
e
1
| _ g
1
|e
1
| b
1
(22)
|y
2
| = |G
2
e
2
| _ g
2
|e
2
| b
2
(23)
Under these conditions, the system is BIBO stable if g
1
g
2
<1,
i.e. any bounded input pair (u
1
; u
2
) produces a bounded output
pair (y
1
; y
2
).
We consider the general case where the process under
control is nonlinear, denoted by A. By dening r(k) = u
1
(k),
e(k) = d(k) = e
1
(k), Du(k) = y
1
(k), u(k 1) = u
2
(k), u(k) =
e
2
(k) and y(k) = y
2
(k) in Fig. 2, we obtain the equivalent
closed loop system in Fig. 7.
Let
M
d
= sup
k _0
[d(k)[; M
v
= sup
k _0
[v(k)[ _
2
T
M
d
;
M
a
= sup
k _0
[a(k)[ _
2
T
M
v
or
4
T
2
M
d
;

M
d
= N
d
M
d
;

M
v
= N
v
M
v
;

M
a
= N
a
M
a
when d
N
(k), v
N
(k) and a
N
(k) are in the interval [l; l[ we have
from Eqs. (10)(14)
|Du(k)| = |y
1
(k)| = |G
1
e
1
(k)|
_
_
2M
3N
Du
__

N
1
DT

N
2
D

N
3
DT
2

N
d
[d(k)[
_

N
1
DT

N
3
DT
2

M
d
_
which is in the form of Eq. (22) with
g
1
=
_
2MN
d
3N
Du
T
2
_
A
B
(24)
where A = (7 u)l
5
l(1 u)[

M
2
a

M
2
d
l
2
(

M
2
a


M
2
d
)[T
(7 u)l
5
l(1 u)[

M
2
v

M
2
a
l
2
(

M
2
v


M
2
a
)[T
2
(7 u)
l
5
l(1 u)[

M
2
d

M
2
v
l
2
(

M
2
d


M
2
v
)[ and B = (15 u)l
6

(1 u)[

M
2
d

M
2
v

M
2
a
(

M
2
d

M
2
v


M
2
v

M
2
a


M
2
a

M
2
d
)l
2
(

M
2
d


M
2
v


M
2
a
)l
4
[. Next, we have
|y(k)| = |y
2
(k)| = |G
2
(k)e
2
(k)| or |Ae
2
(k)| _ |A|[e
2
(k)[
which is in the form of Eq. (23) with
g
2
= |A| < (25)
So the sufcient condition for the nonlinear fuzzy PID control
system to be BIBO stable is the parameters of the fuzzy PID
controller must satisfy the inequality g
1
g
2
<1 where g
1
and g
2
are dened in Eqs. (24) and (25), respectively.
5. Design of the simplest fuzzy PID controller
Eq. (10) can be rewritten as
Du(kT) =
_
2M
3N
Du
__
N
1
N
v
D
v(kT)
N
2
N
d
D
d(kT)

N
3
N
a
D
a(kT)
_
(26)
Now by comparing Eq. (26) with Eq. (1) one can easily
recognize that fuzzy PID controllers are very much similar
in structure to the linear PID controllers. Since N
1
, N
2
, N
3
and
D are not just constants but are nonlinear functions of normal-
ized inputs d
N
(kT), v
N
(kT) and a
N
(kT), the fuzzy PID con-
troller is a nonlinear PID controller with the dynamic gains
dened by
K
Pd
=
2MN
1
N
v
3N
Du
D
;
K
Id
=
2MN
2
N
d
3N
Du
D
and K
Dd
=
2MN
3
N
a
3N
Du
D
(27)
where K
Pd
, K
Id
and K
Dd
are, respectively the dynamic propor-
tional gain, dynamic integral gain and dynamic derivative gain
of fuzzy controller. In the case of linear PID controller the gains
K
d
P
, K
d
I
and K
d
D
are constants, i.e. static. Hence, for comparison
sake, we also dene static gains of fuzzy PID controller by
making d
N
(kT) = v
N
(kT) = a
N
(kT) = 0 in Eq. (27) to get
K
Ps
= bN
v
; K
Is
= bN
d
and K
Ds
= bN
a
(28)
where K
Ps
, K
Is
and K
Ds
are, respectively the static proportional
gain, static integral gain and static derivative gain, and
b =
2M(7 u)l
5
3N
Du
(15 u)l
6
=
2M(7 u)
3N
Du
(15 u)l
(29)
To implement the mathematical model of the simplest fuzzy
PID controller, the design parameters are to be appropriately
chosen. For this a simple procedure is developed here based on
the design methodology of fuzzy two-term controllers [11,22].
Let the proportional gain, integral gain and derivative gain
be K
d
P
; K
d
I
and K
d
D
, respectively, for the best possible process
output corresponding to the incremental control effort in
Eq. (1). Assume that the peak absolute values of error
(displacement), velocity and acceleration of the ne tuned
Fig. 7. Feedback control system.
B.M. Mohan, A. Sinha / Applied Soft Computing 8 (2008) 749758 755
linear PID control system are [d[
max
, [v[
max
and [a[
max
,
respectively.
Assume that the sampling period T of the fuzzy PID control
system is the same as that of the linear PID control system. The
values of the functional parameters like N
d
; N
v
; N
a
; N
Du
; l; L
and M are calculated in the following manner.
Assuming that the static gains of the fuzzy PIDcontroller are
set equal to the respective gains of the linear PID controller, we
have from Eq. (28)
K
Ps
= K
d
P
= bN
v
; K
Is
= K
d
I
= bN
d
and
K
Ds
= K
d
D
= bN
a
(30)
or
K
d
P
K
d
I
=
N
v
N
d
and
K
d
P
K
d
D
=
N
v
N
a
(31)
The values of N
d
, N
v
and N
a
can be any arbitrary numbers as
long as the ratios in Eq. (31) are maintained. For simplicity let
N
v
= 1. Then
N
d
=
K
d
I
K
d
P
and N
a
=
K
d
D
K
d
P
(32)
For simplicity assuming l = M, the value of N
Du
can be
calculated from Eqs. (29) and (30), i.e.
N
Du
=
2(7 u)
3(15 u)K
d
P
(33)
Recall that 0 _ u <1. Theoretically analyzing the inuence of u
value on the controller performance is not easy. Its selection is
empirical and plant dependent.
The value of parameter l can be chosen such that all the three
normalized input variables d
N
, v
N
and a
N
lie in the interval
[l; l[ which is possible if the value of l is found from
l = max N
d
[d[
max
; N
v
[v[
max
; N
a
[a[
max
(34)
The value of L can be chosen such that L_l depending upon the
situation.
Usually for any process (especially a nonlinear one),
the desired performance is not obtained with the aforemen-
tioned design methodology of calculating functional para-
meters. This normally happens if the performance of the
tuned linear PID control system is not satisfactory. In such a
situation, all the functional parameters of the fuzzy PID
controller should be carefully ne tuned locally about their
initial values (obtained from the linear PID controller as
explained above) on trial-error-basis to obtain the desired
performance.
6. Illustrative examples
Comparison of the performances of linear PID controller
and the simplest fuzzy PID controller is done here by
considering the following examples:
(i) a linear third-order nonminimum phase system [2]
G
p
(s) =
s
2
s 2
s
3
3s
2
10s 24
(35)
with unit setpoint, and
(ii) a nonlinear rst-order system [2]
y(t) = y(t) sin
2
(

[y(t)[
_
) u(t) (36)
with step input of magnitude 4 as the reference signal. In
Eq. (35), G
p
(s) represents the transfer function of the plant
to be controlled.
The fuzzy PID controllers are designed for the above two
plants using the design procedure in Section 5. The values of
sampling period T, proportional gain K
d
P
, integral gain K
d
I
,
derivative gain K
d
D
, maximum absolute displacement (error)
[d[
max
, maximum absolute velocity [v[
max
, and maximum
absolute acceleration [a[
max
are given in Table 3 for both the
processes.
The parameters N
d
, N
v
, N
a
, N
Du
, l, and M of the fuzzy
PID controllers, which gave rise to the responses in Figs. 8
and 9, are listed in Table 4 in which M
p
, t
r
and t
s
denote
peak overshoot, rise time and settling time, respectively.
u = 0 is considered for both the plants. Figs. 8 and 9
also show the responses with conventional PID controllers.
Upon comparison, it is evident from the plots that the
fuzzy PID controllers perform better, demonstrating their
Table 4
Attributes and time-domain performance data of plants with linear and fuzzy PID controllers
Process PID controller N
d
N
v
N
a
N
Du
l = M M
p
(%) t
r
(s) t
s
(s)
G
p
(s) Linear 85.2357 0.009 0.4020
Fuzzy 1800 4.0 0:41 10
4
1.6 3800 0.2687 0.0022 0.0044
Nonlinear Linear 37.4442 1.0 5.6
Fuzzy 3 3 0.045 2.1 20 0.184 1.164 1.6
Table 3
Attributes of the nonminimum phase, and nonlinear processes
Process T K
d
P
K
d
I
K
d
D
[d[
max
[v[
max
[a[
max
G
p
(s) 0.001 10.5T 20,000T 0.0005T 1 131.343 19960.067
Nonlinear 0.1 1:8T 1:8T 0:008T 4 4.10676 994.592
B.M. Mohan, A. Sinha / Applied Soft Computing 8 (2008) 749758 756
superiority over their counterparts-conventional PID con-
trollers.
7. Conclusions
In this paper, analytical structure for fuzzy PID controllers
has been derived using L-type and G-type input fuzzy sets,
trapezoidal output fuzzy sets, algebraic product triangular
norm, bounded sum triangular co-norm, Mamdani minimum
inference method and COS defuzzication method. It has been
found that the maximum (M) and minimum (M) values of
normalized incremental control effort Du(kT) do not change for
any value of u in [0; 1). BIBO stability results for fuzzy PID
control have also been derived.
Since analytical structure of a fuzzy controller, in general,
depends on the type of triangular norm, and minimum
triangular norm has already been proved to be inappropriate for
realizing fuzzy PID control [13], the important task that
remains to be done is nding analytical structures for fuzzy PID
controllers via other types of triangular norms, if possible, and
studying their properties. It is believed that research in this
direction is desirable to have a clear picture of fuzzy PID
control systems.
Acknowledgments
Both the authors are highly grateful to the anonymous
Referees, and the Editor in Chief for their helpful suggestions
and encouragement. The second author would like to thank the
Council of Scientic and Industrial Research, India for
providing fellowship for the above work.
Appendix A
Consider Fig. 5. Trapezoid ABEF represents inferred fuzzy
set whose area is calculated as follows:
Area of inferred fuzzy set
= area of triangle ABC area of rectangle BEDC
area of triangle DEF:
Since triangles ABC and DEF are identical, we have
Area of inferred fuzzy set
= area of rectangle BEDC 2 area of triangle DEF
= 2a m 2
1
2
(b a) m = m(a b)
where a = m(b m) b from Fig. 5. Substituting the value
of a, we obtain
Area of inferred fuzzy set = m(2b mb mm)
= mb(2 m mu)
where u = m=b.
Assuming b = (2M 3m)=3, we have
Area of inferred fuzzy set =
m(2 m mu)(2M 3m)
3
and u = 3m=(2M 3m):
References
[1] J. Aracil, F. Gordillo (Eds.), Stability Issues in Fuzzy Control, Physica-
Verlag, Heidelberg, 2000.
[2] J. Carvajal, G. Chen, H. Ogmen, Fuzzy PID controller: design, perfor-
mance evaluation, and stability analysis, Inform. Sci. 123 (2000) 249270.
[3] G. Chen, Conventional and fuzzy PID controllers: an overview, Int. J.
Intell. Control Syst. 1 (1996) 235246.
[4] M. Golob, Decomposed fuzzy proportional-integral-derivative control-
lers, Appl. Soft Comput. 1 (2001) 201214.
[5] M. Guzelkaya, I. Eksin, E. Yesil, Self-tuning of PID-type fuzzy logic
controller coefcients via relative rate observer, Eng. Appl. Artif. Intell.
16 (2003) 227236.
[6] B.G. Hu, G.K.I. Mann, R.G. Gosine, A systematic study of fuzzy PID
controllersfunction-based evaluation approach, IEEE Trans. Fuzzy
Syst. 9 (2001) 699712.
[7] T.T. Huang, H.Y. Chung, J.J. Lin, A fuzzy PID controller being like
parameter varying PID, in: Proceedings of the IEEE Int. Fuzzy Systems
Conference, Korea, (1999), pp. 269276.
Fig. 9. Step (magnitude 4) response of closed loop system with nonlinear
process.
Fig. 8. Unit step response of closed loop system with G
p
(s).
B.M. Mohan, A. Sinha / Applied Soft Computing 8 (2008) 749758 757
[8] J.H. Kim, S.J. Oh, A fuzzy PID controller for nonlinear and uncertain
systems, in: Soft Computing, vol. 4, Springer Verlag, 2000, pp. 123129.
[9] H.X. Li, H.B. Gatland, Conventional fuzzy control and its enhancement,
IEEE Trans. Syst. Man Cybern.-Part B 26 (1996) 791797.
[10] W. Li, Design of a hybrid fuzzy logic proportional plus conventional
integral-derivative controller, IEEE Trans. Fuzzy Syst. 6 (1998) 449462.
[11] H.A. Malki, H. Li, G. Chen, New design and stability analysis of fuzzy
proportional-derivative control systems, IEEE Trans. Fuzzy Syst. 2 (1994)
245254.
[12] M. Margaliot, G. Langholz, New Approaches to Fuzzy Modeling and
Control: Design and Analysis, World Scientic, 2000.
[13] M. Mizumoto, Realization of PID controls by fuzzy control methods,
Fuzzy Sets Syst. 70 (1995) 171182.
[14] B.M. Mohan, A.V. Patel, Analytical structures and analysis of the simplest
fuzzy PD controllers, IEEE Trans. Syst. Man Cybern.-Part B: Cybern. 32
(2002) 239248.
[15] A.V. Patel, B.M. Mohan, Analytical structures and analysis of the simplest
fuzzy PI controllers, Automatica 38 (2002) 981993.
[16] W.Z. Qiao, M. Mizumoto, PID type fuzzy controller and parameters
adaptive method, Fuzzy Sets Syst. 78 (1996) 2335.
[17] K. Tanaka, H.O. Wang, Fuzzy Control Systems Design and Analysis: A
Linear Matrix Inequality Approach, John Wiley & Sons, NY, 2001.
[18] M. Vidyasagar, Nonlinear Systems Analysis, 2nd edn., Prentice-Hall,
Engle Wood, 1993.
[19] Z.W. Woo, H.Y. Chung, J.J. Lin, A PID type fuzzy controller with self-
tuning scaling factors, Fuzzy Sets Syst. 115 (2000) 321326.
[20] J.X. Xu, C.C. Hang, C. Liu, Parallel structure and tuning of a fuzzy PID
controller, Automatica 36 (2000) 673684.
[21] H. Ying, W. Siler, J.J. Buckley, Fuzzy control theory: a nonlinear case,
Automatica 26 (1990) 513520.
[22] H. Ying, Fuzzy Control & Modeling: Analytical Foundations and
Applications, IEEE Press, 2000.
B.M. Mohan, A. Sinha / Applied Soft Computing 8 (2008) 749758 758