Professional Documents
Culture Documents
net/publication/235428341
CITATIONS READS
14 76
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Alfonso Baños on 01 June 2014.
Published by:
http://www.sagepublications.com
Additional services and information for Journal of Vibration and Control can be found at:
Subscriptions: http://jvc.sagepub.com/subscriptions
Reprints: http://www.sagepub.com/journalsReprints.nav
Permissions: http://www.sagepub.co.uk/journalsPermissions.nav
Citations http://jvc.sagepub.com/cgi/content/refs/14/9-10/1513
JOAQUÍN CERVERA
ALFONSO BAÑOS
Faculty of Computer Engineering, Computer and Systems Engineering Department, University
of Murcia, Campus de Espinardo, 30100 Murcia, Spain (jcervera@um.es)
(Received 15 November 20051 accepted 4 October 2006)
Abstract: This work focuses on the problem of automatic loop shaping in quantitative feedback theory (QFT),
where the search for an optimum design (a non-convex and nonlinear optimization problem) is traditionally
simplified by linearizing and/or convexifying the problem. In this work, the authors propose a suboptimal
solution using a fixed structure in the compensator. In relation to previous work, the main idea consists in
the study of the use of fractional compensators, which give singular properties to automatically shape the
open loop gain function by using a minimum set of parameters. CRONE controllers, in particular CRONE 2
and CRONE 3, are considered as possible candidate structures, being original structures modified for a better
approach to the QFT theoretical optimum. CRONE 3 non-minimum-phase zeros are avoided by constraints
on the structure parameters.
Keywords: Automatic loop shaping, computer-aided control systems design, QFT, CRONE
1. INTRODUCTION
Quantitative feedback theory (QFT) is a robust frequency domain control design methodol-
ogy which has been successfully applied to practical problems from several domains (Horo-
witz, 1993). One of the key design steps is shaping of the nominal open loop gain function
to a meet set of restrictions (or boundaries) given by the design specifications and the (un-
certain) model of the plant. Although this step has traditionally been performed by hand, the
use of CACSD tools (such as the Matlab QFT toolbox (Borghesani et al., 1995), has made
manual loop shaping much easier. However, the problem of automatic loop shaping is of
enormous practical interest, since manual loop shaping can be difficult for an inexperienced
engineer. It has thus received a considerable attention, especially in the last two decades,
with the development of CACSD tools.
Optimal loop computation is a nonlinear and non-convex optimization problem for which
it is difficult to find a satisfactory solution, since there is no optimization algorithm which
can guarantee a globally optimum solution to such a problem.
One possible approach is to simplify the problem in some way, in order to obtain a sim-
pler optimization problem, for which there exists either a closed solution or an optimization
algorithm which does guarantee a global optimum. A trade-off exists, however, between
(necessarily conservative) simplification of the problem and computational solvability. For
instance, Thompson (1990) convexifies the problem by using rectangular templates and then
solves the resulting nonlinear optimization problem. Gera and Horowitz (1980) linearize the
problem in their semi-automatic iterative algorithm, by considering a linear approximation
of boundaries at each iteration. Some authors have also investigated the loop shaping prob-
lem in terms of particular structures, with a certain degree of freedom, which can be shaped
to the particular problem to be solved (see, e.g., Chait et al., 19991 Fransson et al., 20021
Yaniv and Nagurka, 2004).
Another possibility is to use evolutionary algorithms (Spears et al., 1993), which are
able to adapt to nonlinear and non convex optimization problems. This is the approach
adopted by Chen et al. (1998) and Raimúndez et al. (2001). The primary drawbacks of using
evolutionary algorithms are that they are computationally demanding, and do not, in general,
guarantee an optimal solution.
However, through use of information about the specific problem (instead of a brute force
method) computation effort can be reduced and solutions reasonably close to the optimum
can be obtained. In this sense, a good structure for the compensator, in terms of using a
reduced set of parameters while retaining rich frequency domain behavior, is of crucial im-
portance. This article explores the use of evolutionary algorithms, providing a general and
nonconservative design procedure, combined with a flexible structure, which is able to get
close to the optimal solution using a reduced number of parameters, thus allowing the evo-
lutionary algorithms to show good behavior (i.e., to require lower computation effort). In
previous work, the compensator has been fixed to a rational structure, with a finite (but not
necessarily small) number of zeros and poles. In this work, we introduce the use of a frac-
tional compensator that provides, with a minimum number of parameters, a flexible structure
in the frequency domain in terms of automatic loop shaping. This can be approximated by
a rational compensator, but only by using a considerably larger number of parameters. This
dramatic reduction in the number of parameters has proven to have a substantial effect on the
success of evolutionary algorithms in the resolution of the automatic loop shaping problem.
The key idea behind this structure choice is to use as much a priori information about the
optimum as possible, so that the evolutionary search does not have to try every possibility
from a large set of parameters, but can instead choose a small set of values which parameter-
ize the set of possible controllers according to the information already incorporated.
This work, following on from earlier work by the same authors (Cervera and Baños,
2005, 2006), is a first step towards the use of fractional control ideas in automatic loop
shaping in QFT.
This article will focus on the particular case of minimum phase open loop gain functions,
for which the compensators investigated are able to give a good structure with a reduced set
of parameters. The non-minimum phase case will be considered elsewhere.
The rest of the article is structured as follows. In Section 2 we consider the basics of QFT
(Section 2.1) and CRONE (Section 2.2). In Section 3, the proposed solution is developed,
in terms of a CRONE 2-based controller (Section 3.1) and a CRONE 3-based controller
(Section 3.2)1 in section 3.3 conditions for the prevention of NMP zeros are established.
Section 4 contains a benchmark example problem, which is solved by means of the proposed
method. Finally, Section 5 concludes the article.
2. BACKGROUND
The basic idea in QFT (Horowitz, 1993) is to define and take account of, as part of the control
design process, the quantitative relation between the amount of uncertainty about the system
and the amount of control effort to use. This way, QFT guarantees robust specifications for
every plant belonging to the admissible set defined by the initially specified uncertainty.
The typical QFT control system configuration (Figure 1) considers two degrees of free-
dom: A controller C1s2, in the closed loop, which is designed to ensure the satisfaction of
robust specifications despite uncertainty, and a precompensator, F1s2, designed after C1s2,
which allows the system to achieve the desired closed loop tracking specifications.
For a given plant P1s2, its template 1 is defined as the set of possible values of P1 j32,
allowing for the uncertainty, for a given frequency 3. A nominal plant, P0 2 1 , is chosen.
The design of controller C1s2 is accomplished, in terms of the nominal open loop transfer
function L 0 1s2 3 P0 1s2C1s2, using the Nichols chart. A discrete set of design frequencies
4 is chosen. Given quantitative specifications for the robust stability and robust perfor-
mance of the closed loop system, boundaries 23 , 3 2 4, are computed. 23 defines the
permitted regions for L 0 1 j32 in the Nichols chart1 thus, L 0 1 j32 remaining within 23 im-
plies specification satisfaction by L1 j32 3 P1 j32C1 j32, 4P1s2 2 1 .
The basic step in the design process, loop shaping, consists of the design of a function
L 0 1 j 32 which satisfies the boundaries and is reasonable close to optimum. The QFT op-
timization criterion (Horowitz, 1973) is the minimization of the high frequency gain K h f ,
given by
where n pe is the excess of poles over zeros, which is assumed to be constant for any L1s2 3
P1s2C1s2, P 2 1 . Only loops with the same value of n pe can be compared on the basis of
this definition. The UHFB (universal high frequency boundary) is the robust stability bound-
ary which should be satisfied by L 0 1 j 32, 43. It has been shown (Bode, 19451 Horowitz,
19731 Lurie and Enright, 2000) that, to get close to the optimum, L 0 1 j32 should be as close
as possible to the critical point (0 dB, 71808 ) in the vicinity of the crossover frequency, 3cg ,
and thus it should get as close as possible to the right side and bottom of the UHFB. 3cg is
defined as the first frequency such that L1 j3cg 2 3 0 dB.
Loop shaping is traditionally carried out manually, with the help of CACSD tools, such
as the Matlab QFT toolbox (Borghesani et al., 1995). This leads to a trial and error process,
the quality of the result obtained being strongly determined by the experience and intuition of
the designer. Automatic loop shaping is, consequently, a key issue, and one which, although
it has been addressed, has not yet been satisfactorily solved.
The CRONE approach (Oustaloup, 19911 Oustaloup et al., 1996, 1998, 2000) is the basis for
three generations of fractional controllers based on the use of frequency-band non-integer
differentiators. The fact that these differentiators are defined only within a specified fre-
quency band permits their implementation with a finite number of poles and zeros. The
first and second generations (CRONE 1 and CRONE 2) use real non-integer differentia-
tion, whereas the third generation (CRONE 3) uses complex non-integer differentiation. The
difference between CRONE generations 1 and 2 is that in the first, the fractional structure
defined corresponds to the controller, whereas in the second it corresponds to the open loop
transfer function.
In this article, two CRONE structure based compensators are studied. The first is based
on the structure used by CRONE 2 (and also by CRONE 1). The second compensator is
based on the structure used by CRONE 3. For the sake of completeness, CRONE 2 and
CRONE 3 structures will now be briefly described.
The basic component of the CRONE 2 structure is an order n differentiator
51s2 3 ks n 6 n6 k 2 1 (2)
Additionally, an order n I band limited integrator and an order n F low-pass filter are added,
to provide solution to accuracy, robustness and control effort problems. Thus, the open loop
structure is defined as
1 2
33 4n I 1 9 s n 1
l 3h
51s2 3 k 91 3 4n F 7 (4)
s 1 9 3ls
s
91 3h
The CRONE 3 structure replaces the (real) order n integro-differentiator in equation (4)
with the real part of a (complex) order n 3 a 9 ib integro-differentiator
5 6a9ib
s
D1s2 3 6 a6 b6 3u 2 1 (5)
3u
where C0 3 3h 83u 3 3u 83l and 3u 3 3l 3h . D1s2 can be decomposed into its real and
imaginary parts
where
1 2a 7 1 28
19 s
3h
19 s
3h
Dr 1s2 3 C0 cos b Log C0 7 (8)
19 s
3l
19 s
3l
that the plant uncertainty description used in QFT is richer than either of these conditions,
allowing unstructured and parameterized (structured) uncertainty.
The automatic loop shaping method proposed here consists of adapting CRONE structures
to solve the QFT loop shaping problem. The genetic and evolutionary algorithm toolbox for
use with Matlab (GEATbx, see Pohlheim, 2004) was used for the evolutionary optimization
algorithm.
In the first case, based on CRONE 2, the application of the structure to the QFT problem
is quite straightforward. Two variants of the second case, based on CRONE 3, are produced.
One of these uses the original CRONE 3 structure, while in the other, the original structure
is modified in order to obtain greater flexibility. A detailed study of the parameters involved,
their interrelations, and the values required in order to obtain the desired behavior, is pro-
vided. In particular, the conditions for avoiding non-minimum-phase (NMP) zeros with both
the original and modified CRONE 3 structures are defined in section 3.3.
Structure (4) is intended, in the original CRONE 2 context, to be adjusted in such a way
that its middle term is positioned at 3cg 1 that is, in such a way that 3cg is the geometric
mean of 3l and 3h . This achieves a constant stability phase margin with relation to plant
gain uncertainty (vertical straight line template). This original semantic will be lost, in order
to use this structure to obtain the best possible loop to solve the QFT problem, including
uncertainties given by arbitrary templates.
In order to ease K h f optimization, structure (4) is rewritten as
33 4n I 5 3 9 s 6n 1
l h
51s2 3 k 91 7 (10)
s 3l 9 s 1s 9 3h 2n F
The original third generation CRONE structure (9) is modified to produce a decoupled
CRONE 3 structure, by decoupling the first and second terms in equation (8), in order to
obtain higher shaping flexibility. In CRONE 3, these two terms act together to achieve a
loop consisting, in the frequency band [3l 6 3h ], of an arbitrarily orientable segment. The de-
coupling is done in order to allow different orientations of the loop in two different frequency
bands, so that the loop can follow both the right side and the bottom of the UHFB closely, in
order to minimize K h f .
The frequencies 3h and 3l are decoupled by defining new frequencies 3h and 3l for
the second term, allowing a different frequency range to be defined for each term. C0 in
the second term is decoupled by redefining 9 C0 as C0 3 cC0 , where c is a new free para-
meter, C0 3 3h 83u 3 3u 83l , and 3u 3 3l 3h . The right-hand part of the new complex
differentiator is then
1 2a 7 1 28
1 9 3sh 1 9 3s
Dr 1s2 3 C0 cos b Log cC0 h
(11)
1 9 3sl 1 9 3s
l
where 3h for the last term (3h4 ) has also been decoupled, to provide greater flexibility. The
resulting structure has ten parameters. Some of these parameters can be fixed on the basis
of the specifications or re-expressed in term of others. The remaining parameters can be
manually adjusted (they possess intuitive meanings, which lead to an intuitive iterative tuning
process) or optimized by evolutionary computation. The intuitive meaning of the parameters
is one of the reasons why this CRONE structure was chosen.
The behavior of the modified (real part of) the complex differentiator introduced in equation
(11) is quite complex, and needs to be analyzed in detail so that it can be used to satisfy its
loop shaping purpose. In general, special attention should be paid to the (undesired) appear-
ance of non-minimum-phase behavior for some combination of parameters. In particular,
NMP zeros may appear for certain relations between parameters. We will now derive condi-
tions for avoiding such situations. Since the decoupled CRONE 3 structure is a generalized
version of the original CRONE 3 structure (which is equivalent to setting 3l 3 3l , 3h 3 3h ,
and c 3 1), these results also apply to the original CRONE 3 controller, structure (9).
NMP zeros can only be present in the term
7 1 28
19 s
3h
T3 1s2 3 cos b Log cC0 (13)
19 s
3l
T3 1s2 zeros correspond to numbers z 2 1 such that cos z 3 01 i.e., such that
9
z3 12k 7 126 k 2 2 (14)
2
z satisfies (14) when
9
b ln1x2 3 12k 7 12 (15)
2
where
19 s
3h
x 3 cC0 (16)
19 s
3l
To determine the sign of the zeros given by equation (17), s0 numerator and denominator
signs are studied. The numerator is positive when
9
e 2b 12k712 7 cC0
06 (18)
9
ln1C0 2 12k 7 12 7 ln1c27 (20)
2b
i.e.
5 6
c 9 C0 9
7 e 2b 12k712
0 5 ln 7 12k 7 127 (22)
C0 c 2b
Zeros are NMP (positive) for s values for which numerator and denominator have different
signs. So, defining
9
T 3 12k 7 12 7 ln1c2 (24)
2b
and using equations (20) and (23), zeros are NMP for s values satisfying
7T ln1C0 2 T (25)
or
T ln1C0 2 7T (26)
and
5 6
1 2b
k f loor 3 ln c 9 1 (28)
2 9
Depending on the sign of b, conditions (25) and (26) transform, respectively, into (for
b
0)
9
ln C0 12kceil 7 12 7 ln c (29)
2b
and
9
ln C0 7 12k f loor 7 12 9 ln c (30)
2b
or (for b 0)
9
ln C0 12k f loor 7 12 7 ln c (31)
2b
and
9
ln C0 7 12kceil 7 12 9 ln c7 (32)
2b
These equivalences are inferred as follows. In equation (25) (and comparably (26)),
since ln C0
0, the first inequality is always satisfied. The second inequality is satisfied
(4k 2 2) if and only if it is satisfied by k0 , the k value yielding the smallest positive value of
T (7T for (26)). k0 is computed as follows:
For condition (25),
9
12k 7 12 7 ln c
0 (33)
2b
must be satisfied 4k 2 2.
If b
0: k0 is the minimum k satisfying
5 6
1 2b
k
ln c 9 1 (34)
2 9
that is,
5 6
1 2b
k0 3 ln c 9 1 3 kceil 7 (35)
2 9
that is,
5 6
1 2b
k0 3 ln c 9 1 3 k f loor 7 (37)
2 9
must be satisfied 4k 2 2.
Figure 4. Templates for P1s2 at frequencies 0.1, 0.5, 1, 2, 15 and 100 rad/s.
If b
0: k0 is the maximum k satisfying (36) i.e., k0 3 k f loor .
If b 0, k0 is the minimum k satisfying (34) i.e., k0 3 kceil .
4. DESIGN EXAMPLE
To illustrate the behavior of the proposed design procedure, the QFT Toolbox for Matlab
(Borghesani et al., 1995) benchmark example number 2 is used. This example has previously
been used by Chen et al. (1998) and Raimúndez et al. (2001). In the work of Chen et al.
(1998), two rational loops are designed: A second order loop, with K h f 3 13676 dB, and a
third order one, with K h f 3 130 dB, with n pe 3 3 in both cases. n pe 3 3 will be used here,
to allow comparison. The plant is defined as
ka
13 P1s2 3 6 k 2 [16 10]6 a 2 [16 10] 7
s1s 9 a2
Note that corresponding templates (Figure 4) are two-dimensional (unlike the one-dimensio-
nal segment CRONE templates).
The specification is given by the robust stability specification
P1 j32C1 j 32
1 9 P1 j32C1 j32 1726 4P 2 1 6 3 0
120
Tmax 132 3
1 j323 9 171 j322 9 821 j32 9 120
4 2 0716 0756 16 26 156 100] is established, and 3cg 2 [07016 20] specified.
The result obtained for the CRONE 2-based controller is shown in Figures 6 and 7, with
K h f 3 12975 dB, and using the open loop transfer function
5 61728 5 63716
2588 1765 107 9 s 1
L 0a 1s2 3 2798 10 6
91 7 (39)
s 2588 9 s 1s 9 1765 107 23
Although the result is good (compare with K h f 3 130 dB for the best case from the article
by Chen et al. (1998)), the structure is clearly not flexible enough to permit the loop to get
very close to the UHFB.
5 s 6
1 9 825 1
cos 0734 Log 273 s 3 (40)
1 9 153
s
91 825
with K h f 3 126781 dB1 slightly better than the result achieved using CRONE 2. Figures 10
and 11 show a design using the decoupled CRONE 3 structure
61707 5
5 6175
7 1 9 256785
s
L 0c 3 1746 91 11723
s 1 9 7705
s
5 s 6
1 9 210 1
cos 0745 Log 077 1172
4
3 (41)
1 9 166
s s
91
446
with K h f 3 105732 dB. This result significantly improves on that of the unmodified CRONE
3 design (by more than 20 dB). In Figure 10 it can be seen how the loop tightly follows the
UHFB along its right side and round its bottom right corner.
5. CONCLUSIONS
An automatic loop shaping procedure, using evolutionary algorithms for optimization of the
parameters of a fixed structure, has been proposed. The key idea behind this proposal is the
introduction of a structure with few parameters (necessary in order to get good results from
evolutionary optimization) which is, at the same time, flexible enough, thanks to its fractional
nature, to adapt to imposed performance and stability boundaries, and thus achieve results
which are close to the optimum.
In particular, CRONE 2 and 3 structures have been explored as candidate structures. A
new structure, decoupled CRONE 3, based on the CRONE 3 structure, has been proposed.
Conditions for avoiding NMP zeros in both the original and decoupled CRONE 3 structures
have also been defined.
Acknowledgements. This work was supported in part by Ministero de Educación y Ciencia (Spain) under project DPI
2007-66455-C02-01.
REFERENCES
Bode, H. W., 1945, Network Analysis and Feedback Amplif ier Design, Van Nostrand, New York.
Borghesani, C., Chait, Y., and Yaniv, O., 1995, Quantitative Feedback Theory Toolbox, The MathWorks Inc., Natick,
MA.
Cervera, J. and Baños, A., 2005, “Automatic loop shaping in QFT by using a complex fractional order terms
controller,” in Proceedings of the 7th Quantitative Feedback Theory and Robust Domain Methods Symposium,
Lawrence, KA, August 9–11.
Cervera, J. and Baños, A., 2006, “Automatic loop shaping in QFT by using CRONE structures,” in Proceedings of
the 2nd IFAC Workshop on Fractional Differentiation and its Applications, Porto, Portugal, July 19–21.
Chait, Y., Chen, Q., and Hollot, C. V., 1999, “Automatic loop-shaping of QFT controllers via linear programming,”
ASME Journal of Dynamic Systems, Measurement, and Control 121, 351–357.
Chen, W. H., Ballance, D. J., and Li, Y., 1998, “Automatic loop-shaping in QFT using genetic algorithms,” Technical
report, Centre for Systems and Control, University of Glasgow, Glasgow, UK.
Fransson, C. M., Lennartson, B., Wik, T., Holmström, K., Saunders, M., and Gutman, P. O., 2002, “Global con-
troller optimization using Horowitz bounds,” in Proceedings of the IFAC 15th Triennial World Congress,
Barcelona, Spain, July 21–26.
Gera, A. and Horowitz, I., 1980, “Optimization of the loop transfer function,” International Journal of Control 31,
389–398.
Horowitz, I., 1973, “Optimum loop transfer function in single-loop minimum-phase feedback systems,” Interna-
tional Journal of Control 18, 97–113.
Horowitz, I., 1993, Quantitative Feedback Design Theory - QFT (Vol.1). QFT Press, Boulder, CO.
Lurie, B. J. and Enright, P. J., 2000, Classical Feedback Control with MATLAB, Marcel Dekker, New York.
Oustaloup, A., 1991, La commande CRONE (in French), Hermès, Paris, France.
Oustaloup, A., Moreau, X., and Nouillant, M., 1996, “The CRONE suspension,” Control Engineering Practice 4(8),
1101–1108.
Oustaloup, A., Sabatier, J., and Moreau, X., 1998, “From fractal robustness to the CRONE approach,” in Proceed-
ings of the Colloquium Fractional Differential Systems: Models, Methods and Applications ESAIM Vol. 5,
pp. 177–192.
Oustaloup, A., Melchoir, P., Lanusse, P., Cois, C., and Dancla, F., 2000, “The CRONE toolbox for Matlab,” in Pro-
ceedings of the 11th IEEE International Symposium on Computer Aided Control System Design. Anchorage,
AK.
Pohlheim, H., 2004, “Geatbx: Genetic and evolutionary algorithm toolbox for use with Matlab documentation,”
version 3.50, http://www.geatbx.com.
Raimúndez, C., Baños, A., and Barreiro, A., 2001, QFT controller synthesis using evolutive strategies,” in Pro-
ceedings of the 5th International QFT Symposium on Quantitative Feedback Theory and Robust Frequency
Domain Methods, Pamplona, Spain, pp. 291–296.
Spears, W. M., De Jong, K. A., Bäck, T., Fogel, D. B., and de Garis, H., 1993, “An overview of evolutionary
computation,” in Proceedings of the European Conference on Machine Learning, Vol. 667, pp. 442–459.
Thompson, D. F., 1990, “Optimal and sub-optimal loop shaping in quantitative feedback theory,” Ph.D. Thesis,
School of Mechanical Engineering, Purdue University, West Lafayette, IN.
Yaniv, O. and Nagurka, M., 2005, “Automatic loop shaping of structured controllers satisfying QFT performance,”
ASME Journal of Dynamic Systems, Measurement and Control 127, 472–477.