Robust Loop Shaping Controller Design for Spectral Models by

Quadratic Programming
Gorka Galdos, Alireza Karimi and Roland Longchamp

Abstract A quadratic programming approach is proposed are not satisfactory. There are some attempts to modify the
to tune fixed-order linearly parameterized controllers for stable Ziegler-Nichols tuning algorithm which are reported in [2].
LTI plants represented by spectral models. The method is Recently, some iterative methods using specific points of
based on the shaping of the open-loop or closed-loop frequency
functions in the Nyquist diagram. The quadratic error between the frequency response function have been developed. A
a desired open loop transfer function and the actual open loop PI controller tuning method achieving specified maximum
frequency function is minimized in the frequency domain sub- sensitivity and phase margin using a Phase Locked Loop
ject to linear constraints guaranteeing stability and robustness (PLL) identifier module for measuring some frequency points
margins by quadratic programming. Moreover, it is shown that is presented in [3]. A PID controller tuning technique based
the H infinity mixed sensitivity robust performance problem
can be approximated by linear constraints and be integrated on the minimization of the sum of square errors between
in the control design method. The method can directly consider the desired and measured specifications (gain margin, phase
multi-model as well as frequency-domain uncertainties. An margin, maximum sensitivity and crossover frequency) has
application to a difficult benchmark problem illustrates the been proposed in [4], [5] based on simple relay experiments.
effectiveness of the proposed approach. A linear quadratic control criterion in frequency domain
I. I NTRODUCTION is minimized iteratively using only the spectral models of
the closed-loop system in [6]. At each iteration the closed-
Most controller design methods are based on plant para-
loop system (with the controller from previous iteration) is
metric models. A parametric model can be obtained ei-
excited with a reference signal and the gradient and Hessian
ther by first principle modeling or by parameter estimation
of the criterion are estimated using the spectral models
techniques using measured data. However, it is usually too
identified by the measured data. It should be noted that all
difficult or time consuming to obtain a parametric model
of the mentioned iterative methods use the Gauss-Newton
based on physical laws. On the other hand, identification of
algorithm and consequently they converge to a local optimum
parametric models is based on several a priori information
of their criteria. Moreover, they need many experiments
and user choices like sampling period, time-delay, number of
on the real system and cannot consider the multimodel
parameters in numerator and denominator of plant and noise
uncertainty.
model, optimal excitation etc. As a result, identification of
With new progress in numerical methods for solving
frequency response functions or spectral models that need
convex optimization problems, new approaches for controller
less a priori knowledge of plant has attracted the attention
design with convex objectives and constraints have been
of many researchers in recent years [1]. In this type of
developed. In [7] a convex optimization method for PID con-
models the information is not condensed into a small set
troller tuning by open-loop shaping in the frequency-domain
of parameters thus avoiding errors of unmodeled dynamics
is proposed. The infinity-norm of the difference between
that appear in parametric models.
the desired open-loop transfer function and the achieved
Although identification of spectral models has been largely
one weighted by a so-called target sensitivity function is
considered in literature, methods for controller design based
minimized. It is shown based on the small gain theorem
on this type of models are rather limited. The first systematic
that if the infinity norm is less than 1 the nominal closed-
controller design methods were based on graphical tools
loop system is stable. This is a sufficient condition and
in the Bode diagram and are discussed in the classical
depends on the choice of the target sensitivity function. The
textbooks for the design and analysis of control systems.
condition for the stability of multiple models becomes more
The well-known Ziegler-Nichols tuning method based on
conservative as for each model a reasonable target sensitivity
only one point on the frequency response of the plant model
function should be available. In [8] a robust fixed-order
(critical frequency) is still used to tune PID controllers
controller design using linear programming is proposed. The
in many practical situations. This method gives typically
main feature of this method is that the stability and some
good responses to load disturbances for systems that can
robustness margins are guaranteed by linear constraints in
be approximated by a first order model with relatively small
the Nyquist diagram and the method is applicable to multiple
delay. However, for more complicated systems the results
models as well. However, the performance specifications
The authors are with the Laboratoire dAutomatique of Ecole Polytech- are limited to the choice of a lower bound for crossover
nique Federale de Lausanne (EPFL), 1015 Lausanne, Switzerland. frequency and minimization of the integral of the tracking
This research work is financially supported by the Swiss National Science
Foundation under Grant No. 200020-107872. error.
Corresponding author: [email protected] In this paper, the method proposed in [8] is improved.
First, it is shown that by minimizing the two norm of the where
difference between the actual open-loop transfer function and T = [1 , 2 , . . . , n ] (4)
the desired one under the stability constraints, the closed- T
loop performance can be ameliorated. Secondly the H infinity (s) = [1 (s), 2 (s), . . . , n (s)] (5)
mixed sensitivity function shaping for robust performance n is the number of controller parameters and i (s), i = 1, n
is represented by linear constraints in the Nyquist diagram are transfer functions with no RHP pole chosen from a set
and then solved by quadratic programming. In this approach of orthogonal basis functions. The main property of this
frequency-domain uncertainty originating from noise and parameterization is that every point on the Nyquist diagram
numerical error in spectral models and multiple model un- of K(j)Gi (j) can be written as a linear function of the
certainty can be directly taken into account. The proposed controller parameters :
method can be used for PID controllers as well as for
higher order linearly parametrized controllers in discrete or K(jk )Gi (jk ) = T (jk )Gi (jk )
(6)
continuous time. The effectiveness of the proposed approach = T Ri (k ) + jT Ii (k )
is illustrated by application to a benchmark problem for where Ri (k ) and Ii (k ) are respectively the real and
robust digital controller design [9]. imaginary parts of (jk )Gi (jk ).
This paper is organized as follows: In Section II the
class of models, controllers and the control objectives are C. Design Specifications
defined. Section III introduces the linear stability margin Some classical methods for controller tuning are based
and presents the open-loop shaping method. New linear on either open-loop or closed-loop shaping. Open-loop
constraints assuring the robust performance and the closed- shaping methods tune the controller so that the open-loop
loop shaping method are presented in Section IV. A solution transfer function L(s) = K(s)G(s) has an appropriate
to a benchmark problem satisfying all of the specifications shape. Normally high magnitudes at low frequencies are
is given in Section V. Finally, Section VI gives some required for load disturbance attenuation. A -20dB/decade
concluding remarks. slope is desired near the crossover frequency for appropriate
robustness. Finally, small magnitudes at high frequencies
II. P ROBLEM F ORMULATION
is advantageous for a good attenuation of measurement
A. Class of models noise or unmodeled dynamics perturbations. Practically, a
The class of linear time-invariant SISO systems with no pure integrator is a recommended shape satisfying these
pole in the right half plane (RHP) is considered. The model specifications.
belongs to a set P that is the convex combination of m On the other hand, closed-loop shaping methods tune the
spectral models with a finite number of frequency points N : controller giving appropriate shapes to closed-loop transfer
functions:
m m output sensitivity function S(s) = [1 + L(s)]1
! #
" "
P= i Gi (jk ) : i = 1; k = 1, N (1) input sensitivity function U(s) = K(s)S(s)
i=1 i=1 complementary sensitivity function T (s) = L(s)S(s)
where i are real positive numbers. This set represents Commonly, lower and upper bounds are fixed in these closed-
multimodel uncertainty as well as unstructured frequency- loop transfer functions. For example, S(s) should be small
domain uncertainty. Consider, for example, the following at low frequencies and its supremum should not exceed a
multiplicative uncertainty model : certain value to have good performance and robustness.
$ %
G(jk ) = 1 + (jk )W2 (jk ) G0 (jk ) ; k = 1, N (2) III. ROBUST O PEN L OOP S HAPING USING Q UADRATIC
where G0 is the nominal model, W2 a weighting filter P ROGRAMMING
defined in N frequency points (not necessarily represented A linear programming approach has been recently pro-
by a transfer function) and an unknown stable system posed in [8]. In this approach a new stability margin is
satisfying !! < 1. This uncertainty can be approximated defined that guarantees a lower bound on classical margins
by a polytopic model which takes the form of (1). The (gain, phase and modulus margins). This linear margin can
multiplicative uncertainty model at each frequency defines be ensured for all models in P by linear constraints in
a circle of radius |W2 G0 | in the Nyquist diagram. At each the parameters of a linearly parameterized controller. This
frequency a convex polygon circumscribing this circle can margin is defined as follows. A straight line d1 in the
be defined. Each vertex of these polygons define a model in complex plane crossing the negative real axis between 0 and
the set P (see Fig. 2) . -1 with an angle is considered (see Fig. 1). The new linear
stability margin & is the distance between the critical point
B. Class of controllers -1 and the intersection point of d1 with the negative real
Linearly parameterized controllers are used in the pro- axis. If the Nyquist curve of the open-loop transfer function,
posed method : L(j), lies on the right hand side of d1 , a lower bound on
K(s) = T (s) (3) the classical robustness margins is ensured. Furthermore, in
 Im
problem:
d1
Minimize !T (jk )G(jk ) Ld (jk )!22
II
Subject to: (7)
T (cot I(k ) R(k )) + & 1 for all k
-1 & Re
where k = 1, . . . , N .
If L(j) contains two integrators, the constraints should
I be modified such that for low frequencies L(j) can be
III c located in Region III. The optimization problem can be
x formulated as:
Minimize !T (jk )G(jk ) Ld (jk )!22
IV
Subject to:
(8)
d2 T (cot I(k ) R(k )) + & 1 for k > x
T (cos I(k ) + sin R(k )) 1 for k x
for k = 1, . . . , N . This optimization problem can be easily
extended to the polytopic models represented by P in (1). For
example, in the case that L(j) contains only one integrator
Fig. 1. Linear constraints for robustness and performance. New linear we have :
margin assuring classical robustness margins m
& (jk )Gi (jk ) Ldi (jk )&2
" & T &
Minimize 2
i=1 (9)
Subject to:
order to consider a lower bound for the crossover frequency, T cot Ii (k ) Ri (k ) + & 1
' (
for all k
we plot a second straight line d2 with an angle with the
real axis tangent to the unit circle centered at the origin. for i = 1, m, where Ldi is the desired open-loop transfer
The intersection of d2 and L(j) defines x , the lower function for Gi .
bound for crossover frequency. The lines (d1 , d2 ) divide the IV. ROBUST C LOSED -L OOP SHAPING USING LINEAR
complex plane in 4 regions I, II, III and IV. By restricting the CONSTRAINTS
open-loop transfer function to lie in one of the regions for a
In the H control framework, nominal performance is
frequency interval, closed-loop performance and robustness
represented by !W1 S! < 1 where W1 (s) is a known
can be achieved by a set of linear constraints.
weighting filter. On the other hand, robust stability for
Two optimization problems are proposed and solved by multiplicative uncertainty is assured iff !W2 T ! < 1. It
linear programming in [8]. is shown in [10] that, for uncertain systems, the necessary
and sufficient condition for robust performance is :
1) Minimizing the integrated error (IE) in order to op-
timize load disturbance rejection. This minimization !|W1 S| + |W2 T |! < 1 (10)
leads to a linear objective function and therefore the
control problem can be solved by linear programming. The graphical interpretation of this constraint in Nyquist
However, it should be used with caution as it may lead diagram is shown in Figure 2. For all k , the plants
to oscillatory response because a norm of error is not uncertainty circle of radius |W2 (jk )L0 (jk )|, with L0 =
minimized. K(jk )G0 (jk ), should not intersect the circle with radius
2) Maximizing the linear margin & for robustness subject |W1 (jk )| centered at -1. This constraint is a non-convex
to a lower bound for the crossover frequency. constraint on the controller parameters. However, the circle
of radius |W1 (jk )| centered at -1 can be approximated by
In this paper we introduce the open-loop shaping in the a tangential dk line with an angle k between the negative
framework of the controller design method proposed in real axis and dk . This gives a convex constraint for each
[8]. In this section, a desired open-loop transfer function frequency k . If the circle of radius |W2 (jk )L0 (jk )|
Ld (s) corresponding to an appropriate closed-loop behavior centered at L0 (jk ) is located on the right hand side of dk
is chosen. Ld (s) is typically chosen as an integrator, however, for all frequencies k , then robust performance is assured.
more complicated transfer functions also can be considered. This can be presented by a linear constraint for k [0, 90 ]:
The 2-norm of the difference between the open-loop transfer
T cos k I(k ) sin k R(k ) + |W1 (jk )| sin k
' (
function L(j) and Ld (j) is minimized subject to the linear
constraints assuring the stability margins. Minimizing the 2- For low frequencies (k < c ) where typically high values
norm leads to a quadratic cost function which can be solved for |W1 (jk )| are given, k = 0 can be chosen (dk is
by standard quadratic programming algorithms. When L(j) thus horizontal). For high frequencies (k > c ) |W1 (jk )|
contains one integrator, L(j) should be in Region I or IV is usually small, and k = 90 is a good choice (dk
for all k . This can be achieved by a simple optimization being vertical). The use of a progressive variation for k
 Im V. S OLUTION TO THE FLEXIBLE TRANSMISSION
BENCHMARK
|W1 |
The design procedure is applied to the flexible transmis-
sion benchmark proposed in [9]. A number of researchers
k have addressed this benchmark problem. Some of the solu-
tions were presented at the 1995 European Control Confer-
-1 Re ence in Rome and published in a special issue (Vol. 1, No.
2, 1995) of the European Journal of Control. Many other
contributions to this benchmark problem can be found in
KG1 literature. To the best of our knowledge, only two controllers
KG2
meet 100% of specifications up to now. The first one is a
QFT controller with 20 parameters in the controller [11]
which requires a good expertise for designing different
dk KGm k lead-lag filters involved in the QFT method. The second
one uses convex optimization methods for pole placement
and sensitivity function shaping [12] and ends up with a
|W2 L0 | controller with 16 parameters. The parametric or multimodel
KGi
uncertainty cannot be considered in this approach. Several
H controllers are proposed for this benchmark [13], [14],
Fig. 2. Multiplicative uncertainty model approximation with a set of dis- [15] that result in very high-order controllers and do not
crete models. Graphical interpretation of the robust performance constraint satisfy all of the benchmark specifications. The multimodel
for a given frequency k . Linear robustness performance constraints for
low frequencies and high frequencies. uncertainty is considered as unstructured uncertainty and the
weighting filters are designed iteratively by a trial and error
approach.
The benchmark problem consists of designing a robust
in the neighborhood of c could be useful. This adds less
digital controller for a flexible transmission system in three
conservatism but the solution could be quite sensitive to the
different operating points (no load, half load and full load).
variation rate.
Three discrete time transfer functions are provided in the
As mentioned in Section II the multiplicative uncertainty is benchmark :
approximated by a set of models at each frequency k giving
q d Bi (q 1 )
a linear constraint for each Li (jk ) = K(jk )Gi (jk ) Gi (q 1 ) = i = 1, 2, 3
point. If all K(jk )Gi (jk ) points for all k lie at the Ai (q 1 )
right hand side of its corresponding dk line in the Nyquist where q 1 is the backward shift operator and d = 2 is
diagram, then robust performance is assured. the integer number of sampling periods contained in the
The objective function of the optimization is to mini- plant pure time delay with 20Hz as sampling frequency. The
mize the 2-norm of the error between L0 (j) and Ld (j) corresponding identified and validated models are:
subject to linear robust performance constraints. Ld (s) can Unloaded model:
be defined as an integrator or from the given W1 (s) by A1 (q 1 ) = 1 1.14833q 1 + 1.58939q 2
Ld (s) = W1 (s) 1 since W1 (s) is typically chosen as 1.31608q 3 + 0.88642q 4
the inverse of the target output sensitivity function. When B1 (q 1 ) = 0.28261q 1 + 0.50666q 2
L0 (j) contains only one integrator, the optimization can
be defined as follows: Half loaded model:
A2 (q 1 ) = 1 1.99185q 1 + 2.20265q 2
T 1.84083q 3 + 0.89413q 4
Minimize ! (jk )G(jk ) Ld (jk )!22
Subject to: B2 (q 1 ) = 0.1027q 1 + 0.18123q 2
(11)
T I(k ) |W1 (jk )| for k c Fully loaded model:
T R(k ) |W1 (jk )| 1 for k > c
A3 (q 1 ) = 1 2.09679q 1 + 2.31962q 2
1.93353q 3 + 0.87129q 4
This optimization problem can be applied to the case that the
B3 (q 1 ) = 0.06408q 1 + 0.10407q 2
system has both multimodel uncertainty and multiplicative
uncertainty. In this case, only the number of constraints is A controller should be designed to satisfy the following
increased and the cost function will change to sum of the specifications: rise time (90% of the final value in less than
two norms of difference between desired and actual open- 1s), overshoot (less than 10%), rejection of 90% of the
loop transfer functions. In the next section, the proposed output disturbance filtered by 1/Ai in less than 1.2s, perfect
approach is applied to a benchmark problem with multimodel rejection of a constant disturbance, disturbance attenuation at
uncertainty. low frequency, maximum output sensitivity function S less
 TABLE I
than 6dB (modulus margin greater than 0.5), a delay margin
S PECIFICATIONS OF THE CONTROLLER
of at least 40 ms and a maximum value of less than 10dB for
the input sensitivity function U at high frequencies (between Specification No load Half load Full load
8 to 10 Hz). Rise Time [s] 0.923 0.895 0.802
Overshoot[%] 6.1 7.8 6.7
A discrete-time two-degree of freedom polynomial form Disturbance reject. [s] 1.16 1.17 1.00
RST controller is to be designed. The canonical form of the Maximum S [dB] 5.86 5.48 5.96
RST controller is given by: Delay Margin [ms] 76 159 338
Maximum U [dB] 9.59 9.04 8.95

S(q 1 )u(t) = T (q 1 )r(t) R(q 1 )y(t) (12)

where u(t) is the plant input, y(t) the plant output and r(t) formulated:
the desired reference. 3
& (jk )Gi (jk ) ci /(jk )&2
" & T &
We consider the following linearly parameterized con- Minimize 2
troller : i=1
Subject to:
R(q 1 ) = (0 + 1 q 1 + . . . + 6 q 6 )(1 + q 1 ) T Ii |A1
i (jk )| for k ci
i = 1, 2, 3
S(q )
1
= (1 q 1
)
T Ri |A1
i (jk )| 1 for k > ci
T (q 1 ) = t0 i = 1, 2, 3
T (cot Ii Ri ) + & 1 for all k
The fixed term (1 q 1 ) in S(q 1 ) satisfies the integral i = 1, 2, 3
action of the controller and disturbance attenuation at low
frequencies and the fixed term (1 + q 1 ) in R(q 1 ) reduces where
Ri = Re[(jk )Gi (jk )]
the input sensitivity function U at high frequencies. In
Ii = Im[(jk )Gi (jk )]
addition, = 80 and & = 0.5/ sin are chosen in order
to assure the specified bound on the maximum of the output and
sensitivity function (modulus margin of 0.5). The time- 1 + ejk h
domain performances (rise time, overshoot and disturbance (jk ) = [1, ejk h , . . . , e6jk h ]
1 ejk h
rejection time) are tuned using the desired open-loop transfer
functions Ldi (s) = ci /s. Note that since the design and h = 0.05 is the sampling period. Choosing ci = 1.2
is carried out in frequency domain the desired open-loop rad/s for all models gives a controller satisfying almost all
transfer function can be defined either in continuous- or in the specifications but the disturbance rejection time for the
discrete-time. The delay margin cannot be transformed into unloaded case (1.24 s instead of less than 1.2 s). Increasing
a linear constraint. However, this specification is naturally the crossover frequency of the desired open-loop transfer
met when other specifications are satisfied. function for the unloaded system to 2.6 rad/s, a controller
The controller is tuned in two steps. In the first step the satisfying 100% of the specifications is obtained with these
feedback controller is tuned based on the proposed method. parameters :
In the second step the unique parameter in T (q 1 ) is taken R(q 1 ) = (1 + q 1 )(0.4485 1.7163q 1
equal to sum of the parameters of R(q 1 ) giving a unit
+2.9159q 2 3.2385q 3 + 2.6753q 4
gain to the closed-loop system. The frequency response of
the three models are computed at N=8000 equally spaced 1.4738q 5 + 0.4126q 6)
frequency points between 0 and 10Hz (Nyquist frequency). S(q 1 ) = (1 q 1 )
The desired crossover frequency ci = 1.2 rad/s was T (q 1 ) = 0.0474
chosen for all models giving a controller satisfying all the
specifications except the disturbance rejection time. Looking Figures 3 and 4 show that the specifications on the input
at the frequency response of the closed-loop transfer function sensitivity function Ui and output sensitivity function Si
between disturbance and output, Si /Ai , a large peak greater are satisfied for the three models. Figure 5 shows the step
than 35dB is observed for three closed-loop systems. Limit- the disturbance rejection responses. The details of achieved
ing the maximum of Si /Ai can help reducing the disturbance specifications are shown in Table I.
rejection time. This constraint can be represented by :
VI. C ONCLUSION
& &
& Si & Robust fixed-order controller design for spectral models
& & <
& Ai &

is formulated as a quadratic optimization problem. The
proposed method is based on frequency loop shaping in
and be considered in the proposed approach by taking the Nyquist diagram. The classical robustness margins and
W1i = 1/(Ai ). The design variable is fixed to 1027/20 robust performance criterion are represented as linear con-
(equal to 27dB). Then the following optimization problem is straints in the Nyquist diagram. The control objective is
 to minimize the quadratic error between a desired open-
20
loop transfer function and the achieved one subject to the
10
constraints mentioned before. The few design variables are
directly related to robustness (linear margin &) and perfor-
0
mance (crossover frequency of the desired transfer function)
!10
of the closed-loop system. Multimodel uncertainty can be
taken into account straightforwardly. The method is very
dB

!20
simple to apply and requires only the frequency response of
!30
the plant. The method is very appropriate for PID controller
design, as well as being applicable to higher-order linearly
!40
parameterized controllers in discrete or continuous time.
!50
An application of the proposed method to a challenging
benchmark problem illustrates the effectiveness of the pro-
!60
0 0.05 0.1 0.15 0.2 0.25
f/fe
0.3 0.35 0.4 0.45 0.5 posed method. The multimodel uncertainty of this problem is
directly considered in the approach and using a few design
Fig. 3. U of Unloaded (dashed, blue), Half loaded (solid, green) and Fully parameters a fixed-order robust controller is obtained that
loaded (dashed-dotted, red) systems. satisfies all the specifications with only 10 parameters in
the controller. This is the smallest order controller designed
that meets 100% of specifications with a very simple and
straightforward approach.
10

R EFERENCES
5
[1] R. Pintelon and J. Schoukens. System Identification: A Frequency
Domain Approach. IEEE Press, New York, USA, 2001.
0
[2] K. J. Astrom and B. Wittenmark. Adaptive Control. Addison-Wesley,
1989.
[3] J. Crowe and M.A. Johnson. Automated PI control tuning to meet clas-
sical performance specifications using a phase locked loop identifier.
dB

!5

In IEEE American Control Conference, pages 21862191, Arlington,

USA, 2001.
!10
[4] D. Garcia, A. Karimi, and R. Longchamp. Data-driven controller tun-
ing using frequency domain specifications. Industrial & Engineering
!15
Chemistry Research, 45(12):40324042, 2006.
[5] A. Karimi, D. Garcia, and R. Longchamp. PID controller tuning using
Bodes integrals. IEEE Transactions on Control Systems Technology,
!20 11(6):812821, 2003.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
f/fe [6] L. C. Kammer, R. R. Bitmead, and P. L. Bartlett. Direct iterative
tuning via spectral analysis. Automatica, 36(9):13011307, 2000.
Fig. 4. S of Unloaded (dashed, blue), Half loaded (solid, green) and Fully [7] E. Grassi and K. Tsakalis. PID controller tuning by frequency loop
loaded (dashed-dotted, red) systems. shaping. In 35th IEEE Conference on Decision and Control, pages
47764781, Kobe, Japan, 1996.
[8] A. Karimi, M. Kunze, and R. Longchamp. Robust controller design
by linear programming with application to a double-axis positioning
system. Control Engineering Practice, 15(2):197208, February 2007.
[9] I. D. Landau, D. Rey, A. Karimi, A. Voda, and A. Franco. A
flexible transmission system as a benchmark for robust digital control.
1.2
European Journal of Control, 1(2):7796, 1995.
[10] C. J. Doyle, B. A. Francis, and A. R. Tannenbaum. Feedback Control
Theory. Mc Millan, New York, 1992.
1
[11] M. Nordin and P. Gutman. Digital QFT design for the benchmark
problem. European Journal of Control, 1(2):97103, 1995.
0.8
[12] J. Langer and A. Constantinescu. Pole placement design using
convex optimisation criteria for the flexible transmission benchmark.
Output

0.6
European Journal of Control, 5:193207, 1999.
[13] N. W. Jones and D. J. N. Limebeer. A digital H controller for a
0.4
flexible transmission system. European Journal of Control, 1(2):134
140, 1995.
0.2
[14] D. J. Walker. Control of a flexible transmission - a discrete time H
approach. European Journal of Control, 1(2), 1995.
0
[15] G. Ferreres and V. Fromion. H control for a flexible transmission
system. European Journal of Control, 5:185192, 1999.
0 1 2 3 4 5 6 7
Time

Fig. 5. Step and perturbation rejection of Unloaded (dashed, blue), Half

loaded (solid, green) and Fully loaded (dashed-dotted, red) systems. The
dashed lines show the limits where the responses should be inside.