Вы находитесь на странице: 1из 8

Tuning of PID controllers with fuzzy logic

A.Visioli

Abstract: The paper presents a comparison between different methods, based on fuzzy logic, for
the tuning of PID controllers. Specifically considered are different control structures in which a
fuzzy mechanism is adopted to improve the performances given by Ziegler-Nichols parameters.
To verify the full capabilities of each controller, genetic algorithms are used to tune the parameters
of the fuzzy inference systems (scaling coefficients, shape of the membership functions, etc.).
Furthermore, a discussion about the practical implementation issue of the controllers is provided,
and comparisons made with a typical PID-like fuzzy controller and a standard and a nonlinear PID
controller. The results show the superiority of the fuzzy set-point weighting methodology over the
other methods.

1 Introduction fuzzy inference system is used to tune the PID gains


depending on the current operating conditions of the
Although many innovative methodologies have been controlled system. In this way, it is implemented as a
devised in the past 50 years to handle more complex fuzzy supervisor more than a fuzzy controller, a solution
control problems and to achieve better performances, the known to be more appropriate for industrial control
great majority of industrial processes are still controlled by systems [7]. However, a comparison between the different
means of simple porportional-integral-derivative (PID) solutions has not been performed yet and in some cases
controllers. This seems to be because PID controllers, there are unclear points that still need to be discussed, to
despite their simple structure, assure acceptable perfor- verify the practical usage of the controller in industrial
mances for a wide range of industrial plants and their usage settings. For example, it is not clear how to set the
(the tuning of their parameters) is well known among parameters (scaling coefficients, shape of the membership
industrial operators. Hence, PID controllers provide, in functions and so forth) of the fuzzy mechanisms to
industrial environments, a cost/benefit performance that improve performances. Then, it has to be shown that a
is difficult to beat with other kinds of controllers. However, proposed control system is superior to other, simpler, ones,
because of their simple structure, PID controllers are and it is necessary to show that a controller devoted to set-
particularly suited for pure first- or second-order processes, point following is also effective in load-disturbance
while industrial plants often present characteristics such as attenuation.
high order, time delays, nonlinearities and so on. In this In other words, the trade-off between the potential
context, the tuning of the parameters is a crucial issue and improvement in the performance and the increased
the many tuning formulae, that have been devised [l], such complexity in the controller's design and implementation
as the Ziegler-Nichols one, often fail to achieve satisfac- has to be quantified.
tory performances, and therefore the operator has to use The aim of this paper is to clarify these points by means
their experience and might fail to attain the best perfor- of a large number of simulation results and of a thorough
mances. discussion. In particular, different fuzzy logic based meth-
For these reasons it is highly desirable to increase the ods will be compared with a typical fuzzy PID-like
capabilities of PID controllers by adding new features; in controller, with a nonlinear PID controller and a standard
this way they can improve their performances for a wide PID controller with a set-point weight. To make a fair
range of plants while retaining their basic characteristics, comparison, the fuzzy inference systems have been opti-
so that industrial practicioners can still fully exploit their mally tuned separately for each considered plant by means
knowhow. of a genetic algorithm, a technique whose effectiveness has
In this context, the use of fuzzy logic seems to be been recognized in the literature [XI.
particularly appropriate, since it allows us to make use of
the operator's experience and therefore to add some sort of
intelligence to the automatic control [ 2 ] . On this basis, 2 PID tuning w i t h fuzzy logic
several solutions have been proposed [3-61. Specifically, a
2. I Standard PID control
The typical PID control law in its standard form is
0 IEE, 2001
IEE Proceedings online no. 20010232
DOL 10.1049/ip-cta:20010232
Paper first received 10th October 2000
The author is with the Dipartimento di Elettronica per l'Automazione, where e(t)=y,(t) - y(t) is the system error (difference
University of Brescia, Via Branze 38, 1-25123, Brescia, Italy between the reference input and the system output), u(t) the
E-mail: visioli@bsing.ing.unibs.it control variable, Kp the proportional gain, Td the derivative
IEE Proc.-Control Theory Appl., Vol. 148, No. I , January 2001
time constant and Ti the integral time constant. Eqn. 1 can other words, the current values of the proportional, integral
also be written as and derivative gains are increased or decreased by means
of a fuzzy inference system, according to the following
+de(t)
u(t) = Kpe(t) Kd - K;
dt
+ joe(z)dz (2) relations:

where Kd=KpTd and Ki=Kp/Tj. The tuning problem


+
P = P CV(e(t),e(t)} x k,
consists of determining the values of these three para- +
I = I CV{e(t),e(t)) x k2 (4)
meters with the aim of satisfying different control specifi- D = D + CV(e(t),e(t)) x k3
cations such as set-point following, load disturbance
attenuation, robustness to model uncertainties and rejec- where the basic tuning is the Ziegler-Nichols one, CV(e(t),
tion of measurement noise. The several tuning rules t.(t)} is the output of the fuzzy inference system, based on
proposed in the literature along the past 50 years are the Macvicar-Whelan fuzzy rule matrix [ 111, which
generally devoted to meet one in particular of these reflects the typical action of a human controller (for
requirements. For example, the widely known Ziegler- example, the integral action has to be increased at the
Nichols formula [9] assures a good load-disturbance beginning of the transient response to decrease the rise
attenuation, but it generally provides a poor phase time and then it has to be decreased when the system error
margin, and therefore it might produce a large overshoot is negative, to reduce the overshoot). Finally, k , , k, and k3
and settling time in the step response. To avoid these are constant parameters that determine the range of varia-
situations partially, a modification can be applied to eqn. tion of each term. The whole fuzzy system involves 14
1. Specifically, to reduce the overshoot, the set-point for the quantisation levels for both e and b. It has to be stressed
proportional action can be weighted by means of a constant that the tuning of the three parameters k, ,k2 and k3, and of
parameter b < 1 [ 1, 101 so that we have the following more the two scaling factors that multiply the two inputs e and b
general expression: (the range of the input membership functions is assumed to
be normalised between - 1 and l), is left to the user, and it
might be a difficult task, as it is not clear how these
parameters influence the performances of the overall
controller, for a generic system.
where b may be either 0 or 1. In this way a two degrees of
freedom controller is implemented. Namely, a part of the
control structure is devoted to the load disturbances 2.3 Fuzzy self-tuning of a single parameter (SSP)
attenuation (for which a high loop gain is desirable) and The method devised by He et al. [4] consists of parame-
the other part is devoted to the set-point following. Fig. 1 terising the Ziegler-Nichols formula by means of a single
shows this scheme for the following transfer functions: parameter a,then using an online fuzzy inference system to
self-tune the parameter. In this way, the three PID para-
meters can be expressed as
Kp = 1.2a(t)ku
1
T, = 0.75-
+
1 a(t)tu (5)
The main drawback of this method is that it leads to an Td = 0.25Ti
increasing of the rise time, since, obviously, the effective-
ness of the proportional action is somewhat reduced. where k, and tu are the ultimate gain and the ultimate
Fuzzy logic can be used in the above context to vary the period, respectively, of the process. The value of a(t) is
PID parameters values during the transient response, in determined recursively with the following equation:
order to improve the step response performances, main-
taining a basic tuning aimed at properly rejecting load
disturbances (e.g. the Ziegler-Nichols parameters). In this
way, it is possible to implement a nonlinear controller so
that the rise time and the overshoot are reduced at the same where y is a positive constant that has to be chosen in the
time. range [0.2,0.6] and h(t) is the output of the fuzzy inference
system which has seven membership functions both for
each of the two inputs (the system error and its derivative)
2.2 Incremental fuzzy expert PID control (IFE) and for the output. It has to be noted that the initial value of
Tzafestas and Papanikolopoulos [3] proposed to scale the a(t) is set equal to 0.5 which corresponds to the Ziegler-
values of the three controller parameters (initially deter- Nichols formula. Also for the SSP method, the tuning of
mined by the Ziegler-Nichols formula) during the transient the scaling coefficient of the fuzzy module and of the
response depending on the system error and its rate. In parameter y is left to the user and no rules of thumb are
given for this task.

2.4 Fuzzy gain scheduling (FGS)


According to the methodology proposed by Zhao et al. [ 5 ] ,
the three current PID parameters are determined as
follows:

Fig. 1 Two degrees-of-freedom scheme of PID controller with set-point


weighting K; = Kj/(aK,)
L IEE Proc.-Control Theory Appl., Vol. 148, No. 1, January 2001
where Kk, K&,and a are determined by means of a fuzzy Table 1. Meaning of the linguistic variables in the fuzzy
mechanism and Kpsm,, Kp,min,Kd," and Kd:min are inference system
adopted to normalise the values of Kp and Kd into the
range between zero and one. These constants are deter- NVB negative very big
mined by the following rule of thumb: NB negative big
NM negative medium
NS negative small
Z zero
PS postive small
Seven triangular membership functions are used for the PM postive medium
two inputs e(t) and e(t) of the fuzzy module, while only two PB postive big
are used for the outputs Kh and K& and four singletons PVB postive very big
define the output of a. In this scheme, only the two scaling
coefficients of the two inputs of the fuzzy module (the
system error and its derivative) have to be fixed by the user,
so that setting the overall controller appears to be easier
than in the case of the IFE and SSP methods.

2.5 Fuzzy set-point weighting (FSW)


The approach proposed by Visioli [6] consists of fuzzify-
ing the set-point weight, leaving fixed the other three
parameters (again determined with the Ziegler-Nichols Fig. 3 Control scheme with the fuzzy set-point weighting (FSW)
methodology
method to preserve a good load disturbance attenuation).
In this way, the control law can be written as
Finally, Reference [6] shows the methodology is robust
to variation of the Ziegler-Nichols parameters, and there-
fore its practical implementation seems to be quite simple.

2.6 Fuzzy PID-like controller


where w is a positive constant parameter less than or equal
to 1, andf(t) is the output of the fuzzy inference system, The above methodologies, in which fuzzy logic is used to
which consists of five triangular membership functions for determine the parameters of the PID controller, have been
the two inputs e(t) and e(t) and nine triangular membership compared also with a fuzzy PID-like controller, in which
functions for the output. The fuzzy rules are based on the the control variable is determined directly by means of a
Macvicar-Whelan matrix [ I l l , as shown in Fig. 2 (the fuzzy inference system. The evaluated PID-like controller
meaning of the linguistic variables is described in Table 1). is the one proposed in Reference [13],in which the control
Fig. 3 shows the overall control scheme. It is worth variable is the sum of the outputs of a fuzzy PI-like
stressing that in this method the role of the fuzzy mechan- controller and a fuzzy PD-like controller (see Fig. 4).
ism parameters is somewhat intuitive, and it is very similar This choice allows to keep a low number of rules (with
to the one in the typical fuzzy PD-like controller, for which respect to the typical case of the fuzzy PID-like controller,
tuning procedure have been established [ 121. Hence, the in which three inputs have to be considered) without
task of the user is simplified. In any case, a simple decreasing the performances. Two membership functions
empirical procedure for the manual tuning of the fuzzy are used for the two inputs e and 2, and three singletons for
module has been proposed. the output, which is of the Takagi-Sugeno type. It turns out
that two scaling factors for the two inputs and two for the
Ae
two outputs of the PI and PD-like controller have to be
NB NS z PS PB selected by the user. Typical rules of thumb regarding the
fuzzy controllers can be adopted in this case, as for the
fuzzy set-point weighting.
Note that the implemented scheme encompasses both
the single standard PI-like and PD-like controller and
therefore also the well known fuzzy sliding mode control

e . fuzzy
-
.
PI-like
e controller

o u.
- fuzzy
-
.
PD-like
controller

Fig. 2 Basic rules table of the fuzzy inference system for the FSW
scheme Fig. 4 Control scheme with PID-like fuzzy controller
IEE Proc.-Control Theory Appl., Vol. 148, No. 1, Janualy 2001 3
[14]. This can be seen as an improved version of the PD- with respect to the pure Ziegler-Nichols method. Hence, in
like control, in which the full control capabilities are this case, also the rules of the fuzzy mechanism have been
retained despite the simplified design (the number of optimally selected by means of the genetic algorithm.
rules is reduced) [ 141.
4 Simulation results
2.7 Nonlinear PID controller
Finally nonlinear PID controller, in which fuzzy logic is The performances of the different controllers have been
not adopted, has been considered. The implemented evaluated on a wide range of plants. Here, the following
control law is [ 151 transfer functions, with different values of the parameters,
are considered:

4 = 1; = 0.2,0.8; (12)
G1(s) = s2 + 2w5s + CO: ; W,

It comes out that the proportional gain depends on the


current system error, and the parameter L E [0, 11 represent 1
G~(s)= -’
the degree of nonlinearity: if L = 1 we have a typical PID s(1 +s)’
(linear) controller and, on the contrary, if L=O the
controller is highly nonlinear, i.e. the proportional gain is p L
bigger when the error is big than when the error is small. In G~(s)= T = 1, 10; L = 0.1,0.4, 0.8; (14)
this control scheme there are four parameters to tune, the
~

(1 + sT)’ *

PID gains and the scalar constant L.
1
3 Optimal tuning of the fuzzy modules by ) -.
G ~ ( s=
(1 +s)3 ’
means of genetic algorithms
e-sL
To test the full capabilities of the examined control
schemes, for each case and for each considered process G,(s) = (1 +s)(l +0.5s)(l +0.25s)(l +0.125s)’
under control (cf. Section 4), the free parameters have been
determined using a genetic algorithm [ 161, which guaran- L = 0,O.l; (16)
tees in a stochastic sense that a global optimum is
achieved. The selected objective function to minimise is (1 - 0.5s)
the integrated absolute error, defined as G6(S) =
(1 +s)3 .
First, to verify the full potentialities of the investigated
methodologies, it will be assumed that no saturation levels
which somehow takes into account at the same time the are present for the control variable. After the tuning phase,
rise time, the overshoot and the settling time. Note that accomplished using the genetic algorithm, the unit step
when triangular membership functions are employed, the responses have been simulated with Matlab and Simulink.
genetic algorithm also searches for the optimal shapes of The resulting values of the IAE for the different processes
the membership functions. In particular, the optimal posi- and different controllers are reported in Table 2. To
tion of the peaks of all the membership functions is compare the control laws, tuned by a genetic algorithm
searched, except for the central ones and those at the with classical industrial methodologies, it also shows a PID
limits of the variable range, both for the inputs and the tuned with the Ziegler-Nichols method, both without
outputs. For the sake of simplicity, a symmetric shifting of (‘Z-N’) and with (‘Z-N b’) a fixed set-point weight. In
the involved membership functions for each variable has the latter case, the value of b has been optimally selected
been considered. The bases of the other membership again to minimise the IAE parameter) by a simple iterative
functions are updated accordingly to preserve the fuzzy procedure. For example, the step responses of system (eqn.
partition property (i.e. the sum of the membership degrees 15) with the different controllers are plotted in Figs. 6 and
for each value of the variable is always equal to one). This 7. Note that the ‘fixed b’ and ‘SSP’ cases are missing, as
procedure is shown in Fig. 5. they are equal to the ‘PID’ and ‘Z-N’ cases, respectively.
The only exception to this optimisation procedure has To provide a more detailed insight of the results, Tables
been applied to the SSP method, since it has been seen that 3 and 4 shows the values of overshoot and the rise time,
searching for the constant parameters and the shape of the respectively, for the different cases.
membership functions does not lead to any improvement Although the investigated methodologies have been
devised for the control of pure linear systems, a saturation
level U, = f 5 on the control variable has been considered
t
for each case to test the controllers’ capabilities in the
presence of a typical process nonlinearity. For each
scheme, with the exception of the fuzzy PID-like control-
ler, an anti-windup system (Reference 1, page 83) has been
used, where the tracking time constant T, has been tuned by
the rule of thumb Tt = l/(TiTd), where Ti and Td are the
-1 0 1
c
base integral and derivative time constants of the control-
e, e ler. The controller parameters have been retuned, again by
Fig. 5 Tuning o f p e a b of membership functions means of genetic algorithms. Results are reported in Tables
A symmetric shifting of the intermediate membership functions is permitted 5, 6 and 7.
4 IEE Proc.-Control Theory Appl., Vol. 148, No. I , January 2001
Table 2. Value of IAE achieved by the examined controllers tuned by the genetic algorithm (compared also with the
Ziegler-Nichols response) when no saturation is present

Process PID fixed b IFE SSP FGS FSW fuzzy nonlinear Z-N Z-N b

GI( s ) , 5 = 0.2 0.90 0.90 0.83 1.16 0.94 0.60 0.99 1.32 3.34 2.23
G~(s), (~0.8 0.70 0.70 0.83 0.95 1.27 0.56 0.99 0.83 1.27 0.97
G2 ( S ) 0.85 0.79 1.01 2.09 1.56 0.66 1.55 1.06 2.09 1.44
G,(s),T = l , L=0.1 0.72 0.72 0.95 1.05 1.40 0.75 0.86 0.73 1.40 0.99
G,(s),T = l , L=0.4 1.31 1.31 1.48 1.48 1.34 1.38 1.57 1.30 1.79 1.50
G,(s), T = l , L=0.8 2.00 1.99 2.22 1.96 2.81 2.10 3.25 2.08 2.22 2.13
G,(s), T=10, L=0.1 0.85 0.82 2.21 4.91 1.91 0.81 3.92 1.59 4.91 2.50
G,(s),T=10, L=0.4 1.39 1.39 3.69 7.62 3.65 1.66 3.26 3.13 7.62 4.05
G,(s),T=10, L=0.8 2.24 2.24 4.20 9.38 5.17 2.60 3.87 3.70 9.38 5.14
G~(s) 1.13 1.13 1.13 2.34 1.34 0.98 1.49 1.72 2.34 1.89
G,(s),L=O 0.99 0.98 1.01 1.55 1.44 1.22 1.27 1.09 1.83 1.36
G~(s),
L=0.1 1.25 1.25 1.22 1.29 1.31 1.30 1.37 1.31 1.59 1.41
G6(S) 2.12 2.11 2.60 2.86 2.87 2.43 5.28 2.86 2.86 2.53

Table 3. Value of the overshoot (%) achieved by the examined controllers tuned by the genetic algorithm (compared
also with the Ziegler-Nichols response) when no saturation is present

Process PID fixed b IFE SSP FGS FSW fuzzy nonlinear 2-N Z-N b

GI ( s ) , 5 = 0.2 17 16 0 16 0 18 1 22 46 18
GI ( s ) , 5 10.8 31 12 21 20 41 22 2 13 41 6
G,(s) 21 0 1 31 20 15 0 8 31 9
G~(s),
T= 1, L=0.1 13 13 41 19 57 1 0 3 57 3
G~(s),T= 1, L=0.4 8 10 33 4 18 6 11 12 37 4
G~(s),
T= 1, Lz0.8 5 6 23 7 22 10 0 10 23 8
G~(s),
T=10, L=0.1 1 0 30 66 13 2 3 6 66 9
G~(S),
T=10, L r 0 . 4 10 10 11 62 13 3 2 3 62 6
G~(s),T=10, L=0.8 9 12 6 59 13 1 0 3 59 5
G~(s) 14 14 3 49 15 4 2 26 49 5
G~(S),
L=O 24 23 17 15 15 8 5 8 52 6
G~(s),
Lz0.1 7 7 15 6 10 6 14 15 34 6
G~(S) 14 9 22 34 15 21 17 18 34 17

Table 4. Rise times Is)achieved by the examined controllers tuned by the genetic algorithm (compared also with the
Ziegler-Nichols response) when no saturation is present

Process PID fixed b IFE SSP FGS FSW fuzzy nonlinear Z-N Z-N b

GI (s),5 = 0.2 0.38 0.39 0.50 0.93 0.77 0.11 0.66 0.46 0.46 0.72
GI ( s ) ,5 = 0.8 0.11 0.13 0.34 0.45 0.32 0.07 0.78 0.35 0.32 0.46
G2 (s) 0.31 0.59 0.46 0.96 0.58 0.20 1.21 0.84 0.96 1.24
G,(s), T=l, L=0.1 0.17 0.17 0.22 0.44 0.28 0.24 0.37 0.23 0.28 0.48
G,(s), T = l , L=0.4 0.52 0.58 0.44 0.90 0.46 0.68 0.38 0.52 0.60 0.84
G,(s),T=l, L=0.8 0.98 1.00 0.86 1.00 0.82 0.98 3.06 1.oo 0.88 1.02
G,(s),T= IO, L=0.1 0.42 0.42 1.00 0.88 1.22 0.30 5.32 1.20 0.88 1.58
G3(s),T= IO, L=0.4 0.66 0.66 2.14 1.78 2.38 1.16 3.26 3.58 1.78 3.02
G,(s), T= I O , L=0.8 1.24 1.20 2.02 2.50 2.82 1.78 3.82 3.52 2.50 4.26
G4(s) 0.65 0.65 0.73 0.63 0.82 0.54 1.14 0.70 0.68 1.48
G~(s),
L=O 0.34 0.35 0.38 0.74 0.40 0.58 0.72 0.58 0.56 0.88
G~(s),L=0.1 0.56 0.56 0.46 0.60 0.56 0.70 0.46 0.54 0.60 0.80
G6(S) 0.84 1.02 0.90 0.80 0.88 0.92 0.90 1.oo 0.80 1.02

IEE Proc-Control Theory AppL. Vol. 148, No. I , January 2001 5


1.6 r

time, s time, s
Fig. 6 Step responses of G4(s) with the different controllers without Fig. 7 Step responses of G4(s) with the different fizzy logic based
fuzzy logic controllers

Table 5. IAE achieved by the examined controllers tuned by the genetic algorithm (compared also with the Ziegler-
Nichols response) when saturation of the control variable is present
Process PID fixed b IFE SSP FGS FSW fuzzy nonlinear 2-N Z-N b

GI (s),c = 0.2 1.74 1.09 0.98 1.56 1.41 0.96 1.06 1.36 3.33 2.66
GI ( s ) , { = 0.8 1.05 1.00 1.01 1.17 1.28 1.01 1.08 1.07 1.44 1.41
G~(s) 2.53 1.19 1.14 2.35 2.12 0.97 1.12 2.06 3.49 3.01
G~(s),T = l , L=0.1 1.12 1.10 1.11 1.14 1.26 1.10 1.22 1.11 1.41 1.33
G,(S), T = l , LzO.4 1.65 1.59 1.46 2.00 1.64 1.53 1.46 1.52 2.00 1.95
G~(s),T = l , L=0.8 2.46 2.34 2.40 2.32 2.50 2.35 2.94 2.27 2.55 2.54
G,(s), T=10, L=0.1 6.03 5.73 5.70 5.82 5.90 5.70 6.70 5.78 6.20 6.20
G~(s),T=10, L=0.4 6.51 6.01 6.07 6.57 6.98 6.03 7.57 6.17 7.53 7.53
G,(s), T=10, L=0.8 7.28 6.44 6.55 7.29 7.12 6.50 7.69 6.64 9.11 9.11
G~(s) 1.75 1.68 1.65 1.91 2.09 1.68 2.23 1.70 2.73 2.33
G~(s),L = O 1.34 1.28 1.26 1.44 1.45 1.25 1.32 1.25 1.65 1.59
G~(s),Lx0.1 1.55 1.52 1.35 1.50 1.63 1.43 1.78 1.39 1.78 1.75
Gtj(S) 2.98 2.74 2.60 2.81 2.82 2.68 2.85 2.57 3.37 3.11

Table 6. Overshoot (%) achieved by the examined controllers tuned by the genetic algorithm (compared also with the
Ziegler-Nichols response) when saturation of the control variable is present

Process PID fixed b IFE SSP FGS FSW fuzzy nonlinear Z-N 2-N b

GI (s), 5=0.2 0 13 4 22 3 10 2 28 45 17
GI (s), {=0.8 0 1 7 20 15 4 0 8 22 25
G2 (s) 0 7 11 38 32 6 2 38 19 27
G~(s),T=l, L=0.1 2 2 3 5 3 2 0 1 28 19
G3(s), T= 1, L = 0.4 12 5 4 30 10 4 9 13 30 10
G~(s), T=l, L=0.8 28 5 29 10 24 7 2 20 15 11
G,(S), T=10, L=0.1 3 3 0 4 4 0 1 1 8 8
G~(s), T= I O , L=0.4 0 0 1 3 8 0 1 2 11 11
G ~ ( S )T=
, I O , L=0.8 0 0 2 8 2 1 2 1 14 14
G4 (s) 8 7 6 13 23 1 0 9 4 12
G~(s),L=O 8 5 4 16 25 3 4 13 32 12
G~(s),L=0.1 12 8 2 6 11 5 17 15 26 17
G6 (S) 20 4 20 6 10 8 4 14 35 11

6 IEE Proc.-Control Theory Appl., Vol. 148. No. I, January 2001


Table 7. Value of the rise timers (s)achieved by the examined controllers tuned by the genetic algorithm (compared also
with Ziegler-Nichols response) when saturation of the control variable is present
Process PID fixed b IFE SSP FGS FSW fuzzy nonlinear Z-N Z-Nb

G ~ ( s )5, ~ 0 . 2 2.60 0.56 0.54 0.82 0.86 0.46 0.58 0.56 0.68 0.98
GI (s), 5 = 0.8 0.61 0.60 0.54 0.54 0.60 0.60 0.68 0.62 0.66 0.60
G~(s) 4.74 0.84 0.70 0.96 0.88 0.50 0.70 0.90 2.34 1.84
G ~ ( s ) T=
, 1, L=0.1 0.56 0.58 0.58 0.64 0.74 0.62 0.84 0.62 0.56 0.58
G ~ ( s ) T=
, 1, L z 0 . 4 0.70 0.76 0.68 0.78 0.76 0.70 0.64 0.60 0.78 1.10
G ~ ( s ) T=
, 1, L=0.8 0.98 1.20 0.74 1.20 1.20 1.12 2.18 0.86 1.22 1.30
G ~ ( s ) T=
, IO, L=0.1 5.58 5.58 5.64 5.58 5.88 5.70 5.96 5.80 5.56 5.56
G3(s), T= IO, L=0.4 5.68 5.68 5.64 5.70 5.74 5.80 10.3 5.96 5.56 5.56
G ~ ( s )T=
, IO, L=0.8 5.72 5.72 5.70 5.70 5.86 6.00 6.20 6.28 5.98 5.98
G4 ( S ) 0.96 1.oo 1.02 1.06 0.94 1.24 2.56 1.08 1.88 1.48
G ~ ( s )L, = O 0.58 0.64 0.70 0.70 0.54 0.68 0.80 0.58 0.68 0.88
G ~ ( s )L=0.1
, 0.72 0.76 0.78 0.90 0.74 0.74 0.76 0.60 0.76 0.86
G~(S) 0.96 1.24 0.80 1.56 1.16 1.18 1.66 1.02 1.22 1.66

5 Discussion the same performances can be attained by simply applying


a fixed set-point weight, simple to determine and imple-
From the results, it appears that, without an amplitude limit ment.
on the control variable, it is very difficult to improve the Fuzzy-logic-based tuning is more useful if saturations
performances of an optimally tuned PID, even if we use a are significant in the process. In fact, it appears that in this
nonlinear (fuzzy-based) controller, The use of an additional case the performances of a classic PID can be improved
parameter b is almost useless if the PID is optimally tuned using time-varying parameters, and the achieved results are
(in this case, when the overshoot is reduced, the rise time difficult to improve by a standard fuzzy controller.
increases, so that the IAE does not decrease significantly). Between the considered schemes, the FSW seems in
However, it has to be highlighted again that in practical general to ensure better performances, although the
cases it is very difficult to achieve an optimal tuning, so it values of the IAE attained by the different controllers are
is appropriate to compare the results obtained by using closer in this case than when there is no saturation.
fuzzy-logic-based methodologies with those obtained with Finally, one should consider the ease of tuning of the
the Ziegler-Nichols tuning formula. In this context, the controllers (i.e. of the fuzzy module’s parameters), since in
only two techniques that are generally able to improve the practical industrial applications, the use of genetic algo-
performances achieved with the simple use of a fixed set- rithms is possible if a good process model is available.
point weight are the IFE and the FSW. In particular, the Many simulations (not reported here for the sake of
latter appears to be the best method and it gives better step brevity) have been performed for this purpose. It emerges
responses than in a standard nonlinear PID. This is because that, for the fuzzy-logic-based tuning methods for which a
with the FSW method, both the overshoot and the rise time technique for the selection of the parameters of the fuzzy
can be decreased from the Ziegler-Nichols response (see inference system has not been proposed, it might very
Tables 3 and 4). Conversely, results show that the SSP and difficult to perform this task effectively. For example, for
FGS methods are not always worth using, as, in general, the IFE method, the choice of parameter k3 is critical since

Table 8. Value of the IAE on a load disturbance response achieved by the examined controllers turned by the genetic
algorithm (compared also with the Ziegler-Nichols response) when no saturation is present. The lack of a number
means that the system has become unstable
Process PID fixed b IFE SSP FGS FSW fuzzy nonlinear Z-N

0.66 0.46 - - - 0.003 - 0.51 0.39


0.06 0.09 - - - 0.02 - 1.49 2.80
7.99 33.35 - - 8.82 0.08 - - 7.39
0.57 0.17 - - - 0.16 - 0.70 0.27
2.23 1.37 - - 1.99 1.29 - 2.44 1.38
4.10 3.30 3.36 4.21 5.70 3.37 - 4.14 3.33
0.29 0.02 0.20 - 0.09 0.04 - 0.81 0.13
1.71 0.91 - - 0.99 0.44 - 4.46 0.92
3.93 1.14 - - 2.47 1.21 - 6.08 2.04
4.63 0.35 - - 1.74 0.15 - 1.97 1.40
1.84 1.84 - 1.05 1.36 0.83 - 1.94 0.93
1.97 1.84 - 1.77 3.1 1 1.39 - 1.99 1.33
5.30 5.15 - - 6.67 4.33 - 6.06 4.28

IEE Proc.-Control Theory Appl., Vol. 148, No. I , January 2001 7


its value has to be kept very low, otherwise the overall are significant. In other circumstances, it has to be taken
control system can be unstable. The derivative action is into account that the performances obtained by the fuzzy
critical in the FGS scheme as well, while SSP control based schemes might be achieved with simpler techniques.
structure is very difficult to set, as it depends on the results The ease of tuning of the fuzzy mechanism parameters
obtained from the genetic algorithm, which is not always plays a key role in the practical applicability of the
able to improve on the Ziegler-Nichols response and, when methodologies, since it determines the improvement in
it succeeds, the selected rules are very difficult to interpret. the costlbenefit ratio with respect to standard methods.
The only scheme in which the setting of the fuzzy module’s In this context, the fuzzy set-point weighting technique
parameters is straightforward is the FSW one, for which a appears superior to the others, as it guarantees in general
manual tuning procedure can be applied [6]. very good performances in the set-point and load distur-
Finally, the evaluation of the load disturbance attenua- bance step responses and it requires a modest implementa-
tion capabilities of the proposed schemes has been eval- tion effort; therefore its practical implementation in
uated. Specifically, a load disturbance step has been industrial environments appears to be very promising.
applied to the plant at steady-state initial conditions, with-
out limits on the control variable. The controllers are the
same as in the case of no saturation, i.e. tuned by the 7 Acknowledgment
genetic algorithm to minimise the integrated absolute error
on the set-point step response. The resulting IAE in the This work was supported in part by MURST scientific
different cases have been reported in Table 8. Obviously, research funds.
there is no difference in the results between two PID
controllers in which only the fixed set-point weight value 8 References
changes (remember that, on the contrary, all the PID
parameters are different between the ‘PID’ and the ‘fixed 1 ASTROM, K., and HAGGLUND, T.: ‘PID controllers: Theory, design
and tuning’ (ISA Press, 1995)
b’ columns). Gaps in the table mean that the controlled 2 TZAFESTAS, S.G.: ‘Fuzzy systems and fuzzy expert control: an over-
system has become unstable. Hence it is apparent that, for view’, Knowl. Eng. Rev., 1994, 9, (3), pp. 229-268
some methodologies, great care has to be taken in the 3 TZAFESTAS, S.G., and PAPANIKOLOPOULOS, N.P.: ‘Incremental
fuzzy expert PID control’, IEEE Trans. Ind. Electron., 1990,31, (5), pp.
tuning of the overall controller, as highly undesirable 365-371
effects might occur for some aspects of the control 4 HE, S.-Z., TAN, S., and XU, F.-L.: ‘Fuzzy self-tuning of PID control-
lers’, Fuzzy Sets Sys., 1993, 3 7 4 6
system. It has to be stressed again that the system has 5 ZHAO, Z.-Y., TOMIZUKA, M., and ISAKA, S.: ‘Fuzzy gain scheduling
become unstable with the tuning performed by the genetic of PID controllers’, IEEE Trans. Syst. Man, Cybern., 1993, 23, (5), pp.
algorithm in order to optimise the IAE of the set-point step 1392-1 398
6 VISIOLI, A.: ‘Fuzzy logic based set-point weighting for PID control-
response, so that the system can be stabilised provided that lers’, IEEE Trans. Syst. Man, Cybern. - Pt. A , 1999, 29, pp. 587-592
the controller is detuned. However, it is worth noting that 7 CHIU, S.: ‘Using fuzzy logic in control applications: Beyond fuzzy PID
the reported results about the following setpoint might not control’, IEEE Control Syst. Mag., 1998, 18, (5), pp. 100-105
8 HOMAIFAR, A., and McCORMICK, E.: ‘Simultaneous design of
be achieved in practical cases; hence, it turns out that those membership hnctions and rule sets for fuzzy controllers using genetic
controllers that present stability problems are less effective algorithms’, IEEE Trans. Fuzzy Syst., 1995, 3, (2), pp. 129-139
9 ZIEGLER, J.G., and NICHOLS, N.B.: ‘Optimum setting for automatic
than previously thought. From the load-disturbance controllers’, ASME Trans., 1942, 759-768
attenuation viewpoint, again, the FSW scheme is better 10 LEVA, A., and COLOMBO, A.M.: ‘Method for optimising set-point
than the others. weights in ISA-PID autotuner’, IEE Proc. Control Theory Appl., 1999,
~~

146; (2), pp. 137-146


11 MACVICAR-WHELAN,P.J.: ‘Fuzzy sets for man-machine interaction’,
Int. 1 Man-Mach. Stud., 1976, 8, PP. 687-697
6 Conclusions 12 ZHENG, L.: ‘A practical guide to (he tuning ofproportional and integral
(PI)-like fuzzy controllers’. Proc. IEEE Int. Conf. on Fuzzy systems, New
In this paper, a comparison between different methodolo- Orleans, pp. 633-640, 1992
13 LIU, C., XU, J.-X., and HANG, C.-C.: ‘Comparison between a fuzzy
gies regarding the tuning of PID controllers by means of PID controller and a kind of nonliner PID controller’, Proc. 36th IEEE
fuzzy logic has been presented. The exposed results have Int. Conf. on Decision and control, S. Diego (CA), pp. 2736-2741,1997
14 PALM, R.: ‘Fuzzy sliding mode control’, in BONIVENTO, C.,
allowed a detailed discussion about the use of fuzzy FANTUZZI, C., and ROVATTI, R. (Eds.) ‘Fuzzy logic control-
inference systems in this context. It has turned out that, Advances in methodology’ (World Scientific, 1998), pp. 75-109
despite the investigated schemes have been conceived for 15 SHINSKEY. F.G.: ‘Process control svstems: Auulication. design. and
pure linear systems, the main benefits in the use of fizzy
logic appear when process nonlinearities such as saturation optimization: A Matlab implementation’: NCSU-IE

8 IEE Proc.-Control Theory Appl., Vol. 148, No. 1, January 2001

Вам также может понравиться