You are on page 1of 6

Proceedings of the 2nd IFAC

Workshop on Fractional Differentiation and its Applications


Porto, Portugal, July 19-21, 2006

A ROBUST TUNING METHOD FOR


FRACTIONAL ORDER PI CONTROLLERS

YangQuan Chen ,1 Huifang Dou


Blas M. Vinagre Concha A. Monje


Center for Self-Organizing and Intelligent Systems
(CSOIS), Department of Electrical and Computer
Engineering, Utah State University, Logan, UT
84322-4160, USA.

Department of Electronic & Electromechanical
Engineering, Industrial Engineering School, University of
Extremadura, Avda. De, Elvas s/n, 06071-Badajoz, Spain

Abstract: The application of fractional controller attracts more attention in the


recent years. In this paper, a new tuning method for PI controller design is
proposed for a class of unknown, stable, and minimum phase plants. We are able
to design a PI controller to ensure that the phase Bode plot is flat, i.e., the phase
derivative w.r.t. the frequency is zero, at a given gain crossover frequency so that
the closed-loop system is robust to gain variations and the step responses exhibit
an iso-damping property. Several relay feedback tests can be used to identify the
plant gain and phase at the given frequency in an iterative way. The identified
plant gain and phase at the desired tangent frequency are used to estimate the
derivatives of amplitude and phase of the plant with respect to frequency at the
same frequency point by Bodes integral relationship. Then, these derivatives are
used to design a PI controller for slope adjustment of the Nyquist plot to achieve
the robustness of the system to gain variations. No plant model is assumed during
the PI controller design. Only several relay tests are needed.

Keywords: Fractional order controller, PI controller tuning, relay feedback test,


Bodes integral, flat phase condition, iso-damping property.

1. INTRODUCTION Zerganoh, 2000; Podlubny, 1999; Vinagre and


Chen, 2002). This is due to a better understanding
In recent years, an increasing number of stud- of the fractional order calculus potentials revealed
ies can be found related to the application of by many phenomena such as viscoelasticity and
fractional controllers in many areas of science damping, chaos, diffusion and wave propagation.
and engineering (Manabe, 1961; Oustaloup et
al., 1995; Oustaloup et al., 1996; Raynaud and In theory, the control systems can include both
the fractional order dynamic system to be con-
trolled and the fractional-order controller. How-
1 Corresponding author: Dr. YangQuan ever, in control practice, it is more common to
Chen. E-mail: yqchen@ece.usu.edu; Tel. 01- consider the fractional-order controller. This is
435-7970148; Fax: 01-435-7973054. URL: due to the fact that the plant model may have
http://www.csois.usu.edu/people/yqchen. YangQuan
Chen is supported in part by the TCO Bridging Fund of
already been obtained as an integer order model
Utah State University (2005-2006). Blas M. Vinagre is in the classical sense. In most cases, our objective
partially supported by the Research Grant 2PR02A024 is to apply the fractional-order control (FOC) to
(Junta de Extremadura and FEDER).
enhance the system control performance (Vinagre is derived to ensure that the slope of the Nyquist
and Chen, 2002). For example, a generalization curve is equal to the phase of the open loop sys-
of the PID controller, namely the PI D con- tem at a given frequency. Section 3 presents an
troller, involving an integrator of order and a approximation for solving the fractional order .
differentiator of order where and can be The controller design procedure are summarized
real numbers, was proposed in (Podlubny, 1999), in Sec. 4. In Sec. 5, extensive illustrative simula-
where the better response of this type of controller tions are given to demonstrate the effectiveness
was demonstrated in comparison with the classical of the proposed design method. Finally, Sec. 6
PID controller, when used for the control of frac- concludes this paper with some remarks on further
tional order systems. However, in general, there is investigations.
no systematic way for setting the fractional orders
and .
In this paper, we will concentrate on the fractional 2. SLOPE ADJUSTMENT OF PHASE BODE
order PI controller, i.e., PI controller PLOT
1 In this section, we will show how Ki and are
C(s) = Kp (1 + Ki ), (1)
s related under the new condition (2).
where is a real number and (0, 2). For the Substitute s by jw in close loop system (3) so
systematic design of , a new tuning condition, that the close loop system can be written as
called flat phase condition, first proposed in (Chen G(jw) = C(jw)P (jw), where
et al., 2003), will be used which can give a relation- 1
ship between Ki and . Specifically, in addition to C(jw) = Kp (1 + Ki )
(jw)
the gain and phase margin specifications, we pro-
pose to add an additional condition that the phase Ki Ki
Bode plot at a specified frequency wc where the = Kp [(1 +
cos ) j sin ] (4)
w 2 w 2
sensitivity circle tangentially touches the Nyquist
curve is locally flat. When achieved, this new
is the PI controller obtained from (1). The phase
condition will make the system more robust to
of the closed-loop system is given by
gain variations. This additional condition can be
6 G(s)
expressed as d ds |s=jwc = 0 with its equivalent 6 G(jw) = 6 C(jw) + 6 P (jw)
expression given as following:
w+1 sin (+1)
2 + Ki w ( + 1)
dG(s) = 0 + tan1 [ ] .(5)
6 |s=jwc = 6 G(s)|s=jwc , (2) w+1 cos (+1) 2
ds 2

where wc is the frequency at the point of tangency where 0 = 6 P (jw). Then, the derivative of the
where the sensitivity circle tangentially touches closed-loop system G(jw) with respect to w can
the Nyquist curve. In (2), be written as follows:
G(s) = C(s)P (s) (3) dG(jw) dC(jw) dP (jw)
= P (jw) + C(jw) . (6)
dw dw dw
is the transfer function of the open loop system
including the controller C(s) and the plant P (s). From (2), the phase of the derivative of the open
At the first look of (2), it seems complicated since loop system should be known in advance which
the derivative of the phase of the system at wc obviously can not be obtained directly from (6).
has to be known. Fortunately, Bodes integrals So, we need to simplify (6).
(Bode, 1945) can be used to approximate the Consider (6). The PI controller C(jw) is given
derivatives of the amplitude and the phase of by (4) whose derivative with respect to w is that
a system with respect to frequency at a given
frequency. To obtain the approximate derivatives, dC(jw) jKp Ki
= . (7)
the knowledge of the amplitude and the phase of dw (jw)+1
the system at the given frequency together with
the static gain of the system. dP (jw)
For calculation of dw , we have
In practice, wc can be set as the gain crossover
frequency. Our objective in this paper is to devise lnP (jw) = ln|P (jw)| + j 6 P (jw). (8)
a way to retrieve the parameters Kp , Ki and
of the controller C(s) to ensure the flat phase Differentiating (8) with respect to w gives
condition (2). Then, we can adjust Kp to make
the sensitivity circle exactly tangentially touches dlnP (jw) 1 dP (jw)
=
the Nyquist curve on the flat phase. dw P (jw) dw
The remaining part of this paper is organized as dln|P (jw)| d6 P (jw)
follows. In Sec. 2, the relationship between Ki and = +j . (9)
dw dw
Straightforwardly, we arrive at 4

2
dP (jw) dln|P (jw)| d6 P (jw)
= P (jw)[ +j ]. (10) 0
dw dw dw
2

Substituting (4), (7) and (10) into (6) gives 4

dG(jw) Ki sa sp 8
= Kp P (jw)[(1 +
)( + j )
dw (jw) w w 10

jKi 12
], (11)
(jw)+1 14

16
where sa (w) and sp (w), first introduced in (Karimi 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

et al., 2002), are defined as follows:


Fig. 1. vs. . The illustration for zero with
dln|P (jw)| different sp : -0.5, -1, -1.5, -2
sa (w) = w , (12)
dw 1.8
origin
approximation
d6 P (jw) 1.6
sp (w) = w . (13)
dw 1.4

1.2
Hence, the slope of the Nyquist curve at any
specific frequency w is given by 1

dG(jw) ( + 1) 0.8

6 = 0 + tan1 [
dw 2 0.6

sa w sin (+1) + sa Ki + sp w cos (+1)


0.4

2 2 Ki
].(14) 0.2
sa w cos (+1)
2 sp w sin (+1)
2 sp K i
0
3 2.5 2 1.5 1 0.5 0

Following the condition (2), one obtains the rela-


Fig. 2. Comparison of the original zero and sp
tionship between Ki and as follows:
relationship and zero (sp )
w ( + 1) ( + 1)
Ki = [cos 2sp sin + ],(15) are performed. However, from (15), we notice that
2sp 2 2 can not be negative so as to ensure that Ki is
a real number. Therefore, for each sp , must be
where = 2 cos2 (+1) 2 4sp cos (+1)
2 sin (+1)
2 + limited within a more restricted interval. In Fig. 1,
4s2p sin2 (+1)
4s2p . It should be mentioned that it is shown that for different sp , there exists an
2 zero , such that, when > zero , > 0. Clearly,
due to the nature of the quadratic equation, an al-
w (+1) there exist a relationship between zero and sp .
ternative relationship, i.e., Ki = 2sp [cos 2 Using least squares fitting techniques, zero (sp )

2sp sin (+1)
2 ], has been discarded. Also can be approximately expressed by
noted is that in (15) only sp presents. So, there
is no need to compute sa . 2.0093s2p 0.5211sp + 0.0035
zero = . (17)
s2p 0.9359sp + 0.0474
The approximation of sp can be given as follows
(Karimi et al., 2002):
Figure 4 shows that the accuracy of the approx-
d6 P (jw) imate function (17) to the actual zero and sp
sp (w0 ) = w0 |w0
dw relationship is practically acceptable.
2
6 P (jw0 ) + [ln|Kg | ln|P (jw0 )|] (16)
3. PHASE MARGIN ADJUSTMENT
where |Kg | = P (0) is the static gain of the plant, To determine all the three parameters for PI
6 P (jw0 ) is the phase and |P (jw0 )| is the gain controller, we have already established the rela-
of the plant at the specific frequency w0 . For the tionship (15) in the previous section. However, we
systems containing an integrator, because of the still need two other equations.
phase of the integrator is constant and its deriva-
tive is zero, sp should be estimated by using the Assume that the phase of the open loop system at
partial model of the system without the integra- the gain crossover frequency wc is
tor. Note that, the pure time delay has no effect
(+1)
on the estimation of sp . For most of the plants, wc+1 sin2
+ Ki wc
6 G(s)|s=jwc = 0 + tan1 [ ]
sp can be selected between -3 and 0. In general, +1 (+1)
wc cos 2
sp depends on the system dynamics and the fre- ( + 1)
quency at which the simulations or experiments . (18)
2
The corresponding gain is i) Given wc , the gain crossover frequency;
r ii) Given m , the desired phase margin;
Ki 2 Ki 2
|G(jwc )| = Kp |P (jwc )| (1 + cos ) + ( sin ) = 1.(19) iii) From the real plant, obtain the measure-
wc 2 wc 2 ments of 6 P (jwc ) and |P (jwc )| using the
iterative relay tests proposed in (Chen et
Denote m the desired phase margin, i.e., 6 G(s)|s=jwc = al., 2003);
m . Straightforwardly, we have iv) Calculate an estimation of sp (wc ) accord-
controller = m 0 ing to (16);
v) Compute and Ki from (22) and (15),
w+1 sin (+1) + Ki wc ( + 1) respectively;
= tan1 [ c +1 2 (+1) ] .(20) vi) Obtain Kp from (19).
wc cos 2
2
Remark 4.1. Due to the constraint in zero (sp ),
However, it is very complex to solve (20) together wc should not be chosen too aggressively. As
with (15) to get , Kp and Ki . However, from an usual, m should be selected from 30 to 60 .
important observation that by substituting (15)
into (20), controller is the function only of sp and
, not explicitly of w any more, we can proceed
to use the LS fitting again to approximate the 5. ILLUSTRATIVE EXAMPLES
function in (20). We propose to use the following
form of approximation: The PI controller design method presented above
will be illustrated via simulation examples. In
A(sp )2 + B(sp ) + C(sp ) the simulation, the following plants, studied in
controller , (21)
2 + D(sp ) + E(sp ) (Wallen et al., 2002), will be used.
(zero (sp ), 2), 1
Pn (s) = , n = 1, 2, 3, 4; (23)
(s + 1)(n+3)
where A(sp ), B(sp ), C(sp ), D(sp ) and E(sp ) are 1
polynomial functions of sp . Our fitting results are P5 (s) = ; (24)
s(s + 1)3
summarized below for completeness:
1
P6 (s) = es ; (25)
(s + 1)3
A(sp ) = 0.00652s7p 0.07259s6p 0.32682s5p 0.7568s4p
1
0.92446s3p 0.44551s2p + 0.19469sp + 0.00283, P7 (s) = es ; (26)
s(s + 1)3
B(sp ) = 0.0273s7p + 0.30814s6p + 1.41817s5p + 3.42016s4p
1
+4.57371s3p + 3.04877s2p + 0.30284sp 0.01085, P8 (s) = es ; (27)
(s + 1)
C(sp ) = 0.02871s7p 0.32823s6p 1.54191s5p 3.85236s4p

5.52107s3p 4.39267s2p 1.42674sp + 0.01003,


5.1 General plants Pn (s)
D(sp ) = 0.02154s7p + 0.2571s6p + 1.26183s5p + 3.3037s4p

+5.04888s3p + 4.74463s2p + 3.03777sp 2.09475, Let us consider the following fifth order plant first,
i.e., P2 (s), which is also discussed in (Karimi et
E(sp ) = 0.02433s7p 0.29619s6p 1.49144s5p 4.05076s4p
al., 2002). The specifications are set as wc =0.295
6.55861s3p 6.81121s2p 5.17001sp + 0.10642. rad./sec. and m = 45 . The PI controller
designed by using the proposed tuning formulae
So, can be sovled from the following approxi- is C2 (s) = 1.378(1 + s0.168
1.383 ). The Bode and the
mate relationship: Nyquist diagrams are compared in Fig. 3.
A(sp )2 + B(sp ) + C(sp )
= m 0 (22) 1

2 + D(sp ) + E(sp )
200

0 0.5

200
0
400

Clearly, given sp , it is much easier to obtain by 600

3
10
2
10
1
10
0
10
1
10
2
10
3
10
0.5

solving (22) than by solving (20). 100


1.5
200

2
300

Remark 3.1. For (zero (sp ), 2) and sp 400 2.5

(3, 0), the precision of the estimation is found 500


3
10
2
10
1
10
0
10
1
10
2
10
3
10
3
3 2.5 2 1.5 1 0.5 0 0.5 1

to be acceptable via our extensive numerical ex- (a) Bode plot (b) Nyquist plot
periments.
Fig. 3. Bode and Nyquist plots for C2 (s)P2 (s).

4. THE PI CONTROLLER DESIGN From the Bode diagram in Fig. 3(a), it is seen that
the phase curve near the gain crossover frequency
The procedures to determine the PI controller is flat due to the proposed design method. The
parameters are briefly summarized in the follow- phase margin exactly equals 45 . That means
ing: the controller moves the point P (0.295j) of the
Nyquist curve to a point of C(jw)P (jw) on the The other plants shown in (23) have similar sim-
unit circle having a phase of 135 and at the ulation results. We briefly summarized the results
same time makes the Nyquist curve match the as follows for further illustrations:
constraint of (2). A bad new is, from Fig. 3(b), 1
For the fourth order plant: P1 (s) = (s+1) 4 , the
the Nyquist curve of the open loop system is
0.2512
not tangential to the sensitivity circle at the flat proposed controller is 0.695(1 + s1.369 ) with re-
phase region. But if we are allowed to adjust the spect to =0.5, wc =0.374 rad./sec. and m =45 .
open loop gain, we can shift the gain Bode plot The controller designed by the modified Ziegler-
1
get a different gain crossover frequency. Define Nichols method is 0.062(1 + 0.22s ). The results are
the frequency interval corresponding to the flat summarized in Fig. 5.
phase is [wl , wh ]. So, the gain crossover frequency Nyquist Diagram Step Response

wc can be moved in [wl , wh ] by adjusting Kp by


1 1.6

0.5 1.4

wl wh
Kp = Kp where belongs to [ w c
, wc ]. In this 0 1.2

case, setting = 0.5 gives the modified proposed 0.5 1

Imaginary Axis

Amplitude
controller C2
(s) = 0.689(1 + s0.168
1.383 ). For compar-
1

1.5
0.8

0.6

ison, the PI controller designed by the modified 2 0.4

Ziegler-Nichols method is C2mZN (s) = 0.344(1 + 2.5 0.2

1
1.237s ). The Bode plots are compared in Fig. 4(a).
3 0
3 2.5 2 1.5 1 0.5 0 0.5 1 0 20 40 60 80 100 120 140 160
Real Axis Time (sec)

The step responses of the close loop system are (a) Nyquist plots (b) Step responses
compared in Fig. 4(b). Comparing the closed-loop
system with the proposed modified controller to Fig. 5. Comparisons of Bode plots and step re-
that with the modified Ziegler-Nichols controller, sponses for P1 (s) (Bode plots - Dashed line:
the overshoots of the step response from the pro- The modified Ziegler-Nichols, Solid line: The
posed scheme remain invariant under gain vari- proposed; Step responses - Solid line: The
ations. However, the overshoots of the modified modified proposed controller with gain vari-
Ziegler-Nichols controller change remarkably. ations 1, 1.1, 1.3; Dotted line: The modified
Ziegler-Nichols controller with the same gain
200
1.6
Step Response

variations 1, 1.1, 1.3)


0 1.4

1
200

400
1.2
For the sixth order plant: P3 (s) = (s+1) 6 , the
600 1
0.132
3
10
2
10
1
10
0
10
1
10
2
10
3
10 proposed controller is 0.526(1 + s1.385 ) with re-
Amplitude

0.8

100
0.6 spect to =0.4, wc =0.242 rad./sec. and m =45 .
200
0.4
The controller designed by the modified Ziegler-
300
1
400
0.2

0
Nichols method is 0.289(1 + 1.327s ). The results
are summarized in Fig. 6.
500 0 20 40 60 80 100 120 140 160
3 2 1 0 1 2 3
10 10 10 10 10 10 10 Time (sec)

(a) Bode plots (b) Step responses


Nyquist Diagram Step Response
1 1.6

Fig. 4. Comparisons of Bode plots and step re- 0.5

0
1.4

1.2

sponses (Bode plots - Dashed line: The mod- 0.5 1

ified Ziegler-Nichols C2mZN (s)P2 (s), Solid


Imaginary Axis

Amplitude

1 0.8


line: The proposed C2 (s))P2 (s); Step re- 1.5 0.6

sponses - Solid line: The modified proposed


2 0.4

2.5 0.2

controller with gain variations 1, 1.1, 1.3; 3


3 2.5 2 1.5 1 0.5 0 0.5 1
0
0 20 40 60 80 100 120 140 160

Dotted line: The modified Ziegler-Nichols


Real Axis Time (sec)

(a) Nyquist plots (b) Step responses


controller with the same gain variations 1,
1.1, 1.3)
Fig. 6. Comparisons of Bode plots and step re-
In practice, the fractional order integrator in sponses for P3 (s) (Bode plots - Dashed line:
the proposed PI controller can not be exactly The modified Ziegler-Nichols, Solid line: The
achieved since it is an infinite dimensional filter. proposed; Step responses - Solid line: The
A band-limit implementation of the fractional or- modified proposed controller with gain vari-
der integrator is important in practice, i.e., the ations 1, 1.1, 1.3; Dotted line: The modified
finite-dimensional approximation of the fractional Ziegler-Nichols controller with the same gain
order system should be done in a proper range variations 1, 1.1, 1.3)
of frequencies of practical interest (Chen and 1
For the seventh order plant: P4 (s) = (s+1) 7 , the
Moore, 2002). The approximation method we use 0.105
proposed controller is 0.516(1 + s1.389 ) with re-
in this paper is the Oustaloup Recursive Algo-
spect to =0.4, wc =0.206 rad./sec. and m =45 .
rithm (Oustaloup et al., 2000). In our simulations,
The controller designed by the modified Ziegler-
for approximation of the fractional order inte- 1
Nichols method is 0.164(1 + 0.949s ). The results
grator, the frequency range of practical interest
are summarized in Fig. 7.
is selected to be from 0.001Hz to 1000Hz. The
sampling time and the number of the recursive From these general plant class Pn (s), the effec-
zero-pole pairs are assigned as 0.001 sec and 13, tiveness of the proposed PI controller is clearly
respectively. demonstrated.
1
Nyquist Diagram
1.6
Step Response
phase, open loop unstable systems. We are par-
0.5 1.4
ticularly interested in fractional-order PI control
0 1.2
of biomimetic systems.
0.5 1
Imaginary Axis

Amplitude
1 0.8

1.5 0.6

2 0.4
7. ACKNOWLEDGMENTS
2.5 0.2

3
3 2.5 2 1.5 1
Real Axis
0.5 0 0.5 1
0
0 20 40 60 80
Time (sec)
100 120 140 160 The authors are grateful to Professor Li-Chen Fu, Editor-
in-Chief of Asian Journal of Control for providing a com-
(a) Nyquist plots (b) Step responses
plimentary copy of the Special Issue on Advances in PID
Control, Asian J. of Control (vol. 4, no. 4). The simulation
Fig. 7. Comparisons of Bode plots and step re- study was helped by C. H. Hu.
sponses for P4 (s) (Bode plots - Dashed line:
The modified Ziegler-Nichols, Solid line: The
proposed; Step responses - Solid line: The REFERENCES
modified proposed controller with gain vari-
Bode, H. W. (1945). Network Analysis and Feed-
ations 1, 1.1, 1.3; Dotted line: The modified
back Amplifier Design. Van Nostrand. New
Ziegler-Nichols controller with the same gain
York.
variations 1, 1.1, 1.3)
Chen, Y. Q. and K. L. Moore (2002). Discretiza-
5.2 Plant with an integrator P5 (s) tion schemes for fractional-order differentia-
tors and integrators. IEEE Trans. Circuits
(Omitted due to space limit) Syst. I 49, 363367.
Chen, Y. Q., C. H. Hu and K. L. Moore (2003).
Relay feedback tuning of robust PID con-
5.3 Plant with a time delay P6 (s) trollers with iso-damping property. In: Pro-
ceedings of The 42nd IEEE Conference on
(Omitted due to space limit) Decision and Control. Hawaii.
Karimi, A., D. Garcia and R. Longchamp (2002).
PID controller design using Bodes integrals.
5.4 Plant with an integrator and a time delay In: Proceedings of the American Control Con-
P7 (s) ference. Anchorage, AK. pp. 50075012.
Manabe, S. (1961). The non-integer integral and
(Omitted due to space limit) its application to control systems. ETJ of
Japan 6(3-4), 8387.
Oustaloup, A., B. Mathieu and P. Lanusse (1995).
5.5 First order plus time-delay plant P8 (s) The crone control of resonant plants: Applica-
tion to a flexible transmission. Eur. J. Contr.
(Omitted due to space limit) Oustaloup, A., F. Levron, F. Nanot and B. Math-
ieu (2000). Frequency band complex non inte-
ger differentiator: Characterization and syn-
thesis. IEEE Trans. Circuits Syst. I 47, 25
6. CONCLUSION
40.
Oustaloup, A., X. Moreau and M. Nouillant
A new PI tuning method is proposed for a class
(1996). The crone suspension. Control Eng.
of unknown plants in this paper. Given the gain
Pract. 4(8), 11011108.
crossover frequency, the phase margin and with
Podlubny, Igor (1999). Fractional-order systems
an additional condition that the phase Bode plot
and PI D -controllers. IEEE Trans. Auto-
at the specified frequency is locally flat, we can
matic Control 44(1), 208214.
design the PI controller to ensure that the closed-
Raynaud, H. F. and A. Zerganoh (2000). State-
loop system is robust to gain variations and to
space representation for fractional order con-
ensure that the step responses exhibit an iso-
trollers. Automatica 36(7), 10171021.
damping property.
Vinagre, Blas M. and YangQuan Chen (2002).
Comparing with the flat phase PID controller Lecture note on fractional calculus applica-
proposed in (Chen et al., 2003), PI , although tions in automatic control and robotics. In:
also only having three parameters, can achieve The 41st IEEE CDC2002 Tutorial Workshop
very good performance. Most importantly, PI # 2 (Blas M. Vinagre and YangQuan Chen,
can be easily applied for the first order system Eds.). Las Vegas, Nevada, USA. pp. 1310.
while in (Chen et al., 2003), the FOPTD plants [Online]
cannot be well handled. This makes PI more http://mechatronics.ece.usu.edu
advantages in practice because every system can /foc/cdc02 tw2 ln.pdf.
be approximated by the first order plus a time Wallen, A., K. J. Astrom and T. Hagglund (2002).
delay model. Loop-shaping design of PID controllers with
constant ti /td ratio. Asian Journal of Control
Further investigations include the experimen- 4(4), 403409.
tal verification and exploration of nonminimum

You might also like