&VSPQFBO
$POUSPM
$POGFSFODF
&$$


"VHVTU
o

4FQUFNCFS

,BSMTSVIF
(FSNBOZ

Comparison of fuzzy logic based tuning methods for PID

controllers
Antonio Visioli
Dipartimento di Elettronica per l'Automazione
Faculty of Engineering
Univerity of Brescia
Via Branze, 38 I-25123 Brescia Italy
e-mail: [email protected]

Keywords : PID controllers, fuzzy logic, tuning, ge- operator has to use its experience in tuning the propor-
netic algorithms. tional, derivative and integral gains and may not achieve
the best result. For these reasons, it would be very use-
ful to improve the capabilities of PID controllers in such
Abstract a way that their simple structure be preserved and the
acquired know-how be still fully exploited by industrial
In this paper, dierent methods based on fuzzy logic for practicioners.
the tuning of Proportional- Integral-Derivative PID con- In this context, fuzzy logic seems to be very appropriate
trollers are analyzed and compared. The easiness of im- to be used for the tuning of PID controllers, since non-
plementation and the eectiveness of the examined control linear terms can be implemented in the controller in a
schemes for a wide range of processes is discussed. Com- very intuitive way, reecting the operator experience in
parisons are made with respect to performances in the the de nition of the fuzzy rules 3  with this aim, dif-
set-point following step response where the parameters ferent schemes have been proposed 4, 5, 6, 7 . However,
of the employed fuzzy inference systems have been opti- some problems may arise in general with these method-
mally selected by means of genetic algorithms. Moreover, ologies for example, they are mainly devoted to increase
step signals have also been applied as a load disturbance in the set-point following performance and their inuence on
order to test the load disturbance attenuation capabilities the load disturbance attenuation capabilities is not clear.
and to verify if stability problems may arise. Moreover, although the role of the various fuzzy rules is
generally easy to understand, the physical meaning of the
parameters of the fuzzy inference systems e.g. scaling fac-
1 Introduction tors might not be intuitive and consequently their tuning
could be a dicult task. Finally, it has also to be taken
It is well-known that the most adopted controllers in in- into account that the study of the stability often results
dustrial processes are the Proportional-Integral-Derivative to be a hard task.
PID controllers which can assure satisfactory perfor- The aim of this paper is to analyze and compare these
mances for a large range of processes and, due to the sim- techniques, pointing out these problems, in order to ver-
plicity of their structure, which represent a solution that ify their potential use in industrial settings. To investigate
is dicult to ameliorate from a cost bene t ratio point of the full capacity of the considered methodologies and to
view. However, PID controllers are particularly adequate compare them from an impartial point of view which not
in the control of rst and second order systems, whilst for depends on the operator's skillness in the tuning of the
systems with a more complicated dynamics, better per- overall controller, the fuzzy modules have been optimally
formances can be achieved by using other and more so- tuned by means of genetic algorithms 8 , which have been
phisticated controllers 1 . Unfortunately, real industrial recently recognized to be very eective for this task 9 .
processes have often characteristics such as high-order, The paper is organized as follows. In Section 2 the exam-
dead-time, nonlinearities, and they are generally subject ined methodologies are exposed, whilst in Section 3 the
to load disturbances and measurement noise. Hence, the use of genetic algorithms to optimally x the controller
dierent tuning methods, such as the Ziegler-Nichols one parameters is presented. Results are presented and dis-
2 , that have been devised along the last fty years, can cussed in Section 4. Conclusions are drawn in the last
fail to lead to an acceptable performance. In this case, the section.

*4#/



Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 10,2025 at 07:50:50 UTC from IEEE Xplore. Restrictions apply.
2 Tuning methods based on fuzzy systems involves 14 quantization levels for both e and e.
logic It has to be stressed that the tuning of the three param-
eters k1 , k2 and k3 and of the two scaling factors that
The well-known expression of a PID controller, in the time multiply the two inputs e and e is in general a crucial
domain, is the following: issue, since it is not clear how they inuence the perfor-
Z t 
mances of the overall controller, for a generic system. In
det 1 particular, it has been noticed from the experiments that
ut = Kp et + Td +T e d 1 the coecient k3 is very critical, since if it is not xed at
dt i 0
a very low value it might lead the system to instability.
where et = ysp t , yt is the system error, ut is the
control variable, Kp is the proportional gain, Td is the 2.2 Fuzzy Self-tuning of a Single Param-
derivative time constant and Ti is the integral time con- eter SSP
stant. We can also write 1 as
Z t The method described in the work of He et al. 5 consists
det of parametrizing the Ziegler-Nichols formula by means of
ut = Kp et + Kd + Ki e d 2
dt 0 a single parameter  and then to adopt an on-line fuzzy
inference system to self-tune the parameter. In this way,
where it is obviously Kd = Kp Td derivative gain and the three PID parameters can be expressed as
Ki = Kp =Ti integral gain. The tuning problem consists
of selecting the values of these three parameters. Many Kp = 1:2tku
methods have been provided in the literature for this pur- Ti = 0:75 1+
1
t tu 4
pose see for example the classical works 10, 11, 12. Td = 0:25Ti
Generally, they are devoted to satisfy in particular one where k and t are the ultimate gain and the ultimate
u u
of the dierent specications e.g. set-point following, at- period, respectively,
tenuation of load disturbances, robustness to model un- determined recursively of the process. The value of t is
certainties that are typically given on the control system, with the following equation:

and they provide constant coecients as a result. How- t + 1 = t + ht1 , t for t  0:5
ever, having time-varying gains, depending on the system t + htt for t  0:5
error and its time derivative, is useful to improve per- 5
formances. Fuzzy logic can be exploited in the design where  is a positive constant that has to be chosen in the
of such nonlinear control schemes, since it is a valuable range 0:2 0:6 and ht is the output of the fuzzy infer-
tool to implement the typical operator experience which ence system which has seven membership functions both
determines the fuzzy rules. Specically, the fuzzy rules for each of the two inputs the system error and its deriva-
can be designed in such a way that the controller parame- tive and for the output. It has to be noticed that the ini-
ters are appropriately varied during the transient response tial value of t is set equal to 0.5 which corresponds to
in order to generally decrease the rise time and to limit the Ziegler-Nichols formula. Also for the SSP method, the
the overshoot at the same time. Dierent schemes have manual tuning of the scaling coecient of the fuzzy mod-
been devised in this framework and they will be briey ule and of the parameter  is not an easy task, since no
described in the next. rules of thumb are given and their values can signicantly
vary depending on the system to be controlled.
2.1 Incremental Fuzzy Expert PID con-
trol IFE 2.3 Fuzzy Gain Scheduling FGS
The idea proposed by Tzafestas and Papanikolopoulos 4 In the scheme devised by Zhao et al. 6, the three current
is to slightly modify the values of the three controller pa- PID parameters are determined by the following relation-
rameters determined by the Ziegler-Nichols formula dur- ships:
ing the system operation, i.e. the current gains are given Kp = Kpmax , Kpmin Kp + Kpmin 0

by: Kd = Kdmax , Kdmin Kd + Kdmin 6

P = P + CV fet etg  k1
0

Ki = Kp2 =Kd 
I = I + CV fet etg  k2 3
D = D + CV fet etg  k3 where Kp, Kd and  are determined by means of the fuzzy
0 0

where et is the system error, et is its rate, mechanism and Kpmax , Kpmin , Kdmax and Kdmin are
CV fet etg is the output of the fuzzy inference sys-
adopted to normalize the values of Kp and Kd into the
tem, based on the Macvicar-Whelan fuzzy rule matrix 13, range between zero and one. These are xed by the fol-
which reects the typical action of a human controller. lowing rule of thumb:
Finally, k1 , k2 and k3 are constant parameters that deter- Kpmin = 0:32ku  Kpmax = 0:6ku
7
mine the range of variation of each term. The whole fuzzy Kdmin = 0:08ku tu  Kdmax = 0:15ku tu



Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 10,2025 at 07:50:50 UTC from IEEE Xplore. Restrictions apply.
Seven triangular membership functions are used for the
two inputs et and e_t of the fuzzy module, whilst only Δe
two are used for the outputs Kp and Kd and four single-
0 0
NB NS Z PS PB
tons dene the output of . In this scheme only the two
scaling coe cients of the two inputs of the fuzzy mod- NB NVB NB NM NS Z
ule the system error and its derivative have to be xed
and in case also the shape of the membership functions,
so that setting the overall controller appears to be easier NS NB NM NS Z PS
than in the case of the IFE and SSP systems.
e Z NM NS Z PS PM

2.4 Fuzzy Set-point Weighting FSW PS NS Z PS PM PB

The approach proposed by Visioli and Finzi 7 consists of PB Z PS PM PB PVB

fuzzifying the weight that multiplies the set-point for the
proportional action. In this way, the expression of the PID
whose parameters are calculated with the Ziegler-Nichols Figure 1: Basic rules table of the fuzzy inference system
formula in order to preserve a good load disturbance at- for the FSW scheme.
tenuation is the following:
Rt
ut = Kpbtysp t , yt + Kd dedtt + Ki 0
e d f b

bt = w + f t d * P
dt w

8
y sp e u y
I PLANT

where w is a positive constant parameter less than or equal D

to one and f t is the output of the fuzzy mechanism which
consists of ve triangular membership functions for the
two inputs et and e_t and nine triangular membership Figure 2: The control scheme with the fuzzy set-point
functions for the output. The fuzzy rules are based on the weighting FSW methodology.
Macvicar-Whelan matrix 13 , as shown in Figure 1 the
meaning of the linguistic variables is described in Table 1.
The overall control scheme is shown in Figure 2. In this
methodology, the role of the fuzzy mechanism parameters
3 Optimization by means of ge-
is somewhat intuitive, so that the tuning of the controller netic algorithms
is quite easy. In any case, a simple empirical procedure for
the manual tuning of the fuzzy module has been proposed. In order to investigate the full capabilities of the pro-
Moreover, the methodology is also robust to variation of posed approaches in the set-point following, disregarding
the Ziegler-Nichols parameters, so that its practical im- the practical problems of their implementation, each con-
plementation does not present particular problems. troller has been tuned by means of genetic algorithms,
that stochastically coverge to the global minimum of a
predened performance index. In the simulation exper-
iments performed with Matlab, the Genetic Algorithms
NVB Negative Very Big Optimization Toolbox made by Houck et al. 14 has been
NB Negative Big employed. Specically, for a single control scheme and for
NM Negative Medium each controlled system separately, it has been searched the
NS Negative Small optimal values of the following parameters:
Z Zero  the scaling factors of the fuzzy inference system with
PS Positive Small the exception of the scaling factor of the input et,
PM Positive Medium that has been set equal to the inverse of the set-point
PB Positive Big value,
PVB Positive Very Big
 the peak values of the triangular membership func-
tions they have been shifted from their original po-
Table 1: Meaning of the linguistic variables in the fuzzy sition, maintaining the symmetry with respect to the
inference system. zero value, as shown in Figure 3



Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 10,2025 at 07:50:50 UTC from IEEE Xplore. Restrictions apply.
 others. The resulting IAE for the considered controllers,
for the classical PID tuned with the Ziegler-Nichols Z-N
formula only and with a xed weight b which has been
optimally selected in order to minimize the IAE by means
of a simple iterative procedure has been reported in Ta-
ble 3. The label `Z-N' means that the genetic algorithm
has not found a tuning solution for the fuzzy mechanism
that is able to improve the IAE value obtained applying
Figure 3: The modication of the peak values of the mem- the Ziegler-Nichols formula. In order to better evaluate
bership function. the performances achieved by the controllers, the rise time
and the overshoot resulting for the examined processes are
also reported in Tables 4 and 5, respectively.
any other remaining parameter, whereas it was not It turns out that the FSW controller is superior to the
strictly determined by the methodology. others, since the value of IAE is lower for almost all the
processes. It comes out how in this case the use of a fuzzy
The objective function to be minimized by the genetic inference system signicantly increases the set-point fol-
algorithm is the integrated absolute error lowing performances, since a time-varying set-point weight
Z 1 value generally allows to both reduce the rise time and the
IAE = jysp t ,y t jdt: 9 overshoot with respect to the Ziegler-Nichols tuning.
0 The IFE scheme seems also to perform well, since it gener-
for the unit step response. It has to be noticed that in ad- ally improves the both the classic Ziegler-Nichols and the
dition to the above parametrs, also the tuning of the fuzzy xed set-point weight IAE performances. In particular,
rules has been tried with the genetic algorithm. However, this is due to the fact that the overhoot is limited without
this appears to be necessary, in order to improve the per- signicantly impairing the rise time.
formances, only for the SSP controller, whereas otherwise On the contrary, the FGS does not permit to obtain of-
it results  = 0 that means that the fuzzy mechanism is ten signicantly better results than the ones guaranteed
not able to improve the Ziegler-Nichols response. For the by the xed set-point weight and the SSP controller is
other schemes, the tuning of the rules does not provide yet less eective since it often fails to be better than the
signicant results, so that this procedure is not worthy in Ziegler-Nichols one. It has also to be noticed that look-
general, due to its relevant complexity. ing at the values of the parameter which results from the
genetic algorithm not reported for brevity , it comes out
that they might signicantly dier from one process to an-
4 Simulation results other and it is not generally easy to somehow determine
a relationship between them and the controlled process.
The parameters of the four considered controllers have Hence, the practical implementation of the examined con-
been optimally tuned by means of genetic algorithms, as trollers appears dicult, apart for the FSW scheme for
described in the previous section, evaluating the step re- which a manual tuning procedure has been provided.
sponse of the following processes: To complete the analysis, a step of amplitude equal to
!n 2 three has been applied to the input of the process with
G1 s = s2 +2!s +!n
2 !n = 1  = 0:2 0:8 10 the system at the steady-state without changing the tun-
ing of the dierent controllers. The resulting values of the
G2 s = s 1 1+ s 11 IAE , where the initial time is when the load disturbance
step has been applied, are shown in Table 6 the result
G3 s = 1+e,sTsL2 T = 1 10 L = 0:1 0:4 0:8 12 of the xed b is obviously equal to the one obtained by
Ziegler-Nichols , where the lack of a number means that
G4 s = 1 +1 s 3 13 the system has become unstable in that case. It is neces-
sary to stress that the system has become unstable with
e,sL
G5 s = 1+s1+0:5s1+0 the tuning performed by using the genetic algorithm in
:25s1+0:125s L = 0 0:1 order to optimize the IAE of the set-point step response,
14 so that the system can be stabilized provided that the
1 , 0:5s controller is detuned. However, it is worth noticing that
G6 s = 1 + s 3 : 15
the reported results about the set-point following might
For the sake of clarity in presenting the following results, not be achieved in practical cases hence, it turns out that
the dierent analyzed processes are labeled according to those controllers that present stability problems are again
Table 2. It has to be stressed again that results regarding less eective than stated before. Also from the load dis-
a single plant have been obatined by performing a spe- turbance attenuation viewpoint, it appears the superiority
cic optimization on the same plant, separately from the of the FSW scheme with respect to the others.



Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 10,2025 at 07:50:50 UTC from IEEE Xplore. Restrictions apply.
 Process Z-N Fixed b IFE SSP FGS FSW
Number Process 1 0.46 0.72 0.50 0.87 0.77 0.11
1 G1 s,  = 0:2 2 0.32 0.46 0.34 0.52 Z-N 0.07
2 G1 s,  = 0:8 3 0.96 1.24 0.46 Z-N 0.58 0.20
3 G2 s 4 0.28 0.48 0.22 0.46 Z-N 0.24
4 G3 s, T = 1, L = 0:1 5 0.60 0.84 0.44 0.68 0.46 0.68
5 G3 s, T = 1, L = 0:4 6 0.88 1.02 0.86 1.08 0.82 0.98
6 G3 s, T = 1, L = 0:8 7 0.88 1.58 1.00 Z-N 1.22 0.30
7 G3 s, T = 10, L = 0:1 8 1.78 3.02 2.14 Z-N 2.38 1.16
8 G3 s, T = 10, L = 0:4 9 2.50 4.26 2.02 Z-N 2.82 1.78
9 G3 s, T = 10, L = 0:8 10 0.68 1.48 0.73 Z-N 0.82 0.54
10 G4 s 11 0.56 0.88 0.38 0.46 0.40 0.58
11 G5 s, L = 0 12 0.60 0.80 0.46 0.72 0.56 0.70
12 G5 s, L = 0:1 13 0.80 1.02 0.90 Z-N 0.88 0.92
13 G6 s
Table 4: Value of the rise time s of the unit step response
Table 2: Labeling numbers for the examined processes. for the analyzed controllers. The `Z-N' label means that
the fuzzy module has not improved the Ziegler-Nichols
response.

Process Z-N Fixed b IFE SSP FGS FSW Process Z-N Fixed b IFE SSP FGS FSW
1 3.34 2.23 0.83 1.29 0.93 0.60 1 46 18 0 18 0 18
2 2.12 0.97 0.83 0.87 Z-N 0.56 2 41 6 21 14 Z-N 22
3 2.09 1.44 1.01 Z-N 1.56 0.66 3 31 9 1 Z-N 20 15
4 1.40 0.99 0.95 1.03 Z-N 0.75 4 57 3 41 14 Z-N 1
5 1.79 1.49 1.48 1.51 1.34 1.38 5 37 4 33 13 18 6
6 2.22 2.12 2.18 1.96 2.22 2.10 6 23 8 23 2 22 10
7 4.91 2.50 2.21 Z-N 1.91 0.81 7 66 9 30 Z-N 13 2
8 7.62 4.05 3.38 Z-N 3.47 1.66 8 62 6 11 Z-N 13 3
9 9.38 5.14 4.20 Z-N 4.50 2.60 9 59 5 6 Z-N 13 1
10 2.34 1.89 1.13 Z-N 1.34 0.98 10 49 5 3 Z-N 15 0
11 1.83 1.36 1.01 1.56 1.44 1.22 11 52 6 17 41 15 8
12 1.58 1.41 1.22 1.30 1.31 1.30 12 34 6 15 8 10 6
13 2.86 2.53 2.60 Z-N 2.87 2.43 13 34 17 22 Z-N 15 21

Table 3: Value of IAE of the unit step response for the Table 5: Value of the overshoot  of the unit step
analyzed controllers. The `Z-N' label means that the fuzzy response for the analyzed controllers. The `Z-N' label
module has not improved the Ziegler-Nichols response. means that the fuzzy module has not improved the Ziegler-
Nichols response.



Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 10,2025 at 07:50:50 UTC from IEEE Xplore. Restrictions apply.
 6 Zhao Z.-Y., Tomizuka M., Isaka S., Fuzzy gain
Process Z-N IFE SSP FGS FSW scheduling of PID controllers", IEEE Trans. on Sys-
1 2.80 - - - 0.023 tems, Man, and Cybernetics, 23,1392-1398, 1993.
2 0.39 - - 0.39 0.003
3 7.39 - 7.39 8.82 0.081 7 Visioli A., Finzi G., PID tuning with fuzzy set-point
4 0.26 - - 0.26 0.16 weighting", Proc. of the IEEE Int. Conf. on Control
5 1.38 - - 1.99 1.29 Applications, Trieste I, 638-642, 1998.
6 3.33 3.36 4.21 5.70 3.37 8 Holland J.H., Genetic algorithms and the optimal
7 0.13 0.20 0.13 0.089 0.039 allocations of trials", SIAL Journal of Computing, 2,
8 0.92 - 0.92 0.99 0.44 88-105, 1973.
9 2.04 - 2.04 2.47 1.20
10 1.40 - 1.40 1.74 0.15 9 Homaifar A., McCormick E., Simultaneous Design
11 0.93 - 1.05 1.36 0.83 of membership functions and rule sets for fuzzy con-
12 1.33 - 1.77 3.11 1.39 trollers using genetic algorithms", IEEE Trans. on
13 4.28 - 4.28 6.67 4.32 Fuzzy Systems, 3, 129-139, 1995.
10 Chien K.L., Hrones J.A., Reswick J.B., On the auto-
Table 6: Value of IAE for the response of the load distur- matic control of generalized passive systems", Trans.
bance. The lack of a number means that the system has ASME, 74, 175-185, 1952.
become unstable. 11 Cohen G.H., Coon G.A., Theoretical consideration
of retarded control", Trans. ASME, 75, 827-834,
5 Conclusions 1953.
In this paper, a comparison between dierent tuning 12 Dahlin E.B., Designing and tuning digital con-
methodologies, based on fuzzy logic, for PID controllers trollers", Instruments and control systems, 42, 77-83,
has been performed. The proposed schemes required a 1963.
small extra computational eort, so that their implemen-
tation on a microprocessor does not present any particular 13 Macvicar-Whelan P.J., Fuzzy sets for man-machine
problem. However, the easiness of the tuning of the pa- interaction", Int. J. of Man-Machine Studies, 8, 687-
rameters has been discussed. In order to verify the full po- 697, 1976.
tentiality of the examined techniques, genetic algorithms 14 Houck C., Joines J., Kay M., A genetic algorithm for
have been adopted to optimally select the controller pa- function optimization: a Matlab implementation",
rameters. NCSU-IE Technical Report 95-09, 1995.
In this context, applying the controllers to a large number
of processes, the superiority of the fuzzy set-point weight-
ing method has been shown, since it provides the best
performances both for the step response and for the load
disturbance attenuation, it is easy to tune and it does not
present particular stability problems.

References
1 Astrom K., Hagglund T., PID Controllers: Theory,
Design and Tuning", ISA, 1995.
2 Ziegler J.G., Nichols N.B., Optimum setting for au-
tomatic controllers", ASME Trans., 759-768, 1942.
3 Tzafestas S.G., Fuzzy systems and fuzzy expert con-
trol: an overview", The Knowledge Engineering Re-
view, 9, 229-268, 1994.
4 Tzafestas S.G., Papanikolopoulos N.P., Incremental
Fuzzy Expert PID Control", IEEE Trans. on Indus-
trial Electronics, 37, 365-371, 1990.
5 He S.-Z., Tan S., Xu F.-L., Fuzzy self-tuning of PID
controllers", Fuzzy Sets and Systems, 37-46, 1993.



Authorized licensed use limited to: Universitas Brawijaya. Downloaded on February 10,2025 at 07:50:50 UTC from IEEE Xplore. Restrictions apply.