Proceedings of the 4th World Congress on Intelligent Control 第四届全球智能与自动化大会论文集

and Automation, June 10 – 14, 2002, Shanghai, P.R. China 2002 年 6 月 10 日-14 日，中国上海

A Comparative Introduction of
1
Four Fractional Order Controllers
Dingyü Xue† and YangQuan Chen ‡

†
School of Information Science and Engineering
Northeastern University
Shenyang 110004, P R China
‡
Center for Self-Organizing & Intelligent Systems (CSOIS)
Dept. of Electrical and Computer Engineering
College of Engineering, Utah State University, UT 84322-4160, USA.

Abstract degree of freedom from the use of fractional-order in-

Using the differentiation and integration of fractional tegrator and differentiator made it possible to further
order or non-integer order in systems control is gaining improve the performance of traditional PID controllers.
more and more interests from the systems control com-
munity. In this paper, four representative fractional- In theory, the control systems can include both the frac-
order controllers in the literature are briefly introduced, tional order dynamic system or plant to be controlled
namely, TID (Tilted Proportional and Integral) con- and the fractional-order controller. However, in control
troller, CRONE controller (Contrôle Robuste d’Ordre practice, more common is to consider the fractional-
order controller. This is due to the fact that the plant
Non Entier), PIλ Dµ controller and fractional lead-lag model may have already been obtained as an integer
compensator. The basic ideas and technical formula- order model in classical sense. In most cases, our ob-
tions are presented with some comparative comments. jective is to apply the fractional order control (FOC)
The major purpose of this paper is to draw attention to enhance the system control performance. Therefore,
to the non-conventional way of robust control based on in this paper, we will concentrate on this scenario -
the fractional order calculus. controller being fractional-order.
Key words: Fractional order calculus, fractional or-
der control, Bode’s ideal loop transfer function, robust This paper serves as a comparative introduction to four
controller design. existing schemes: TID controller [11], CRONE con-
troller [12, 13], PIλ Dµ controller [14] and fractional
1 Introduction lead-lag compensator [15]. The basic ideas and techni-
cal formulations are presented with some comparative
Using the notion of fractional-order, it may be a step comments. The major purpose of this paper is to draw
closer to the real world life because the real processes attention to the non-conventional way of robust control
are generally or most likely fractional [1]. However, based on fractional order calculus.
for many of them, the fractionality may be very small.
A typical example of a non-integer (fractional) order This paper is organized as follows. In Sec. 2, we briefly
system is the voltage-current relation of a semi-infinite introduce the definition of fractional order operator and
lossy RC line or diffusion of the heat into a semi-infinite FODE. Four FOC schemes are introduced in Sec. 3 with
solid, where the heat flow q(t) in nature is equal to the some of our comparative comments in Sec. 4. Sec. 5
semi-derivative of the temperature T (t) [2] concludes this paper with some additional remarks on
the possible future developments in FOC.
d0.5 T (t)
= q(t).
dt0.5
Clearly, using an integer order ordinary differential
equation (ODE) description for the above system may
differ significantly to the actual situation. However,
the fact that the integer-order dynamic models are
more welcome is probably due to the absence of solu- 2 Fractional-order Calculus, Fractional-order
tion methods for fractional-order differential equations Differential Equations and Their Laplace
(FODEs). Recently, some progresses in analysis of dy- Transformation
namic systems modeled by FODEs have been made in
[3, 4, 5, 6, 7, 8, 9, 10]. For example, PID controllers,
which have been dominating industrial controllers,
have been modified using the notion of fractional-order Fractional order calculus is as old as the calculus of dif-
integrator and differentiator. It is shown that extra ferentiation. The theory of fractional-order derivative
1 Corresponding author: Dr Xue, Dingyü. Tel. +86-24- was developed mainly in the 19-th century. For more
23914152; E-mail: xue dy@21cn.com. references, see [6, 16, 17, 8].

0-7803-7268-9/01/$10.00 © 2001 IEEE 3228

2.1 Definitions of Fractional-order Differentia- where ak (k = 0, 1, · · · , n) are constant coefficients of
tor the FODE; βk , (k = 0, 1, 2, · · · , n) are real numbers.
A fundamental operator a Dα t , a generalization of differ- Without loss of generality, assume that βn > βn−1 >
ential and integral operators, is introduced as follows: · · · > β1 > β0 ≥ 0.
 α α
 d /dt , ℜ(α) > 0, It is possible to solve the above FODE analytically by
α 1, ℜ(α) = 0, using Mittag-Leffler function in two parameters which
D
a t = (1)
  t (dτ )−α , ℜ(α) < 0, is a generalization of exponential function ez . The
a
Mittag-Leffler function in two parameters is defined by
where α can be a complex number but it is assumed to ∞
a real number in this paper, a is a real number related
 zk
Eα,β (z) = , (α, β > 0). (7)
to initial value which can be usually taken as 0. There Γ(αk + β)
k=0
are two commonly used definitions for the general frac-
tional differentiation and integral, i.e., the Grünwald- Clearly, ez is a particular case of the Mittag-Leffler
Letnikov definition and the Riemann-Liouville defini- function [7]
tion [6, 17, 8]. The Grünwald-Letnikov definition is
∞ ∞
that  zk  zk
E1,1 (z) = = = ez .
[(t−a)/h] Γ(k + 1) k!
k=0 k=0
 
α 1 
j α
D
a t f (t) = lim (−1) (2)
h→0 hα j The analytical solution of the n-term FODE (6) is given
j=0
in general form in [7].
where [·] is a flooring-operator while the Riemann-
Liouville definition is given by 2.4 Laplace Transformation Method
 t The Laplace transform formula for the Riemann-
α 1 dn f (τ ) Liouville fractional derivative (3) has the form [7]:
D
a t f (t) = dτ, (3)
Γ(n − α) dtn a (t − τ )α−n+1  ∞ n−1

α−k−1
e−st 0 Dα α
t f (t) dt = s F (s)− sk 0 Dt f (t) , (8)
0 t=0
for (n − 1 < α < n) where Γ(x) is the well known k=0
Euler’s gamma function. One can observe that by intro- for (n−1 < α ≤ n) where F (s) = L[f (t)] is the normal
ducing notion of the fractional-order operator a Dα
t , the Laplace transformation.
differentiator and integrator can be unified. Therefore,
in this paper, we shall use the term “fractional-order Consider a control function which acts on the FODE
differentiator” or “fractional derivative” alone which system (6) as follows:
should be understood to imply both differentiator and
integrator as shown in (1). an Dβt n y(t) + · · · + a1 Dβt 1 y(t) + a0 Dβt 0 y(t) = u(t). (9)
By Laplace transform, we can get a fractional transfer
2.2 Some Properties of Fractional-order Differ-
function [18]:
entiator
Here we introduce two general properties of fractional Y (s) 1
derivative. The first is the composition of fractional Gp (s) = = . (10)
with integer-order derivative and the second is the U (s) an s + · · · + a1 sβ1 + a0 sβ0
βn

property of linearity. For more properties of fractional

derivative, refer to [17, 8]. In general, a fractional-order dynamic system can be
represented by [7, 19, 20]
The fractional-order derivative commutes with integer-
order derivation [8], Y (jω) bm (jω)αm + · · · +b1 (jω)α1 +b0 (jω)α0
=
U (jω) an (jω)βn + · · · +a1 (jω)β1 +a0 (jω)β0
dn
 n 
p p d f (t)
( D
a t f (t)) = D
a t = a Dtp+n f (t), (4) m
k=0 bk (jω)
αk
dtn dtn = n (11)
βk
k=0 ak (jω)
under the condition t = a one gets f (k) (a) = 0, (k =
0, 1, 2, . . . , n − 1). The relationship (4) says the opera- in frequency domain where s = jω = j(2πf ) and f is the
tors dn /dtn and a Dpt commute. frequency in Hertz. It should be pointed out that, for
fractional-order control, in the literature, discussions in
Similar to the integer-order differentiation, a Dα the frequency-domain dominate.
t is a
linear operator
p
a Dt (λf (t) + µg(t)) = λ a Dpt f (t) + µ a Dpt g(t). (5) 3 Four Fractional Order Controllers

2.3 Linear Fractional-order Differential Equa- The early attempts to apply fractional-order derivative
tions (FODEs) to systems control can be found in [21, 3, 22]. In this
A typical n-term linear FODE in time domain is give section, four representative fractional-order controllers
by in the literature will be briefly introduced, namely, TID
controller, CRONE controller, PIλ Dµ controller and
an Dβt n y(t) + · · · + a1 Dβt 1 y(t) + a0 Dβt 0 y(t) = 0 (6) fractional lead-lag compensator.

3229
3.1 TID Controller
In [11], a feedback control system compensator of the
PID type is provided, wherein the proportional compo-
nent of the compensator is replaced with a tilted com-
1
ponent having a transfer function s− n . The resulting
transfer function of the entire compensator more closely
approximates an optimal transfer function, thereby
achieving improved feedback controller. Further, as
compared to conventional PID compensators, the TID Figure 2: Block diagram of TID control scheme
compensator allows for simpler tuning, better distur-
bance rejection ratio, and smaller effects of plant para-
meter variations on closed loop response. izes the suboptimal response can be closely approxi-
mated by a rational function. As can be seen from
Fig. 1(b), rather sharp corners occur at the sides of
3.1.1 Basic Motivations: The motivation for
the Bode step. Any smoothing of the corners, espe-
TID (Tilted Proportional and Integral) control is from
cially the left one, caused by an improper or inaccurate
the consideration of the so-called theoretically opti-
rational function approximation, reduces the available
mal loop response due to Bode. Consider the con-
feedback, resulting in reduced performance. A typi-
ventional feedback control system block diagram de-
cal loop gain Bode diagram of the system with a PID
scribed by Fig. 1(a) where C is the feedback controller,
compensator is also shown in Fig. 1(b). When provided
yr is the reference input signal, e is the control error
with the same stability margin and the same average
signal , u and y are input and output signals respec-
loop gain as an optimal Bode controller, the crossover
tively. In Fig. 1(a), the additive disturbance is denoted
frequency fc of the PID controller is about one-half
by v. The major goals for the feedback control sys-
that of the optimal Bode loop response. The feedback
tem are to minimize the effect of disturbances at the
at frequency fc /4 is about 10 dB lower than that of a
output of the system, and to minimize sensitivity of
simplified Bode controller. The conventional PID con-
the closed loop response to plant parameter variations.
trollers in common use when applied to a great variety
To satisfy these requirements, the feedback of the sys-
of plants, are easy to tune to provide robust and fairly
tem, properly weighted in frequency, must be maxi-
good performance. However, the performance is not
mized. These constraints uniquely define the optimal
optimal as explained above.
transfer function for the feedback loop. The purpose
for the compensator of the feedback system is to im- The object of TID is to provide an improved feedback
plement a loop response reasonably close to the op- loop compensator having the advantages of the con-
timal one. A commonly-used compensator employed ventional PID compensator, but providing a response
in feedback control systems is a proportional-integral- which is closer to the theoretically optimal response.
derivative (PID) compensator. In fact, a PID controller
provides varying degrees of gain and phase shift of the
signal according to the frequency contents. The con- 3.1.2 Brief Introduction to TID Control
ventional PID compensator transfer function typically Scheme: Similar to PID control, TID scheme is
has two real zeros. Typically, the P-term dominates shown in Fig. 2. where the the proportional compen-
near fc , the D-term dominates at frequencies over 4fc , sating unit is replaced with a compensator having a
1
and the I-term dominates at frequencies up to fc /4, transfer function characterized by 1/s n or s−1/n . This
where fc is the crossover frequency at which loop gain compensator is herein referred to as a ”Tilt” compen-
is 0 dB as shown in Fig. 1(b). sator, as it provides a feedback gain as a function of
frequency which is tilted or shaped with respect to
the gain/frequency of a conventional or positional com-
pensation unit. The entire compensator is herein re-
ferred to as a Tilt-Integral-Derivative (TID) compen-
sator. For the Tilt compensator, n is a nonzero real
number, preferably between 2 and 3. Thus, unlike the
conventional PID controller, wherein exponent coeffi-
cients of the transfer functions of the elements of the
compensator are either 0, -1, or +1, TID scheme ex-
ploits an exponent coefficient of −1/n. By replacing
(a) Block diagram of the (b) Bode plots for PID the conventional proportional compensator with the
classic feedback control controlled plant and the tilt compensator of the invention, an overall response is
system with disturbance ideal loop response achieved which is closer to the theoretical optimal re-
sponse determined by Bode as illustrated in Fig. 1(b).
Figure 1: Classic control system and its ideal Bode plot. In Fig. 2, R(s) is a prefilter provided for proper com-
mand signal prefiltering which is commonly seen in
Referring to Fig. 1(b), a theoretically optimal loop re- practice. A preferred transfer function for the prefilter
sponse has been determined by Bode. For the purpose is
of industrial control, a simplified suboptimal Bode loop s2 + 2ωc s + ωc2
response can be employed. The suboptimal response is R(s) = 2
illustrated in Fig. 1(b) by a solid line. The slope of this s + 5.25ωc s + ωc2
suboptimal gain response is about -10 dB/octave. The Since the T-term eliminates static error, the coefficient
transcendental loop transfer function which character- of the I-term can be set to zero for many problems,

3230
thus simplifying controller tuning. A suggested tuning French abbreviation for “Contrôle Robuste d’Ordre
procedure for the TID compensator is: Non Entier” (which means non-integer order robust
control). In this section, we shall follow the basic con-
cept of fractal robustness, which motivated the CRONE
(a) set KI =0, KD =0, and set the coefficient KT for control, and then mainly focus on the second genera-
the loop gain to be 0 dB at a desired crossover tion CRONE control scheme and its synthesis based
frequency fc ; on the desired frequency template which leads to frac-
(b) set KD such that the phase stability margin at tional transmittance [24, 25].
the crossover frequency is about 5 degrees larger
than desired; and 3.2.1 Fractal Robustness: In [26], “fractal
robustness” is used to describe the following two char-
(1−1/n)
(c) set KI =0.25KT fc . acteristics: the isodamping and the vertical sliding
form of frequency template in the Nichols chart. This
desired robustness motivated the use of fractional-order
Taking n = 1/3 as an example, the transfer function controller in classical control systems to enhance their
1/s1/3 can be approximated by a transfer function hav- performance.
ing alternating real poles and zeros in a complex plane
representation. Three poles and three zeros per decade
generally suffice to achieve the phase error of less than 1. Isodamping lines. Consider the characteristic
1 degree and the amplitude error of less than 0.1 db equation
which is given by α
1 + (τ s) = 0 (12)
.442s6 +2.23s5 +1.86s4 +0.428s3 +.0295s2 where τ is a constant. The two poles are given
+.000568s+2.18 × 10−6 by
T6/6 (s)= . 1
s +2.42s5 +1.304s4 +.201s3 +.0092s2
6
s = e±jπ/α ; (13)
+.0001098s+1.98 × 10−7 τ
for 1 < α < 2. The poles are complex and
conjugated and form a center angle 2Θ with
Enter the coefficients for the above approximated
c Θ = (π − π/α) as shown in Fig. 4(a). Clearly,
transfer function T6/6 (s) for 1/s1/3 into CtrlLAB [23], the poles move at a constant angle (fixed by the
three mouse clicks give the Bode plot, Nichols chart and order α) when τ varies. The robustness in plane s
root locus as shown in Fig. 3. is then illustrated by two half-straight lines which
form the same angle Θ in relation to the real axis
Bode Diagram
and are called isodamping half-straight lines.
30

The natural frequency and the damping ratio are

Magnitude (dB)

Nichols Chart
20 40
0 dB
10

0
30 0.25 dB
0.5 dB
directly deducible from the poles, through their
−10
20 1 dB −1 dB
modulus 1/τ and the half-center angle Θ as fol-
Magnitude (dB)

−6 −4 −2 0 2 4 10 3 dB
10

0
10 10 10
Frequenct ω (Radian)
10 10
0
6 dB −3 dB
−6 dB
lows:
Phase (degrees)

−10 −10 −12 dB

1 1 π 1 π
−20 −20 −20 dB ωp = sin Θ = sin π − = sin (14)
−30 −30
τ τ α τ α
Phase (degree)
−40
10
−6 −4
10
−2
10 10
0 2
10
4
10 −40
−350 −300 −250 −200 −150 −100 −50
−40 dB
0 and
π π
ζ (α) = cos Θ = cos π − = − cos . (15)
(a) Bode plot (b) Nichols chart α α
It can be clearly seen that the damping ratio ζ
is exclusively a function of fractionality order α,
Figure 3: Frequency responses of transfer function T6/6 (s)
thus allowing the introduction of the notion of
robust oscillatory mode.
From the Bode plot, we can see that T6/6 (s) is a good 2. Frequency template.
approximate for 1/s1/3 in both magnitude and in par- With a unit negative feedback, the forward path
ticular the phase (constant phase angle). The vertical transfer function, or open-loop transmittance, for
line in Nichols chart Fig. 3(b), is a desired robustness the characteristic equation (12) is
property for controller design.
 α
In TID patent [11], an analog circuit using op-amps 1 ωu α
β (s) = = , (16)
plus capacitors and resistors is introduced with a de- τs s
tailed component list which is useful in some cases
where the computing power to implementing T3 (s) dig- which is the transmittance of a non integer inte-
itally is not possible. An example is given in [11] to grator in which ωu = 1/τ denotes the unit gain
illustrate the benefits from TID over conventional PID (or transitional) frequency.
in both time and frequency domain. As arg β (jω) = −απ/2 with 1 < α < 2, the
Nichols chart of β (jω) is a vertical straight
3.2 CRONE Controller line between −π/2 and −π. This is illustrated
The CRONE control was proposed by Oustaloup in in Fig. 4(b). When τ , the system parame-
pursuing fractal robustness [12, 13]. CRONE is a ter, changes, the vertical straight line shown in

3231
 Referring to Fig. 1(a), the desired or ideal T (s) and
S (s) are to set as follows:

Y (s) β (s) 1
T (s) = = = (19)
Yr (s) V (s)=0 1 + β (s) 1 + (s/ωu )α
and
(s/ωu )α

Y (s) 1
S (s) = = = . (20)
V (s) Yr (s)=0 1 + β (s) 1 + (s/ωu )α
In tracking, gain reaches a maximum for resonance frequency
π 1/α
ωt = − cos α ωu , (21)
2
(a) Isodamping half- (b) Robustness in and in regulation, gain reaches a maximum for reso-
straight lines Nichols chart nance frequency
π −1/α
Figure 4: Illustrations of fractal robustness. ωr = − cos α ωu . (22)
2
This result reveals the existence of a resonance when
Fig. 4(b) slides. Such a vertical displacement cos (απ/2) < 0, namely for 1 < α < 3 and therefore for
ensures a constant phase margin Φm , and thus the CRONE control since 1 < α < 2. The resonance
correspondingly a constant damping ratio in the ratio in tracking is
time domain.
|T (jωt )| 1
Qt (α) = = , (23)
|T (j0)| sin α π2
In controller design, the objective is to achieve such
a similar frequency behavior, in a medium frequency while the resonance ratio in regulation is
range around ωu , knowing that the closed loop dynamic
behavior is exclusively linked to the open loop behav- |S (jωr )| 1
ior around ωu . Therefore, the ideal controller design Qr (α) = = . (24)
|S (j∞)| sin α π2
comprises

• an open loop Nichols locus which forms a vertical These results show that the resonance ratio depends ex-
straight line segment around ωu for the nominal clusively on control order α, thus allowing the introduc-
parametric state of the plant, called open loop tion of the notion of robust resonance. By observation,
frequency template (or more simply template) it can be found that
(Fig. 4(b)); (ωt ωr )
1/2
= ωu (25)
• and a sliding of the template on itself when there and
exist parameter changes in the plant (assume that Qt (α) = Qr (α) . (26)
the parameter change will lead to gain variations
around ωu ).
From (25)-(26), the resonance frequencies in tracking
and in regulation are symmetrically distributed with re-
Synthesizing such a template defines the non-integer gard to the open loop unit gain frequency while the reso-
approach that the second generation CRONE control nance ratios in tracking and in regulation are identical.
uses.
3.2.3 The Second Generation CRONE
3.2.2 The Second Generation CRONE Control - Design Steps: Usually, descriptive spec-
Control - Basic Concept: For typical disturbed ifications of the open loop behavior (for the nominal
feedback control system as shown in Fig. 1(a), its con- plant) will be given such as
trol performance is fully characterized by the sensitiv-
ity function S (s), also known as the transmittance in
regulation, or the complementary sensitivity function • the accuracy specifications at low frequencies ;
T (s), also known as the transmittances in tracking and
we know that S (s) + T (s) = 1. It is practically true • the vertical template around unit gain frequency
that given the open loop behavior around the unit gain ωu ;
frequency, one can determine the dynamic behavior in • the input sensitivity specifications at high frequen-
closed loop. Therefore, we use the transmittance fre- cies.
quency template , β (s), as shown in Fig. 4, to define
the desired behavior of T (s) or S (s). Let’s choose a
template such that For a stable minimum phase plant, it turns out that
the behavior thus defined can be described by a trans-
β (s) = β (s) ∀ω ∈ [ωA , ωB ] , (17) mittance based on the frequency-limited real non integer
differentiator, i.e.,
where α
ωu ω
b
 n


b  1+(ωu /ωb )2 1+s/ωh
α 
Kh
n
h
β (s) = , α ∈ [1, 2] . (18) β(s)= Kb +1 
1+(ωu /ωh )2 1+s/ωb
 (27)
s 1 1+s/ωh

3232
with 3.3.2 A Simple Controller Synthesis

Kb = 1 + (ωb /ωu ) 2 −1/2
 
and Kh = 1 + (ωu /ωh ) 2 1/2

.
Scheme: Unlike conventional PID controller, there
(28)
is no systematic and yet rigor design or tuning method
In the particular case where transitional frequencies existing for PIλ Dδ controller. Here, a simple scheme
ωb and ωh are sufficiently distant from frequency ωu , based on the dominant root principle to design PIλ Dδ
around this frequency (i.e. ωb ≪ ω ≪ ωh ), β (s) can controller is briefly introduced. The pole distribution
be reduced to transmittance of the characteristic equation of the controlled system
in the complex plane should be located at the desired
β(s) = (ωu /s)α , (29) dominant roots which are designed based on the
control performance requirement. Assume that the
which is the same as that described by the template desired dominant roots are a pair of complex conjugate
(relation (18)). root as follows:
The order α transmittance of relation (27) describes p1,2 = −r ± jω. (33)
the frequency truncation of the template defined by
the transitional frequencies ωb and ωh . This transmit-
tance results from the substitution of the part raised It is clear that the above dominant root defines the sta-
at power α for the transmittance ωb /p which is used in bility measure St and damping measure Tl . In this sim-
the description of the template between frequencies ωA plified situation, the parameters design of the PIλ Dδ
and ωB , as shown in Fig. 4(b). controller can be divided into two steps, i.e.,

Finally, referring to Fig. 1(a), the controller C(s) in

cascade with the plant is synthesized from its frequency 1. The design of Kp . Proportional gain Kp is re-
response according to lated to the static error Et [%], settling time Tr
[sec.], and overshoot Pr [%]. In general, the larger
β (jω) the K, the smaller the control time Tr [s] as well
C (jω) = , (30) as the static error Et [%]. Therefore, Kp can be
G0 (jω)
simply set via
where G0 (jω) denotes the frequency response of the
nominal plant. Kp ≥ (100/Et ).

There are a number of real life applications of CRONE 2. The design of Td , δ, Ti , λ. From the com-
controller such as the car suspension control [27, 13], plex conjugate roots (33), the (required) stability
flexible transmission [12], hydraulic actuator [28] etc. measure St = r and damping measure Tl = r/ω
CRONE control has been evolved to a powerful non- can be computed. Given St and Tl , using classi-
conventional control design tool with a dedicate MAT- cal root locus method, we can numerically solve
LAB toolbox for it [29]. For an extensive overview, Td , δ, Ti , λ from the characteristic equation with
refer to [30] and the references therein. fractional-order controller C(s) which is given by
C(s)P0 (s) + 1 = 0 (34)
3.3 PIλ Dµ Controller
3.3.1 Basic Formulae: PIλ Dµ controller, also where P0 (s) is a nominal model of P (s) as shown
known as PIλ Dδ controller, was studied in time domain in Fig. 1(a). More specifically, for simple plant
in [14] and in frequency domain in [31]. In general form, models, this can now be done by solving
referring to Fig. 1(a), the transfer function of PIλ Dδ is
given by min C(s)P0 (s) + 1s=−r±jω .
Td ,δ,Ti ,λ

U (s)
C(s) = = Kp + Ti s−λ + Td sδ , (31)
E(s) 3.4 Fractional Lead-Lag Compensator
In the previous subsections, fractional controllers are
where λ and δ are positive real numbers; Kp is the directly related to the use of fractional-order differen-
proportional gain, Ti the integration constant and Td tiator or integrator. It is possible to extend the clas-
the differentiation constant. Clearly, taking λ = 1 and sical lead-lag compensator to the fractional-order case
δ = 1, we obtain a classical PID controller. If λ = 0 which was studied in [15]. The fractional lead-lag com-
(Ti = 0) we obtain a PDδ controller, etc. All these pensator is given by
types of controllers are particular cases of the PIλ Dδ  r
controller. The time domain formula is that 1 + s/ωb
Cr (s) = C0 (35)
u(t) = Kp e(t)+Ti D−λ δ (∗) (∗) 1 + s/ωh
t e(t)+Td Dt e(t). (Dt ≡0 Dt ). (32)
The digital realization techniques for the above FOC where 0 < ωb < ωh , C0 > 0 and r ∈ (0, 1).
will be introduced in detail in the next section.
Consider the feedback control loop in Fig.1. A robust
It can be expected that PIλ Dδ controller (32) may en- control problem of interest is to find C guaranteeing
hance the systems control performance due to more robust Q-factor (amplitude magnification factor at the
tuning knobs introduced. Actually, in theory, PIλ Dδ resonance frequency) for the transfer functions from yy
itself is an infinite dimensional linear filter due to the to y and from v to y for all plants in the form P = αP0 ,
fractional order in differentiator or integrator. with P0 the nominal plant model and α ∈ [αm , αM ].

3233
An ideal solution to this problem is to make the The 4/4 fitting result is that
nominal loop transfer function L0 (s) = C(s)P0 (s) =
(s/ω0 )n , where 1 < n < 2. For L(s) = C(s)P (s), its 
1 + 2s
0.65
Nyquist plot is similar to L0 (s). Therefore, the phase C0.65 (s) = 4280.1
margin and Q-factors for the closed-loop transfer func- 1 + 0.005s
tions from yr to y and from v to y will be independent 9.457×10−11 s4 +1.218×10−8 s3 +3.07×10−7 s2
of α and uniquely determined by n. Furthermore, since +1.476×10−6 s+9.794×10−7
the transfer function from yr to y can be approximated ≈
4.5×10−16 s4 +1.161×10−13 s3 +6.99×10−12 s2
by a second-order stable system, the maximum closed- +9.516×10−11 s+2.14×10−10
loop step response overshoot is almost independent of
α. Thus, a change of α will result in a slower or faster with its Bode plot and Nichols chart drawn in CtrlLAB
but equally damped response, which is a desirable ro- in Fig. 5. We can see that Fig. 5 is quite similar to the
bustness property in some applications, such as the car characteristic of a frequency-band fractional differen-
suspension design problem [15]. tiator.

3.4.1 Design Concept: The ideal loop trans- 110

Bode Diagram
120

fer function can be approximated by shaping L0 (jω)

Magnitude (dB)
100
100

close to (jω/ω0 )n around the cross-over frequency ω0 at 90

80
80

which L0 (jω0 ) = 1. To achieve this, when the plant

Magnitude (dB)
70 60
−4 −2 0 2 4 6

P (s) = M sm , it is required that arg(C(jω)) is close

10 10 10 10 10 10
Frequenct ω (Radian) 40
60 0 dB

Phase (degrees)
to m − 2 + φ/π, where φ is the desired phase margin,
dB
20 −1 dB
40

in a given frequency range around ω0 . One could obvi- 20

0
−3 dB
−6 dB
−12 dB

ously seek to achieve this by using the classical rational 0

10
−4 −2
10 10
0
10
2 4
10
6
10
−20
−100 −50 0
−20 dB
50 100

lead-lag filter corresponding to (35) with r = 1 due to Phase (degree)

the fact that arg[C1 (jω)]

√ is close to arg[C1 (jω0 )] around (a) Bode plot (b) Nichols chart
the frequency ω0 = ωb ωh . Clearly, when r = 1, the
width of the frequency range over which the condition
arg[C1 (jω)] ≈ arg[C1 (jω0 )] = m − 2 + φ/π holds is en- Figure 5: Stable minimum phase 4/4 frequency domain
tirely determined by the choice of φ, and thus cannot fitting of C0.65 (s)
be adjusted to meet robustness requirements.

When a non-integer value of r in (35) is used, it is

possible to guarantee robust phase margin, closed-loop
resonance and overshoot for any given range of varia- 4 Comparative Comments
tion of α.
The four representative fractional-order controllers
To apply this control strategy to more general linear briefly introduced in Sec. 3 have their respective advan-
systems, the control C(s) in Fig. 1 should take the form tages. We offer our comparative comments as follows.
C(s) = Cr (s)G(s), where G(s) is to be chosen such that
• In terms of readiness for real applications,
G(s) ≈ M sm /P0 (s). (36) CRONE method is the best choice since it has
a clear design interpretation with connections to
the familiar conventional controller design meth-
ods based on Bode plot and Nichols chart. In ad-
3.4.2 Realization of Fractional Lead-Lag dition, CRONE has many industrial applications
Compensator: A state-space representation of the especially for the car suspension systems. The re-
fractional lead-lag compensator is proposed in [15] to- cently developed MATLAB Toolbox for CRONE
gether with an error bound estimate. However, the design [29] makes it ready for commercial appli-
stable minimum-phase frequency-domain fitting is an cations.
easier and more effective method. • Since PID control is popular in many industry
sections, PIλ Dµ controller should provide addi-
In the suspension controller design example of [15], tional potentials to achieve better performance.
which is cited from [13], the plant is P (s) = 1/(M s2 ). Compared to many well-proven PID parame-
The range of change in M is from 100 kg to 900 kg. ter setting techniques, development of setting or
The parameters [15] of the designed fractional lead- autotuning techniques for the 5 parameters in
lag compensator are: ωb = 0.5, ωh = 200, r = 0.65. PIλ Dµ is strongly desired.
Nominal frequency ω0 = 10 rad./sec. at the nomi-
nal mass M = M0 = 300kg. λ is set to 20 so that • Although TID can be regarded as a special type
ωb = ω0 /λ and ωh = λω0 . C0 is determined from of PIλ Dµ controller, it is observed that a system-
P (jω)Cr (jω) = 1 which gives C0 = M0 ω02 λ−r such atic parameter setting method has been proposed
that ω0 is the gain crossover frequency for the nominal and tested. Therefore, TID should find its wide
case when M = M0 . applications in process control industry.
Here we give out the fitting result for C0.65 (s) using • Lead-lag compensator is also a popular control
the stable frequency fitting method introduced in [32]. system design method. Fractional order lead-lag

3234
 compensator should have its equal value com- theory and group theory to quantum field and string
pared to CRONE or PIλ Dµ controller. However, theories,” Rev. Math. Phys., vol. 11, no. 4, pp. 463–501,
1999.
more intuitive systematic design and parameter [10] P. Zavada, “Operator of fractional derivative in the complex
tuning method are needed. plane,” Commun. Math. Phys., vol. 192, no. 2, pp. 261–
285, 1998.
[11] B. J. Lurie, “Three-parameter tunable tilt-integral-
In summary, the introduction of fractional order calcu- derivative (TID) controller,” US Patent US5371670,
lus idea to conventional controller design extends the 1994.
opportunity of added performance improvement. For [12] A. Oustaloup, B. Mathieu, and P. Lanusse, “The CRONE
example, one may extend the another popular con- control of resonant plants: application to a flexible
transmission,” European Journal of Control, vol. 1,
troller design method - “pole-placement” to “fractional no. 2, 1995.
pole-placement”. The success depends on the system- [13] A. Oustaloup, X. Moreau, and M. Nouillant, “The CRONE
atic FOC tuning/design method. suspension,” Control Engineering Practice, vol. 4, no. 8,
pp. 1101–1108, 1996.
[14] I. Podlubny, “Fractional-order systems and PIλ Dµ -
controllers,” IEEE Trans. Automatic Control, vol. 44,
5 Concluding Remarks no. 1, pp. 208–214, 1999.
[15] H.-F. Raynaud and A. ZergaInoh, “State-space represen-
tation for fractional order controllers,” Automatica,
We have presented an introduction of four repre- vol. 36, pp. 1017–1021, 2000.
sentative fractional-order controllers in the literature, [16] S. G. Samko, A. A. Kilbas, and O. I. Marichev, Fractional
namely, TID (Tilted Proportional and Integral) con- integrals and derivatives and some of their applications.
troller, CRONE controller (Contrôle Robuste d’Ordre Minsk: Nauka i technika, 1987.
[17] K. S. Miller and B. Ross, An Introduction to the Fractional
Non Entier), PIλ Dµ controller and fractional lead-lag Calculus and Fractional Differential Equations. New
compensator. The basic ideas and technical formula- York: Wiley, 1993.
tions are presented with some comparative comments. [18] I. Podlubny, “Fractional-order systems and fractional-order
The major purpose of this paper is to draw attention controllers,” in UEF-03-94, The Academy of Sciences
to the non-conventional way of robust control based on Institute of Experimental Physics, (Kosice, Slovak Re-
public), 1994.
fractional order calculus. [19] L. D. I. Petráš, “The frequency method for stability investi-
gation of fractional control systems,” SACTA journal,
We believe that fractional order control (FOC) can vol. 2, no. 1-2, pp. 75–85, 1999.
benefit control engineering practitioners in a number [20] I. Petras, L. Dorcak, P. O’Leary, V. B. M., and I. Pod-
of ways. A significant benefit may be due to the re- lubny, “The modelling and analysis of fractional-order
control systems in frequency domain,” in Proceedings of
duced number of tuning knobs compared to other ro- ICCC’2000, (High Tatras, Slovak Republic), pp. 261 –
bust control design methods. Further research efforts 264, May 23-26 2000.
will be on the efficient digital implementation methods [21] S. Manabe, “The non-integer integral and its application to
(discretization method) and the investigation the FOC control systems,” JIEE (Japanese Institute of Electrical
performance for systems with actuator nonlinearities. Engineers) Journal, vol. 80, no. 860, pp. 589–597, 1960.
Another possibility for future work is the use of FOC [22] A. Oustaloup, “Linear feedback control systems of fractional
order between 1 and 2,” in Proc. of the IEEE Sympo-
with real or complex time varying order. sium on Circuit and Systems, (Chicago, USA), 4 1981.
[23] D. Xue, Computer aided design of control systems - MAT-
LAB language and its applications. Beijing (in Chi-
References nese): Tsinghua University Press, 1995.
[24] A. Oustaloup, J. Sabatier, and P. Lanusse, “From fractal
[1] P. J. Torvik and R. L. Bagley, “On the appearance of the robustness to CRONE control,” Fractionnal Calculus
fractional derivative in the behavior of real materials,” and applied Analysis, vol. 2, no. 1, pp. 1–30, 1999.
Transactions of the ASME, vol. 51, pp. 294–298, 6 1984. [25] A. Oustaloup and B. Mathieu, La commande CRONE : du
[2] I. Podlubny, L. Dorcak, and J. Misanek, “Application of scalaire au multivariable. Paris: HERMES, 1999.
fractional - order derivatives to calculation of heat load [26] A. Oustaloup, La commande CRONE. Paris: Edition Her-
intensity change in blast furnace walls,” Transactions of mes, 1991.
Tech. Univ. of Kosice, vol. 5, no. 5, pp. 137–144, 1995. [27] A. Oustaloup, “Nouveau systéme de suspension: La suspen-
[3] M. Axtell and E. M. Bise, “Fractional calculus applications sion CRONE,” INPI Patent 90 046 13, 1990.
in control systems,” in Proc. of the IEEE 1990 Nat. [28] P. Lanusse, V. Pommier, and A. Oustaloup, “Fractional
Aerospace and Electronics Conf., (New York, USA), control system design for a hydraulic actuator,” in Proc.
pp. 563–566, 1990. of the 1 st IFAC conference on Mechatronics systems,
[4] L. Dorčák, “Numerical models for simulation the fractional Mechatronic 2000, (Darmstadt, Germany), 9 2000.
- order control systems,” in UEF-04-94, The Academy [29] A. Oustaloup, P. Melchoir, P. Lanusse, C. Cois, and F. Dan-
of Science, Institute of Experimental Physics, (Kosice, cla, “The CRONE toolbox for Matlab,” in Proc. of
Slovak Republic), pp. 1–12, 1994. the 11 th IEEE International Symposium on Computer
[5] D. Matignon, “Stability result on fractional differential Aided Control System Design - CACSD, (Anchorage,
equations with applications to control processing,” in USA), 9 2000.
IMACS-SMC Proceedings, (Lille, France), pp. 963–968, [30] A. Oustaloup, F. Levron, F. Nanot, and B. Mathieu, “Fre-
7 1996. quency band complex non integer differentiator : char-
[6] K. B. Oldham and J. Spanier, The Fractional Calculus. New acterization and synthesis,” IEEE Transactions on Cir-
York: Academic Press, 1974. cuits and Systems I: Fundamental Theory and Applica-
[7] I. Podlubny, “The Laplace transform method for linear dif- tions, vol. 47, pp. 25–40, 1 2000.
ferential equations of the fractional order,” in Proc. of [31] I. Petráš, “The fractional-order controllers: Methods for
the 9th International BERG Conference, (Kosice, Slo- their synthesis and application,” J. of Electrical En-
vak Republic (in Slovak)), pp. 119–119, 9 1997. gineering, vol. 50, no. 9-10, pp. 284–288, 1999.
[8] I. Podlubny, Fractional Differential Equations. San Diego: [32] D. Xue, Y. Q. Chen and D. P. Atherton, Feedback Control
Academic Press, 1999. Systems - analysis and design with MATLAB 6. Ad-
[9] S. C. Woon, “Analytic continuation of operators - operators vanced Textbook Series in Control and Signal Process-
acting complex s - time. applications: From number ing, London: Springer-Verlag (under rev.), 2002.

3235