Documentos de Académico
Documentos de Profesional
Documentos de Cultura
Abstract
This paper deals with the design of fractional order PI l Dm controllers, in which the orders of the integral and derivative parts, l and m,
respectively, are fractional. The purpose is to take advantage of the introduction of these two parameters and fulll additional
specications of design, ensuring a robust performance of the controlled system with respect to gain variations and noise. A method for
tuning the PI l Dm controller is proposed in this paper to fulll ve different design specications. Experimental results show that the
requirements are totally met for the platform to be controlled. Besides, this paper proposes an auto-tuning method for this kind of
controller. Specications of gain crossover frequency and phase margin are fullled, together with the iso-damping property of the time
response of the system. Experimental results are given to illustrate the effectiveness of this method.
r 2007 Elsevier Ltd. All rights reserved.
Keywords: Auto-tuning; PID controller; Fractional order controller; Gain variations; Robust control
related to the applications of fractional order controllers in where means the integer part, while the RL denition is
many areas of science and engineering have led to the Z
a 1 dn t f t
importance of studying aspects such as the analysis, design, a Dt f t dt (2)
Gn a dtn a t tan1
tuning and implementation of these controllers.
Fractional calculus is a generalization of the integration for n 1oaon and where G is Eulers gamma
and differentiation to the non-integer (fractional) order function.
fundamental operator a Dat , where a and t are the limits and For convenience, Laplace domain notion is commonly
a a 2 R is the order of the operation. Among many used to describe the fractional integro-differential opera-
different denitions, two commonly used for the general tion. The Laplace transform of the RL fractional
fractional integro-differential operation are the Grunwal- derivative/integral (2) under zero initial conditions for
dLetnikov (GL) denition and the RiemannLiouville order a 0oao1 is given by
dfa Da a
t f tg s F s. (3)
Corresponding author. Tel.: +34 91 624 8813; fax: +34 91 624 9430. In theory, control systems can include both the
E-mail address: cmonje@ing.uc3m.es (C.A. Monje). fractional order dynamic system to be controlled and the
0967-0661/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2007.08.006
ARTICLE IN PRESS
C.A. Monje et al. / Control Engineering Practice 16 (2008) 798812 799
fractional order controller. However, in control practice, The implementation and application of these fractional
more common is to consider the fractional order controller. order controllers for industrial purposes are other remark-
This is due to the fact that the plant model may have been able aspects aimed in this work, showing the results
already obtained as an integer order model in the classical obtained when testing the controller in different experi-
sense. mental platforms.
In this line, the objective of this work is to apply This paper is organized as follows. First, Section 2
fractional order control (FOC) for industrial applications, shortly reviews the state of the art of FOC and introduces
introducing a fractional order controller to improve the some considerations on the implementation of fractional
system control performance and taking the most of the order controllers. The tuning method proposed for
fractional orders of the controller. fractional order PI l Dm controllers is described in
It is important to realize that there is a very wide range Section 3, showing the results obtained when controlling
of control problems and consequently also a need for a an experimental platform with the controller designed.
wide range of design techniques. There are already many Section 4 presents an auto-tuning method for this kind of
tuning methods available but a replacement of the controller, whose experimental results are also shown in the
ZieglerNichols method is long overdue. On the research section. Finally, some relevant concluding remarks are
side it appears that the development of design methods for presented in Section 5.
integer order control, and specially ProportionalInte-
gralDerivative PID control, is approaching the point of 2. Fractional order control
diminishing returns. There are some difcult problems that
remain to be solved. 2.1. A review
Therefore, this paper proposes the application of
fractional calculus as an alternative option to solve some Even though the idea of fractional order operators is as
of the control problems that can arise when dealing with old as the idea of integer order ones, it has been in the last
industrial applications, as will be commented later. On the decades when the use of fractional order operators and
one hand, a new method for the design of fractional order operations has become more and more popular among
controllers is proposed, and more concretely for the tuning many research areas. The theoretical and practical interest
of a generalized PI l Dm controller of the form: of these operators is nowadays well established, and its
applicability to science and engineering can be considered
ki as an emerging new topic. Even if they can be thought of as
Cs kp k d sm , (4)
sl somehow ideal, they are, in fact, useful tools for both the
description of a more complex reality and the enlargement
where l and m are the fractional orders of the integral and of the practical applicability of the common integer order
derivative parts of the controller, respectively. Since this operators. Among these fractional order operators and
kind of controller has ve parameters to tune operations, the fractional integro-differential operators
(kp ; kd ; ki ; l; m), up to ve design specications for the (fractional calculus) are specially interesting in automatic
controlled system can be met, that is, two more than in the control and robotics, among others, as detailed next.
case of a conventional PID controller, where l 1 and Maybe the rst mention of the interest of considering a
m 1: It is essential to study which specications are more fractional integro-differential operator in a feedback loop,
interesting as far as performance and robustness are though without using the term fractional, was made by
concerned, since it is the aim to obtain a controlled system Bode (1940), and next in a more comprehensive way in
robust to uncertainties of the plant model, load distur- Bode (1945). A key problem in the design of a feedback
bances and high frequency noise. All these constraints will amplier was to come up with a feedback loop so that the
be taken into account in the tuning technique of the performance of the closed loop was invariant to changes in
controller in order to take advantage of the introduction of the amplier gain. Bode presented an elegant solution to
the fractional orders. this robust design problem, which he called the ideal cutoff
On the other hand, another approach of this work refers characteristic, nowadays known as ideal loop transfer
to the auto-tuning of fractional order controllers. As function, whose Nyquist plot is a straight line through the
commented before, nowadays many research efforts related origin giving a phase margin invariant to gain changes.
to the applications of fractional order controllers have Clearly, this ideal system is a fractional integrator with
concentrated on various aspects of control analysis and transfer function Gs ocg =sa , known as Bodes ideal
synthesis. However, in practical industrial settings, a transfer function, where ocg is the gain crossover frequency
similar auto-tuning procedure for this kind of controller and the constant phase margin is jm p ap=2. This
is rarely found but in strong demand. Therefore, the frequency characteristic is very interesting in terms of
ultimate goal is to develop a method to auto-tune a robustness of the system to parameters changes or
generalized PI l Dm controller that allows the fulllment of uncertainties, and several design methods make use of it.
robustness constraints and whose implementation process In fact, the fractional integrator can be used as an
is simple and reliable. alternative reference system for control, considering its
ARTICLE IN PRESS
800 C.A. Monje et al. / Control Engineering Practice 16 (2008) 798812
own properties (Vinagre, Monje, Calderon, Chen, & Feliu, H 2 norm of a fractional SISO system (without applying the
2004). result to the development of controllers), and Petras and
This rst step toward the application of fractional Hypiusova (2002) suggest the tuning of H 1 controllers for
calculus in control led to the adaptation of the FC concepts fractional SISO systems by numerical minimization.
to frequency-based methods. The frequency response and Applications of fractional calculus in control are
the transient response of the non-integer integral (in fact numerous. In Yago Sanchez (1999) the control of
Bodes ideal transfer function) and its application to viscoelastic damped structures is aimed. Control applica-
control systems were introduced by Manabe (1961), and tions to a exible transmission (Oustaloup, Mathieu, &
more recently in Barbosa, Tenreiro, and Ferreira (2003). Lanusse, 1995; Valerio, 2001), an active suspension
Going a step further in automatic control, Oustaloup (Lanusse, Poinot, Cois, Oustaloup, & Trigeassou, 2003),
(1991) studied the fractional order algorithms for the a buck converter (Calderon, 2003; Calderon, Vinagre, &
control of dynamic systems and demonstrated the superior Feliu, 2003) and a hydraulic actuator (Pommier, Musset,
performance of the CRONE (Commande Robuste dOrdre Lanusse, & Oustaloup, 2003) are found in the literature.
Non Entier) method over the PID controller. There are The fractional control of rigid robots is the objective in
three generations of CRONE controllers, and Oustaloup, Fonseca and Tenreiro (2003), Tenreiro and Azenha (1998),
Levron, Nanot, and Mathieu (2000) concentrate on the and the fractional control of a thermal system is the
third generation. Podlubny (1999b) proposed a general- objective in Sabatier and Oustaloup (2003), Petras and
ization of the PID controller, namely the PI l Dm controller, Vinagre (2002), Petras, Vinagre, Dorcak, and Feliu (2002),
involving an integrator of order l and a differentiator of Vinagre, Petras, Merchan, and Dorcak (2001). Besides,
order m. He also demonstrated the better response of this other applications such as the robust control of main
type of controller, in comparison with the classical PID irrigation canals (Feliu, Rivas, & Sanchez, 2007) and
controller, when used for the control of fractional order robustness analysis of a winding system (Laroche &
systems. A frequency domain approach by using fractional Knittel, 2005) can be found.
order PID controllers was also studied in Vinagre, Regarding the implementation of fractional order
Podlubny, Dorcak, and Feliu (2000). controllers, a very good review is given in Valerio (2005)
Further research activities run in order to dene new referring to continuous and discrete approximations of
effective tuning techniques for non-integer order control- fractional order systems. Other related references are Chen
lers by an extension of the classical control theory. To this and Moore (2002), Monje (2006), Oustaloup et al. (2000),
respect, in Caponetto, Fortuna, and Porto (2002,2004) the Podlubny, Petras, Vinagre, OLeary, and Dorcak (2002),
extension of derivation and integration orders from integer Vinagre, Podlubny, Dorcak et al. (2000), Chen, Moore
to non-integer numbers provides a more exible tuning et al. (2004), and Chen, Vinagre et al. (2004).
strategy and therefore an easier achieving of control To sum all this up, it is clear that FOC and its
requirements with respect to classical controllers. In Leu, applications are becoming an important issue. Of course,
Tsay, and Hwang (2002) an optimal fractional order PID there are other published texts related to fractional
controller based on specied gain and phase margins with a calculus. The main reason why they are not cited here is
minimum integral squared error (ISE) criterion is designed. that their subjects are not relevant for the purpose of this
Other works (Vinagre, 2001; Vinagre, Monje, & Calderon, work.
2002) take advantage of the fractional orders introduced in
the control action in order to design a more effective 2.2. Implementation of fractional order controllers
controller to be used in real-life models (see also Chen,
2006). The tuning of integer PID controllers is addressed in Before introducing the essentials of the design method
Barbosa, Tenreiro, and Ferreira (2003,2004a,2004b) by for the fractional order PI l Dm controller, some initial
minimizing a penalty function that reects how far the considerations on its implementation have to be taken into
behavior of the PID is from that of a desired fractional account.
transfer function, and in Chen, Moore, Vinagre, and The generalized transfer function of this controller is
Podlubny (2004) and Chen, Vinagre, and Podlubny (2004) given by
with a somewhat similar strategy. Another approach is the
use of a new control strategy to control rst-order systems ki
Cs kp k d sm . (5)
with long time delay (Chen, Vinagre, & Monje, 2003; sl
Monje, Calderon, & Vinagre, 2002). A robustness con- Next statements are important to be considered. First of
straint is considered in this last work, forcing the phase of all, properly implemented, a fractional integrator of order
the open-loop system to be at at the gain crossover k a; k 2 N; 0oao1; is, for steady-state error cancella-
frequency. tion, as efcient as an integer order integrator of order
Fractional calculus also extends to other kinds of control k 1 (see Axtell & Bise, 1990). However, though the nal
strategies different from PID ones. In what concerns H 2 value theorem states that the fractional system exhibits null
and H 1 controllers, for instance, Malti, Aoun, Cois, steady-state error if a40; the fact of being ao1 makes the
Oustaloup, and Levron (2003) discuss the reckoning of the output converge to its nal value more slowly than in the
ARTICLE IN PRESS
C.A. Monje et al. / Control Engineering Practice 16 (2008) 798812 801
case of an integer controller. Furthermore, the fractional specications can be met, depending on the particular
effect has to be band-limited when it is implemented. requirements of the system. Therefore, the design problem
Therefore, the fractional integrator must be implemented is formulated as follows:
as 1=sa 1=ss1a ; ensuring this way the effect of an
integer integrator 1=s at very low frequency. Phase margin jm and gain crossover frequency ocg
Similarly to the fractional integrator, the fractional specifications: Gain and phase margins have always
differentiator, sm , has also to be band-limited when served as important measures of robustness. It is known
implemented, ensuring this way a nite control effort and that the phase margin is related to the damping of the
noise rejection at high frequencies. system and therefore can also serve as a performance
On the other hand, when fractional order controllers measure (see Franklin, Powell, & Naeini, 1986). The
have to be implemented or simulations have to be equations that dene the phase margin and the gain
performed, fractional transfer functions are usually re- crossover frequency are
placed by integer transfer functions with a behavior close
enough to the one desired, but much easier to handle. jCjocg Gjocg jdB 0 dB, 6
There are many different ways of nding such approxima- argCjocg Gjocg p jm . 7
tions but unfortunately it is not possible to say that one of
them is the best, because even though some of them are Robustness to variations in the gain of the plant: The next
better than others in regard to certain characteristics, the constraint can be considered in this case (see Chen &
relative merits of each approximation depend on the Moore, 2005):
differentiation order, on whether one is more interested
in an accurate frequency behavior or in accurate time dargF s
0. (8)
responses, on how large admissible transfer functions may do oocg
be, and other factors like these. A good review of these
This condition forces the phase of the open-loop system
approximations can be found in Valerio (2005), Vinagre,
F s CsGs to be at at ocg and hence to be almost
Podlubny, Hernandez, and Feliu (2000).
constant within an interval around ocg : It means that
In this work two different ways to approximate
the system is more robust to gain changes and the
fractional order operators to an integer transfer function
overshoot of the response is almost constant within a
have been used: the Oustaloup continuous approximation
gain range (iso-damping property of the time response).
(Oustaloup et al., 2000; Oustaloup, 1995) and a frequency
It must be remarked that the interval of gains for which
identication method performed by the Matlab function
the system is robust is not xed with this condition. That
invfreqs (MathWorks, 2000b). With both methods a
is, the user cannot force the system to be robust for a
rational transfer function is obtained whose frequency
particular gain range. This range depends on the
response ts the frequency response of the original
frequency range around ocg for which the phase of the
irrational transfer function. These two methods are chosen
open-loop system keeps at. This frequency range will
due to their accuracy in the frequency range of interest, and
be longer or shorter, depending on the resulting
any other of the techniques in Valerio (2005), Vinagre,
controller and the plant.
Podlubny, Hernandez et al. (2000) could also be suitable
High frequency noise rejection: A constraint on the
for that purpose.
complementary sensitivity function T can be established:
Once a continuous approximation of the fractional order
operator is obtained, and for the sake of implementation, Tjo CjoGjo pA dB,
the Tustin method with prewarping (Levine, 1996) has been 1 CjoGjo dB
applied in this work for the discretization of the resulting 8oXot rad/s ) jTjot jdB A dB 9
approximation.
whit A the desired noise attenuation for frequencies
3. A tuning method for fractional order PI l Dm controllers oXot rad/s.
To ensure a good output disturbance rejection: A
3.1. Design specifications and tuning problem constraint on the sensitivity function S can be dened:
1
Sjo pB dB,
As commented in the introduction, the objective of this 1 CjoGjo dB
paper is to design a fractional order controller so that the
8opos rad/s ) jSjos jdB B dB 10
system fullls different specications regarding robustness
to plant uncertainties, load disturbances and high fre- with B the desired value of the sensitivity function for
quency noise. For that reason, specications related to frequencies opos rad/s (desired frequency range).
phase margins, sensitivity functions and robustness con- Steady-state error cancellation: As stated before, the
straints are going to be considered in this design method, fractional integrator sl is, for steady-state error
due to their important features regarding performance, cancellation, as efcient as an integer order integrator.
stability and robustness. Of course, other kinds of So, the specication of null steady state-error is fullled
ARTICLE IN PRESS
802 C.A. Monje et al. / Control Engineering Practice 16 (2008) 798812
with the introduction of the fractional integrator, looked for; LB and UB dene a set of lower and upper
properly implemented. bounds on the design variables, X.
In this particular case, the specication in Eq. (6) is taken
as the main function to minimize, and the rest of
Using the fractional order PI l Dm controller of Eq. (5), up specications ((7)(10)) are taken as constrains for the
to ve of these design specications can be fullled, since it minimization, all of them subjected to the optimization
has ve parameters to tune. For fractional order controllers parameters dened within the function FMINCON. The
such as a PI l or a PDm ; three design specications could be success of this design method depends mainly on the initial
met (one for each parameter). Therefore, for the general case conditions considered for the parameters of the controller.
of a PI l Dm controller the design problem is based on solving In Section 4 a different tuning method for this kind of
the system of ve nonlinear equations (given by the controller is proposed in which only the frequency
corresponding design specications) and ve unknown characteristics of the plant at some frequencies of interest
parameters kp , kd , ki , l, m. is enough for the tuning purpose, without considering
However, the complexity of this set of nonlinear initial conditions for the parameters and avoiding the
equations is very signicant, specially when fractional orders nonlinear minimization problem.
of the Laplace variable s are introduced, and nding out the The tuning method proposed here is illustrated next with
solution is not trivial. In fact, a nonlinear optimization the results obtained from an experimental platform
problem must be solved, in which the best solution of a consisting on a liquid level system.
constrained nonlinear equation has to be found.
Global optimization is the task of nding the absolutely
3.2. Experimental results by using the tuning method
best set of admissible conditions to achieve an objective
under given constraints, assuming that both are formulated
The experimental platform Basic Process Rig 38-100
in mathematical terms. Some large-scale global optimiza-
Feedback Unit has been used to test the fractional order
tion problems have been solved by current methods, and a
controllers designed by the optimization tuning method
number of software packages are available that reliably
proposed previously. The platform consists on a low
solve most global optimization problems in small (and
pressure owing water circuit which is bench mounted and
sometimes larger) dimensions. However, nding the global
completely self contained. The water circuit is arranged in
minimum, if one exists, can be a difcult problem (very
front of a vertical panel, as can be seen in Fig. 1.
dependant on the initial conditions). Supercially, global
For the characterization of the plant and implementa-
optimization is a stronger version of local optimization,
tion of the controller a data acquisition board PCL-818H,
whose great usefulness in practice is undisputed. Instead of
by PC-LabCard, has been used, running on Matlab 5.3 and
searching for a locally feasible point one wants the globally
using its real time toolbox Real-Time Windows Target.
best point in the feasible region. However, in many
A computer Pentium II, 350 MHz, 64M RAM, supports
practical applications nding the globally best point,
the data acquisition board and the program in C
though desirable, is not essential, since any sufciently
programming language (from Matlab) corresponding to
good feasible point is useful and usually an improvement
the controller.
over what is available without optimization (this particular
After the characterization of the system the resulting
case). Besides, sometimes, depending on the optimization
transfer function is
problem, there is no guarantee that the optimization
functions will return a global minimum, unless the global k 3:13
Gs eLs e50s , (11)
minimum is the only minimum and the function to ts 1 433:33s 1
minimize is continuous (Pinter, 1996). Taking all these
that is, the liquid level system is modeled by a rst-order
into account, and considering that the set of functions to
transfer function with time delay L 50 s, gain k 3:13
minimize in this case is continuous and can only present
and time constant t 433:33 s. The design specications
one minimum in the feasible region, any of the optimiza-
required for the system are:
tion methods available could be effective, a priori. For this
reason, and taking into account that Matlab is a very
appropriate tool for the analysis and design of control gain crossover frequency, ocg 0:008 rad/s;
systems, the optimization toolbox of Matlab has been used phase margin, jm 60 ;
to reach out the best solution with the minimum error. The robustness to variations in the gain of the plant must be
function used for this purpose is called FMINCON (Math- fullled;
Works, 2000a), which nds the constrained minimum of a sensitivity function: jSjojdB p 20 dB, 8opos
function of several variables. It solves problems of the form 0:001 rad/s;
MINX FX subject to: CX ( 0, C eq X 0, noise rejection: jTjojdB p 20 dB, 8oXot 10 rad/s.
LB ( X ( UB, where F is the function to minimize; C
and C eq represent the nonlinear inequalities and equalities, Applying the optimization method described previously,
respectively (nonlinear constraints); X is the minimum the fractional PI l Dm controller obtained to control the
ARTICLE IN PRESS
C.A. Monje et al. / Control Engineering Practice 16 (2008) 798812 803
system is
0:0100
Cs 0:6152 4:3867s0:4773 . (12)
s0:8968
In this particular case the fractional integral and
derivative parts have been implemented by the Oustaloup
continuous approximation of the fractional integrator
(Oustaloup, 1995; Oustaloup et al., 2000), choosing a
frequency band from 0.001 to 100 rad/s and an order of the
approximation equal to 5 (number of poles and zeros).
Once the continuous fractional controller is obtained, it is
discretized by using the Tustin rule with a sampling time
T s 1 s and a prewarp frequency ocg (Levine, 1996).
The Bode plots of the open-loop system F s CsGs
are shown in Fig. 2. As can be observed, specications of
gain crossover frequency and phase margin are met.
Besides, the phase of the system is forced to be at at ocg
and hence to be almost constant within an interval around
ocg . It means that the system is more robust to gain
changes and the overshoot of the response is almost
constant within this interval, as can be seen in Fig. 3, where Fig. 2. Bode plots of the open-loop system F s CsGs.
a step input of 0.47 has been applied to the closed-loop
system. Variations in the gain of the plant have been
considered from 2:75 to 3:75: The magnitudes of the control laws obtained for each value of gain. As far as the
functions Ss and Ts for the nominal plant are shown in control laws are concerned, only a slight variation in the
Figs. 4 and 5, respectively, fullling the specications. peak value of the signal is produced when the gain changes,
The experimental results obtained when controlling the which is an important feature as far as the saturation of the
liquid level plant in real time are shown next. Fig. 6 shows actuator is concerned. In this case, the peak value is very
the comparison between simulated and experimental levels far from the saturation value of 10 V for the servo valve.
for the nominal gain k 3:13. In Fig. 7 the experimental From these results, the potential of the fractional order
responses for different gains (set by software) are scoped, controllers in practical industrial settings, regarding perfor-
fullling the robustness constraint to gain changes (within mance and robustness aspects, is clear. However, the design
the variation range selected). Fig. 8 shows the experimental method proposed here involves complex equations relating
ARTICLE IN PRESS
804 C.A. Monje et al. / Control Engineering Practice 16 (2008) 798812
Fig. 3. Simulation step responses of the controlled system with controller Fig. 6. Comparison between simulated and experimental levels for k
Cs. 3:13 and controller Cs.
m
m m l2 s 1
PD s kc x . (15)
xl2 s 1
Eq. (14) corresponds to a fractional order PI l controller
and Eq. (15) to a fractional order lead compensator that
can be identied as a PDm controller plus a noise lter. In
this method, the fractional order PI l controller will be used
to cancel the slope of the phase of the plant at the gain
crossover frequency ocg : This way, a at phase around the
frequency of interest is ensured. Once the slope is cancelled,
the PDm controller will be designed to fulll the design
specications of gain crossover frequency, ocg , and phase
margin, jm , following a robustness criterion based on the
atness of the phase curve of this compensator, as will be
explained later. This way, the resulting phase of the open-
loop system will be the attest possible, ensuring the
maximum robustness to plant gain variations.
Let us rstly give some remarks about the relay test used
Fig. 8. Experimental control laws of the controlled system with cont- for the auto-tuning problem.
roller Cs.
4.1. Relay test for auto-tuning
by a skilled engineer or may be acquired automatically by
some kind of experimentation. The second alternative, The relay auto-tuning process has been widely used in
commonly known as auto-tuning, is preferable not only to industrial applications (see Hang, Astrom, & Wang, 2002).
simplify the task of the operator but also for the sake of The choice of relay feedback to solve the design problem is
robustness. justied by the possible integration of system identication
There are a wide variety of auto-tuning methods for and control into the same design strategy, giving birth to
integer controllers. Some of them aim in someway the relay auto-tuning. In this work a variation of the standard
robustness of the controlled system (see Tan, Huang, & relay test is used, shown in Fig. 9, where a delay ya is
Ferdous, 2002), for example, forcing the phase of the open- introduced after the relay function. With this scheme, as
loop system to be at around the crossover frequency so explained in Chen and Moore (2005), the next relations are
that the system is robust to gain variations (see Chen & given:
Moore, 2005; Chen, Moore et al., 2004; Chen, Vinagre
et al., 2004). However, the complexity of the equations argGjoc p oc ya , 16
relating the parameters of the controller increases when pa 1
jGjoc j , 17
some kinds of robustness constraints are required for the 4d Na
controlled system. The implementation of these types of where Gjoc is the transfer function of the plant at the
auto-tuning methods for industrial purposes will be really frequency oc ; which is the frequency of the output signal y
complicate since, in general, industrial devices such as a corresponding to the delay ya , d is the relay output, a is the
PLC cannot solve sets of complex nonlinear equations. amplitude of the output signal (signal y in Fig. 9), and
For that reason, an auto-tuning method for fractional Na is the equivalent relay gain. This way, for each value
order PI l Dm controllers based on the relay test is proposed, of ya a different point on the Nyquist curve of the plant is
that allows the fullment of robustness constrains for the obtained. Therefore, a point on the Nyquist curve of the
controlled system by simple relations among the para- plant at a particular desired frequency oc can be identied,
meters of the controller, simplifying the later implementa-
tion process.
The nal aim is to nd out a method to auto-tune a
fractional order PI l Dm controller formulated as
ml1 s 1 l l2 s 1 m
Cs kc x . (13)
s xl2 s 1
system can be given in the complex plane: system, as it can be seen in (23). Knowing the pair (a; b), the
values of x and l2 are directly obtained by (24), and the
jl2 ocg 1 m
G flat jocg k0 ejpjm compensator design is nished.
jxl2 ocg 1 As shown in Fig. 11, the vector 1 j tan y denes the
jl2 ocg 1 m ejpjm borderline of the lead region. Using the polar form of this
) C 0 jocg a1 jb1
jxl2 ocg 1 G flat jocg k0 vector
p
jl2 ocg 1 1 jy
) a1 jb1 1=m a jb, 23 1 tan2 yejy e , (25)
jxl2 ocg 1 cos y
where k0 kc xm 1 in this case; G flat s is the plant to be and expressing the vector a1 jb1 1=m in its polar form
q
controlled, and (a1 ; b1 ) is called the design point. Parameter 1
x sets the distance between the zero (1=l2 ) and pole (1=xl2 ) a21 b21 1=m ejtan b1 =a1 =m r1=m ejd=m , (26)
q
of the PDm controller, and the value of l2 sets their position where r a21 b21 and d tan1 b1 =a1 ; the following
in the frequency axis. The smaller the value of x, the longer relationships can be established from (23), making (25)
the distance between the zero and pole. These two values equal to (26)
(x; l2 ) depend on the value of m (see Eq. (23)). For a xed
pair x; l2 , the higher the absolute value of m, the higher the d ym, (27)
slope of the magnitude of the PDm controller and the higher m
1 d
the maximum phase that the compensator can give. After r1=m ) 1 r cos .
cos y m
some simple calculations, the expressions for x and l2 can be
given by Then, solving numerically the function 1 rcosd=mm ;
the lead compensation regions in the complex plane for
a1 aa 1 b2 different positive values of m are obtained, as shown in
x ; l2 . (24)
aa 1 b2 boc Fig. 12. The procedure followed to obtain the curves is the
Studying the conditions for a and b to nd a solution, it one described next. For each value of m (a specic curve)
can be concluded that a lead compensator is obtained the pairs (a1 ; b1 ) that form the curve are obtained. Since
q
when a41 and b40, and a lag compensator when
p p r a21 b21 and d tan1 b1 =a1 ; it is clear that a1
1 1 4b2 =2oao1 1 4b2 =2 and 1=2ob r cosd and b1 r sind: Besides, d and r are functions of
o0. Fig. 11 shows these lead and lag regions in the complex m; that is, d ym and r 1= cosd=mm . Now, a vector y is
plane for the integer order compensator C 0 jocg m 1. dened with increasing values in the range 0oyop=2:
Let us focus on the lead compensation. It is clear that for Remember that y denes the borderline of the integer lead
the conventional lead compensator m 1 the vector a region and, therefore, p=2 is its maximum limit. For each
jb a1 jb1 is perfectly known through the knowledge of value of y a pair (d; r is obtained, and then a point (a1 ; b1 )
the plant Gflat jocg (relay test) and the specications of is dened. Repeating these steps for all the values of y, all
phase margin and gain crossover frequency required for the the points (a1 ; b1 ) forming the curve will be obtained. After
Fig. 11. Lead and lag regions for the integer order compensator Fig. 12. Lead regions for the fractional order compensator C 0 joc for
C 0 jocg m 1. 0pmp2.
ARTICLE IN PRESS
808 C.A. Monje et al. / Control Engineering Practice 16 (2008) 798812
that, a new value of m is selected and the process is repeated curve of the compensator (iso-damping constraint). The
to obtain a new curve. value of l2 corresponding to this value mmin is also recorded.
The zone to the right of each curve is the lead region, and
any design point in this zone can be fullled with a Therefore, all the parameters of the PI l Dm controller
fractional order compensator having a value of m equal or have been obtained through this iterative process. Then,
bigger than the one dening the curve which passes the controller is implemented and starts to control the
through the design point (mmin ). For instance, for the process through the switch illustrated in Fig. 9, concluding
design point in Fig. 12, the value of mmin is 0.48. By the auto-tuning procedure.
choosing the minimum value mmin ; the distance between the
zero and the pole of the compensator will be the maximum 4.4. Formulation of the resulting PI tl Dm controller
possible (minimum value of parameter x; a positive value
very close to zero). In this case, the phase curve of the Once the parameters of the fractional order PI l Dm
compensator is the attest possible and variations in a controller of Eq. (13) are obtained by following the design
frequency range centered at ocg will not produce a methods explained above, these parameters can be identied
signicant phase change, improving the robustness of the with those ones of the standard PI l Dm controller given by
open-loop system regarding its iso-damping property. Let m
us remember that the phase of G flat s around ocg was 1 l Tds
C std s kp 1 1 . (28)
already at due to the effect of the controller PI l s and, T is 1 sT d =N
therefore, it is the shape of the phase curve of the fractional
Carrying out some calculations in (28), the following
order lead compensator (PDm ) that affects the robustness
transfer function is obtained:
of the system to gain variations.
Let us then sum up how the PI l Dm controller is auto- kp T i s 1 l T d 1 1=Ns 1 m
tuned. The following steps can be solved by a simple C std s . (29)
T i l s T d =Ns 1
computer, using a data acquisition system to control and
monitor the real process (as explained in the section for Comparing expressions (13) and (28), the relations
experimental results in this paper). A PLC could also be obtained are T i l1 ; kp k0 =l1 l , N 1 x=x and
used for the determination of the parameters of the T d l2 1 x:
controller, due to the simplicity of the equations involved In the next section the auto-tuning method proposed
in the auto-tuning method. here is illustrated by experimental examples of application.
1. Once the specications of design are given (ocg and jm ), 4.5. Experimental results by using the auto-tuning method
the relay test is applied to the plant and the resulting
pairs (yn ,on obtained from the n iterations of the test For the implementation of the auto-tuning method
are saved and used for the calculation of the phase and proposed the following devices have been used, showing
magnitude of the plant at each frequency on (following a connection scheme in Fig. 13:
Eqs. (16) and (17)). As explained previously, these
values are used for the obtaining of the slope of the Data acquisition board AD 512, by Humusoft, running
plant phase u (18). With the value of the slope, the on Matlab 5.3 and using the real time toolbox Real-
parameters l and l1 of the PI l controller are directly Time Windows Target.
obtained by Eqs. (20) and (22). Then, the system
G flat jocg is obtained.
2. Once the system G flat jocg is dened, and according to
Eq. (23), the parameters of the fractional order
compensator in (15) are obtained by simple calculations
summarized next, following the robustness feature
explained in this section.
3. Select a very small initial value of m, for example, m
0:05: For this initial value, calculate the value of x and
l2 using the relations in (23) and (24).
4. If the value of x obtained is negative, then the value of m
is increased a xed step and step 2 is repeated again. The
smaller the xed increase of m the more accurate the
selection of the parameter mmin . Repeat step 2 until the
value of x obtained is positive.
5. Once a positive value of x is obtained, the value of m
must be recorded as mmin . This value of x will be close to
zero and will ensure the maximum atness of the phase Fig. 13. Connection scheme of the experimental platform.
ARTICLE IN PRESS
C.A. Monje et al. / Control Engineering Practice 16 (2008) 798812 809
A computer Pentium II, 350 MHz, 64M RAM, which Measuring experimentally the frequency response of the
supports the data acquisition board and where the system in order to validate these values, a magnitude of
programs run for the implementation of the method 11:8556 dB and a phase of 150:2001 are obtained. So,
proposed. only a slight error is committed in the estimation. Next, a
A servomotor 33-002 by Feedbak, in Fig. 14, that fractional order PI l Dm controller is designed with the
consists of: (a) a mechanical unit 33100, which proposed tuning method to obtain a phase margin jm
constitutes the servo, strictly speaking, (b) an analogue 60 at the gain crossover frequency ou 2:2789 rad/s.
unit 33110, which connects to the mechanical unit The gain of the controller will be xed to 1, that is, k0
through a 34-way ribbon cable which carries all power kc xa 1:
supplies and signals enabling the normal circuit inter- The rst step is the design of the fractional order PI l
connections to be made on the analogue unit and (c) a part, in Eq. (14). For that purpose the slope of the phase of
power supply 01100 for the system. The mechanical the plant, u, is estimated by using the expression (18). The
unit has a brake whose position changes the gain of the slope obtained in this case is u 0:2568 rad=rad/s. With
system, that is, the break acts like a load to the motor. the value of the slope and applying the criterion described
This break will be used to test the robustness of the for the fractional order PI l controller (see Eqs. (20) and
controlled system to gain variations. (22)), the controller that cancels the slope of the phase
curve of the plant is
Fig. 14. Photo of the experimental platform used for the implementation of the auto-tuning method.
ARTICLE IN PRESS
810 C.A. Monje et al. / Control Engineering Practice 16 (2008) 798812
k=2knom
2 k=knom
k=0.5knom
1.5
Control laws
0.5
0.5
0 5 10 15 20 25 30
Time (sec)
Fig. 16. Bode plots of the fractional order controller Cs. Fig. 18. Experimental control laws of the controlled system with
controller Cs.
Step responses of the system with controller CZN(s) parameters of the controller through the knowledge of the
10 magnitude and phase of the plant at the frequency of
k=2knom
9 interest, obtained with the relay test. Specications of gain
k=knom
crossover frequency, ocg , and phase margin, jm , can be
8 k=0.5knom
fullled with a robustness property based on the atness of
7 the phase curve of the open-loop system, guaranteeing the
iso-damping property of the time response of the system to
6
gain variations. Again, the experimental results illustrate
Amplitude
3 Acknowledgment
systems with non-integer derivations. Bordeaux, France: LAP- Petras, I., & Hypiusova, M. (2002). Design of fractional-order controllers
ENSEIRB. via H 1 norm minimisation. Selected Topics in Modelling and Control,
Chen, Y. Q., Vinagre, B. M., & Podlubny, I. (2004). Continued fraction 3, 5054.
expansion approaches to discretizing fractional order derivativesAn Petras, I., & Vinagre, B. M. (2002). Practical application of digital
expository review. Nonlinear Dynamics, 38(14), 155170. fractional-order controller to temperature control. Acta Montanistica
Feliu, V., Rivas, R., & Sanchez, L. (2007). Fractional robust control of Slovaca, 7(2), 131137.
main irrigation canals with variable dynamic parameters. Control Petras, I., Vinagre, B. M., Dorcak, L., & Feliu, V. (2002). Fractional
Engineering Practice, 15(6), 673686. digital control of a heat solid: Experimental results. In International
Fonseca, N. M., & Tenreiro, J. A. (2003). Fractional-order hybrid control Carpathian control conference, Malenovice, Czech Republic
of robotic manipulators. In 11th international conference on advanced (pp. 365370).
robotics, Coimbra, Portugal (pp. 393398). Pinter, J. (1996). Global optimization in action. Dordrecht, The Nether-
Franklin, G., Powell, J., & Naeini, A. (1986). Feedback control of dynamic lands: Kluwer Academic Publishers.
systems. Reading, MA: Addison-Wesley. Podlubny, I. (1999a). Fractional differential equations. San Diego:
Hang, C. C., Astrom, K. J., & Wang, Q. G. (2002). Relay feedback auto- Academic Press.
tuning of process controllersA tutorial review. Journal of Process Podlubny, I. (1999b). Fractional-order systems and PI l Dm controllers.
Control, 12, 143162. IEEE Transactions on Automatic Control, 44(1), 208214.
Lanusse, P., Poinot, T., Cois, O., Oustaloup, A., & Trigeassou, J. (2003). Podlubny, I., Petras, I., Vinagre, B. M., OLeary, P., & Dorcak, L. (2002).
Tuning of an active suspension system using a fractional controller and Analogue realizations of fractional-order controllers. Nonlinear
a closed-loop tuning. In 11th international conference on advanced Dynamics, 29(14), 281296.
robotics, Coimbra, Portugal (pp. 258263). Pommier, V., Musset, R., Lanusse, P., & Oustaloup, A. (2003). Study of
Laroche, E., & Knittel, D. (2005). An improved linear fractional model for two robust control for an hydraulic actuator. In ECC 2003: European
robustness analysis of a winding system. Control Engineering Practice, control conference 2003, Cambridge, UK.
13(5), 659666. Sabatier, J., & Oustaloup, A. (2003). Implementation of a thermal
Leu, J. F., Tsay, S. Y., & Hwang, C. (2002). Design of optimal fractional- platform for the test of non-integer systems. In Conference on
order PID controllers. Journal of the Chinese Institute of Chemical education of technologies and information and systems sciences
Engineers, 33(2), 193202. (pp. 361364). Toulouse, France: Univerite Paul Sabatier.
Levine, W. S. (1996). The control handbook. Boca Raton, New York: CRC Tan, K. K., Huang, S., & Ferdous, R. (2002). Robust self-tuning
Press, IEEE Press. PID controller for nonlinear systems. Journal of Process Control, 12(7).
Malti, R., Aoun, M., Cois, O., Oustaloup, A., & Levron, F. (2003). H 2 Tenreiro, J. A., & Azenha, A. (1998). Fractional-order hybrid control of
norm of fractional differential systems. In ASME 2003 design robot manipulators. In 1998 IEEE international conference on systems,
engineering technical conferences and computers and information in man and cybernetics: Intelligent systems for humans in a cyberworld,
engineering conference, Chicago, USA. San Diego, California, USA (pp. 788793).
Manabe, S. (1961). The non-integer integral and its application to control Valerio, D. (2001). Fractional order robust control: An application. In
systems. ETJ of Japan, 6(34), 8387. Student forum, Porto (pp. 2528).
MathWorks. (2000a). Matlab optimization toolbox, version 2.1. Users Valerio, D. (2005). Fractional robust system control, Ph.D. thesis, Instituto
guide. The MathWorks, Inc. Superior Tecnico, Universidade Tecnica de Lisboa.
MathWorks. (2000b). Matlab signal processing toolbox. Users guide. The Vinagre, B., Podlubny, I., Dorcak, L., & Feliu, V. (2000). On fractional
MathWorks, Inc. PID controllers: A frequency domain approach. In IFAC workshop on
Monje, C. A. (2006). Design methods of fractional order controllers for digital control. Past, present and future of PID control, Terrasa, Spain
industrial applications, Ph.D. thesis, Escuela de Ingenieras Indus- (pp. 5358).
triales, Universidad de Extremadura, Badajoz, Spain. Vinagre, B., Podlubny, I., Hernandez, A., & Feliu, V. (2000). Some
Monje, C. A., Calderon, A. J., & Vinagre, B. M. (2002). PI vs fractional approximations of fractional order operators used in control theory
DI control: First results. In CONTROLO 2002: 5th Portuguese and applications. Fractional Calculus and Applied Analysis, 3(3),
conference on automatic control, Aveiro, Portugal (pp. 359364). 231248.
Monje, C. A., Calderon, A. J., Vinagre, B. M., Chen, Y. Q., & Feliu, V. Vinagre, B. M. (2001). Modelling and control of dynamic systems
(2004). On fractional PI l controllers: Some tuning rules for characterized by integro-differential equations of fractional order,
robustness to plant uncertainties. Nonlinear Dynamics, 38(14), Ph.D. thesis, UNED, Madrid, Spain.
369381. Vinagre, B. M., Monje, C. A., & Calderon, A. J. (2002). Fractional order
Monje, C. A., Vinagre, B. M., Calderon, A. J., Feliu, V., & Chen, Y. Q. systems and fractional order control actions. In 41st conference on
(2005). Auto-tuning of fractional leadlag compensators. In 16th IFAC decision and control. Tutorial workshop 2: Fractional calculus applica-
world congress, Prague. tions in automatic control and robotics, Las Vegas.
Oustaloup, A. (1991). CRONE control: Robust control of non-integer Vinagre, B. M., Monje, C. A., Calderon, A. J., Chen, Y. Q., & Feliu, V.
order. Paris: Hermes. (2004). The fractional integrator as a reference function. In First IFAC
Oustaloup, A. (1995). Non-integer derivation. Paris: Hermes. workshop on fractional differentiation and its applications (pp. 150155)
Oustaloup, A., Levron, F., Nanot, F., & Mathieu, B. (January 2000). Bordeaux, France: ENSEIRB.
Frequency-band complex noninteger differentiator: Characterization Vinagre, B. M., Petras, I., Merchan, P., & Dorcak, L. (2001). Two digital
and synthesis. IEEE Transactions on Circuits and Systems I: Funda- realizations of fractional controllers: Application to temperature
mental Theory and Applications, 47 (1), 2540. control of a solid. In ECC 2001: European control conference 2001,
Oustaloup, A., Mathieu, B., & Lanusse, P. (1995). The CRONE control of Porto, Portugal (pp. 17641767).
resonant plants: Application to a exible transmission. European Yago Sanchez, B. S. (1999). Fractional-PID control for active reduction of
Journal of Control, 1(2), 113121. vertical tail buffeting. Technical Report, Saint Louis University.