INTERNATIONAL JOURNAL OF ROBUST AND NONLINEAR CONTROL, VOL.

8, 191—210 (1998)
Int. J. Robust Nonlinear Control, 8, 191—210 (1998)

ROBUST DIGITAL CONTROL USING POLE PLACEMENT

WITH SENSITIVITY FUNCTION SHAPING METHOD

I.D. LANDAU* AND A. KARIMI

Laboratoire d’Automatique de Grenoble, ENSIEG, BP. 46, 38402 Saint Martin d’He% res, France

SUMMARY
It is shown in this paper that based on identified discrete-time models a robust digital linear controller can
be designed using the pole placement method combined with sensitivity function shaping in the frequency
domain. Two different design techniques are presented. The first one is based on the shaping of the
sensitivity functions using the fixed parts in the controller and the auxiliary poles of the closed loop while
keeping the dominant poles in the desired places. The main idea of the second one is to determine
a weighting filter for the output sensitivity function in an H optimization approach which assures partial
=
pole placement and desired performances. In this technique the weighting filter is interpreted as the inverse
of the desired output sensitivity function and is computed using a constrained optimization program. The
application of this technique on a flexible transmission system is presented. ( 1998 John Wiley & Sons, Ltd.
Key words: digital control; robust control; pole placement; H optimization
=

1. INTRODUCTION
The controller design methodology considered in this paper is based on pole placement combined
with the shaping of the sensitivity function. The computation of the controller in the pole
placement technique requires the specification of the desired closed-loop poles (the nominal
stability problem) and of some fixed parts of the controller for the rejection of disturbances at
various frequencies (the nominal performance problem). However, this is not enough to guarantee
the robustness of the design with respect to the plant model uncertainties (the robust stability and
robust performance problem). A robust controller design requires also the shaping of the sensitivity
functions. The sensitivity functions, particularly the output sensitivity function, are key indicators
for the nominal and robust performance as well as for the robust stability of the closed-loop
system. The inverse of the maximum value of the output sensitivity function, i.e., the inverse of its
H norm, gives the minimum distance between the Nyquist plot of the open-loop system and the
=
critical point [!1, j0]. This quantity, called the modulus margin, is a much more significant
robustness indicator than the phase and gain margins. On the other hand, conditions for assuring
a certain delay margin which is also a very important robustness indicator, particularly in the high
frequency region, can also be expressed in terms of the shape of the output sensitivity function. It
seems therefore reasonable to combine the pole placement with the shaping of the output

* Correspondence to: I.D. Landau, Laboratoire d’Automatique de Grenoble, ENSIEG, BP. 46, 38402 Saint Martin
d’Héres, France

CCC 1049-8923/98/020191—20$17.50
( 1998 John Wiley & Sons, Ltd.
192 I.D. LANDAU AND A. KARIMI

sensitivity function (or eventually with other sensitivity functions) in order to design robust digital
controllers for SISO plants.
This paper addresses the robust controller design problem for two cases:

(a) Only one plant model is available. In this case we consider certain robustness margins
which translate into templates for the sensitivity functions.
(b) Several plant models are given or an estimation of the model uncertainty in the frequency
domain is available. In this case again the uncertainties will translate into templates for the
sensitivity functions.

For this purpose two different techniques are given. In the first one an iterative method is
presented for correctly specifying the desired closed-loop poles and the fixed parts of the
controller in order to assure the nominal performance, the desired robustness margins, as well as
the robust stability of the closed loop for given classes of plant model uncertainties. This method
has been applied on a large number of real systems.1~4 In the second one a systematic procedure
to compute a controller based on partial pole placement with sensitivity function shaping via
H optimization is presented. In this technique the weighting filter in the H approach is
= =
interpreted as the inverse of the desired output sensitivity function. Therefore, all of the conditions
and the constraints on the output sensitivity function can be considered for the inverse of the
weighting filter. One of the properties of the output sensitivity function is that it satisfies the
Zames—Francis equality.5,6 This equality then is used as a performance index in a constrained
optimization program in order to find the unknown parameters of the desired sensitivity function.
Once the desired sensitivity function is obtained, an H approach is used for computing the
=
controller. The main differences between this technique and the partial pole placement presented
in References 7 and 8 are as follows:

d The weighting filter is interpreted as the inverse of the desired output sensitivity function.
d This filter is the result of an optimization problem so it is not necessary to find the
parameters of the filter by trial and error.
d The controller design can be carried out using only one weighting filter.
d The conditions on the weighting filter in order to obtain a stable controller are intro-
duced.

The paper is organized as follows: Section 2 reviews the R-S-T controller and the notation.
Section 3 gives the various sensitivity functions associated with the R-S-T controller. In Section 4
a template for the output sensitivity function is defined. The shaping of the sensitivity function
with pole placement technique is described in Section 5. The shaping of the sensitivity function
with partial pole placement using H optimization is presented in Section 6. The application of
=
this technique on a flexible transmission system is given in Section 7. Section 8 gives the
concluding remarks.

2. R-S-T CONTROLLER
Figure 1 shows a schematic diagram of an R-S-T controller. The discrete time plant is described
by the following transfer operator:

q~dB(q~1)
H(q~1)" (1)
A(q~1)

Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.
 ROBUST DIGITAL CONTROL 193

Figure 1. Closed-loop system with an R-S-¹ controller in the presence of output disturbances and measurement noise

where q~1 is the backward shift operator, i.e., y(t!1)"q~1y(t), d is the integer number of
sampling periods (¹ ) contained in the plant pure time delay and
s
A(q~1)"1#a q~1#2#a q~nA (2)
1 nA
B(q~1)"b q~1#b q~1#2#b q~nB (3)
1 2 nB
It is assumed that the polynomials A(q~1) and B(q~1) do not have common factors. The
canonical form of the R-S-T controller is given by
S(q~1) u(t)"¹(q~1) y*(t#d#1)!R(q~1) y(t) (4)
where u(t) is the plant input, y(t) is the plant output and y*(t#d#1) is the desired tracking
(reference) trajectory. This trajectory may be generated by a tracking reference model.
B (q~1)
y*(t#d#1)" m r(t) (5)
A (q~1)
m
where r(t) is the reference signal. The closed-loop transfer function between the reference
trajectory and the plant output is given by
z~dB(z~1) ¹(z~1)
H (z~1)" (6)
CL P(z~1)
where
P(z~1)"A(z~1) S(z~1)#z~dB(z~1) R(z~1)"P (z~1) · P (z~1) (7)
D F
defines the closed-loop poles. P (z~1) and P (z~1) correspond to the dominant and auxiliary
D F
closed-loop poles, respectively.

3. THE SENSITIVITY FUNCTIONS

The sensitivity functions play a crucial role in the robustness analysis of the closed-loop system
with respect to modelling errors. These functions will be shaped in order to assure nominal
performance for the rejection of the disturbances and the stability of the closed-loop system in the
presence of model mismatch. Two types of disturbances are considered: output disturbance and
measurement noise.

( 1998 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 8, 191—210 (1998)
194 I.D. LANDAU AND A. KARIMI

The various sensitivity functions can be established from Figure 1. The transfer function
between the disturbance p(t) and the plant output y(t) (output sensitivity function) is given by
A(z~1) S(z~1) A(z~1) S(z~1)
S (z~1)" " (8)
yp A(z~1) S(z~1)#z~dB(z~1) R(z~1) P(z~1)
This function is often called simply the sensitivity function in the literature and is denoted by S.
The transfer function between the disturbance p(t) and the plant input u(t) (input sensitivity
function) is given by
!A(z~1) R(z~1) !A(z~1) R(z~1)
S (z~1)" " (9)
up A(z~1) S(z~1)#z~dB(z~1) R(z~1) P(z~1)
This function is often denoted by º in the literature. The transfer function between the
measurement noise b(t) and the plant output y(t) (noise sensitivity function) is given by
!z~dB(z~1) R(z~1) !z~dB(z~1) R(z~1)
S (z~1)" " (10)
yb A(z~1) S(z~1)#z~dB(z~1) R(z~1) P(z~1)
This function with a positive sign is called the complementary sensitivity function denoted by ¹ in
the literature. From equations (8) and (10) it is obvious that
S (z~1)!S (z~1)"1 (11)
yp yb

4. DEFINITION OF A ‘TEMPLATE’ FOR THE OUTPUT SENSITIVITY FUNCTION

Using the small gain theorem and various representations of the plant uncertainties, the modulus
margin and the delay margin can be converted into robust stability conditions. On the other hand,
the robust stability conditions allow the definition of upper templates for the modulus of the
various sensitivity functions.1,9,10
For a delay margin of one sampling period the robust stability condition is expressed as
follows:11
DS (z~1)D(D(1!z~1)D~1, z"e~ju, 0)u)n (12)
yb
But, from equation (11) we have
1!DS (z~1)D(DS (z~1)D(1#DS (z~1)D (13)
yb yp yb
If S (z~1) satisfies the condition (12) then S (z~1) will satisfy the following condition:
yb yp
1!D1!z~1D~1(DS (z~1)D(1#D1!z~1 D~1; z"e~ju ; 0)u)n (14)
yp
Therefore in order to assure the delay margin *q"¹ , it is required that the modulus of S (z~1)
s yp
lies between the bounds defined by a lower template D¼~1D "1!D1!z~1D~1 and an upper
inf
template defined by D¼~1 D "1#D1!z~1D~1.
sup
The nominal performance requirements and the robust stability conditions lead to the definition
of a desired template for the sensitivity function. We will subsequently consider the definition of
such a template for the output sensitivity function. The desired template takes, in general, the
form shown in Figure 2. Regarding the robust stability, the chosen modulus margin will define the
maximum value of the modulus of the output sensitivity function (upper template) and the chosen
delay margin will define an upper and a lower template starting, for example, around 0·15 f (for
s
*q"¹ ).
s
Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.
 ROBUST DIGITAL CONTROL 195

Figure 2. Desired template for the output sensitivity function (the case of low frequencies disturbance rejection)

A template for the input sensitivity function also may be defined in the same way. Additive
uncertainty can be used as an upper template to assure the robust stability of the closed-loop
system. Another upper bound should be considered for the frequencies where the plant gain is
very low, usually in the high frequencies and at the frequencies of the low damped zeros of the
plant model, in order to prevent the unwanted excitation of control input which has no effect in
the output.

5. METHOD A
5.1. Pole placement
In this section the pole placement technique is reviewed; for more details see Reference 12. The
dominant closed-loop poles are chosen to satisfy the desired regulation performance. The
auxiliary poles can be introduced either for filtering effects in certain frequency regions or for
reducing the stress on the actuator as well as for improving the robustness of the closed-loop
system (as will be shown subsequently).
For different reasons, i.e., disturbance rejection, signal blocking, etc. the polynomials R(z~1)
and S(z~1) generally contain some fixed parts which are specified before solving equation (7) (for
example the polynomial S(z~1) should contain a factor (1!z~1) for a null steady state error). In
order to take into account these prespecified parts, the polynomials R(z~1) and S(z~1) are
factored as:

R(z~1)"R@(z~1) · H (z~1) (15)

R
S(z~1)"S@(z~1) · H (z~1) (16)
S
where H (z~1) and H (z~1) are the prespecified polynomials.The closed-loop poles in equation
R S
(7) become

P(z~1)"A(z~1) · H (z~1) · S@(z~1)#z~dB(z~1) · H (z~1) · R@(z~1) (17)

S R
( 1998 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 8, 191—210 (1998)
196 I.D. LANDAU AND A. KARIMI

The design procedure can be summarized as follows:

(a) Choose the desired closed-loop poles P(z~1), the fixed parts of the controller H (z~1),
R
H (z~1) and the desired tracking dynamics B (z~1)/A (z~1).
S m m
(b) Compute S@(z~1) and R@(z~1) by solving equation (17) and consequently compute S(z~1)
and R(z~1) using equations (15) and (16).
(c) Compute the prefilter ¹(z~1)"P(z~1)/B(1). This choice leads to the following transfer
function between the reference and the plant output:
z~dB (z~1) B(z~1)
H (z~1)" m · (18)
yr A (z~1) B(1)
m
Another method consists of taking ¹(z~1)"R(1).
z~dB (z~1) B(z~1) P(1)
H (z~1)" m · · (19)
yr A (z~1) P(z~1) B(1)
m
Of course intermediate solutions between (18)- and (19)- can also be obtained, that is, to
choose ¹(z~1) such that it compensates only a part of the poles of P(z~1).

5.2. Shaping the sensitivity functions

The dominant poles P (z~1), the auxiliary poles P (z~1) and the two filters H (z~1) and
D F R
H (z~1) are the tools for shaping the sensitivity functions in order to meet the specifications
S
defined by the desired templates. Placing a pair of complex zeros in H (z~1) or in H (z~1)
S R
decreases the magnitude of the modulus of the sensitivity functions around the frequencies of
these zeros. Introducing a pair of pure imaginary zeros at a certain frequency in H (z~1) will
R
make DS D"1 and DS D"0 at this frequency. Complex zeros will bring DS D closer to the 0 dB
yp up yp
axis. Putting some asymptotically stable real high frequency poles in P(z~1) will cause an
attenuation of DS D in the domain of attenuation of 1/P (z~1). Placing a pair of complex poles at
yp F
a certain frequency in P (z~1) will lead to the increase of DS D around this frequency. Introducing
F yp
a term of (1#z~1) in H (z~1) will open the loop in the high frequencies.
R
The underlying philosophy of the design is to choose the closed-loop poles and the fixed parts
of the controller in order to bring the sensitivity functions within the defined templates. This
controller will meet the conditions on modulus and delay margin as well as the attenuation band
and will assure the robust stability of the closed-loop system against the parameter uncertainty.
The shaping method which is an iterative procedure (generally converging in few steps) is as
follows:

Step I
d Choose the dominant closed-loop poles of P(z~1) and the fixed part of R(z~1) and S(z~1) in
order to meet the nominal performance specifications.
d Compute the controller.
d Check the shape of the output sensitivity function. If the upper bound of the template is not
satisfied, generally two situations are identified:
(a) the maximum of the output sensitivity function is located in the frequency range next to
the attenuation band. In this case, go to step II;
(b) the maximum of the output sensitivity function is located in the high frequency range
(near 0·5 f ). In this case jump to step III.
s
Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.
 ROBUST DIGITAL CONTROL 197

Step II
d Add a pair of complex zeros to H (z~1):
S
H (z~1)"1#a z~1#a z~2 (20)
S 1 2
The frequency of these zeros is chosen close to the frequency where the maximum of the DS D
yp
occurs (between the maximum frequency of the attenuation band and the frequency where
DS D occurs). The damping factor is chosen such that the introduced attenuation brings
yp max
DS D below the admissible value. Typical values are between f"0·3 to 0·8.
yp max
d Recompute the controller. If DS D is too large in the high frequency range, go to step III.
yp

Remark
The effects of these additional zeros are in general:
(a) a reduction in the modulus of the sensitivity function in the region next to the attenuation
band;
(b) an increase in the maximum of the modulus of the sensitivity function in the high frequency
range (in general its maximum is shifted in the high frequency range);
(c) an increase in the attenuation band.

Step III
d Add auxiliary high frequency poles of the form:

P (z~1)"(1!p z~1)nF ; 0·05)p )0·5 (21)

F 1 1
where
n )n !n ; n "(deg P) ; n "deg P (22)
F P D P max D D
with increasing values of p starting from 0·05 till the specifications on the DS D in the high
1 yp
frequency range are met. It should be mentioned that high frequency auxiliary poles shift, in
general, the maximum of the DS D towards the lower frequency range. If the DS D is
yp yp max
satisfactory, stop here, if not go to step IV.

Step IV This step is similar to step II

d If the new maximum is close to the one resulting after step I, change accordingly the
frequency and the damping of the zeros introduced in H .
S
d If the new maximum is at a significantly different frequency, add a new pair of complex zeros
in H (z~1) at a frequency close to the maximum of DS D. The damping is chosen in order to
S yp
bring DS D below the maximum acceptable value. If the results are unsatisfactory, go to
yp max
step V.

Step V
d The value of the real auxiliary poles (defined in step III) can be increased.
d If necessary, the dominant poles may also be slowed down.
d End of the procedure for the shaping of the output sensitivity function. Go to step VI.

Step VI
d Check the input sensitivity function. Add, if necessary, high frequency complex zeros in
H (z~1) (often at 0·5 f ).
R s
( 1998 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 8, 191—210 (1998)
198 I.D. LANDAU AND A. KARIMI

d Check again the output sensitivity function. Three different situations may occur:
(a) the output sensitivity function is over the template in the high frequencies. Go to step V;
(b) the output sensitivity function is over the template in the low frequencies. Go to step IV;
(c) the output sensitivity function is inside the template. End of the procedure.
The flowchart of this procedure is illustrated in Figure 3.

Figure 3. The flow chart of the iterative procedure of Method A

Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.
 ROBUST DIGITAL CONTROL 199

6. METHOD B
The problem is to find a controller which guarantees the internal stability of the closed loop,
assures desired dominant closed-loop poles and satisfies certain constraints on the output
sensitivity function. To solve this problem, first a desired sensitivity function which satisfies the
constraints on the output sensitivity function is computed. The use of an appropriate fixed part
in the desired sensitivity function will assure a partial pole placement to desired values. Then the
inverse of the desired sensitivity function will be used as a weighting filter to minimize the infinity
norm of the output sensitivity function. Therefore, the real output sensitivity function will
approximate the desired one. The internal stability of the closed loop will be guaranteed by using
the Youla parametrization in the H optimization.
=
The demonstration will be carried out in discrete time but it could be done for continuous time
as well.

6.1. Desired sensitivity function

It was shown in Section 3 that the output sensitivity function may be written in the following
form:
A(z~1) S(z~1) A(z~1) S(z~1)
S (z~1)" " (23)
yp A(z~1) S(z~1)#z~dB(z~1) R(z~1) P (z~1) P (z~1)
D F
Where P and P are the dominant and auxiliary poles of the closed loop. Since A and the desired
D F
closed-loop poles P are known, we can fix them in the desired sensitivity function. Therefore we
D
can define
A(z~1) X(z~1)
S (z~1)" · (24)
d P (z~1) ½(z~1)
D
where S is the desired sensitivity function and the polynomials X and ½ should be computed
d
using an optimization method.
It can be proved7 that the denominator of S will appear as a part of the final closed-loop poles
d
if the inverse of S is utilized as a weighting filter in the H optimization method. This technique,
d =
known as partial pole placement,7,8 rejustifies the interpretation of the weighting filter as the
inverse of the desired sensitivity function. On the other hand, natural choice of the polynomial
A in the numerator of S will prevent the presence of the open-loop poles as closed-loop poles.7
d
In the next step, the polynomials X and ½ will be parametrized. They are parametrized
preferably as products of real zeros and poles, respectively, in order firstly to prevent the
unwanted cancellation of the desired poles P or of the open-loop poles A which may contain low
D
damped complex poles, and secondly to avoid the oscillatory poles in the polynomial ½ which
will appear consequently as the closed-loop poles.
Then the polynomials X and ½ are defined as follows:
X(z~1)"(1!x z~1)(1!x z~1) 2 (1!x z~1) (25)
1 2 nX
½(z~1)"(1!y z~1)(1!y z~1) 2 (1!y z~1) (26)
1 2 nY
h"[x , x , 2 , x , y , y , 2 , y ] (27)
1 2 nX 1 2 nY
where x to x are the real zeros of X, y to y are the real zeros of ½ and h is the vector of the
1 nX 1 nY
parameters. It may be considered that S (z~1, h) is proper and inversely proper, i.e., the order of
d
the numerator is equal to the order of denominator. This leads to a proper weighting filter for

( 1998 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 8, 191—210 (1998)
200 I.D. LANDAU AND A. KARIMI

H optimization. It should be mentioned that the properness of the weighting filter can be
=
relaxed.13
It can be shown that the denominator of the weighting filter will appear as a part of the
denominator of the controller. Then, in this approach, the desired fixed parts in the denominator
of the controller can be put directly in the numerator of the S . Thus, for example, to reject the
d
constant disturbances in the output, an integrator may be used in S (z~1, h). This can be easily
d
done by choosing x "1.
1
The minimum order of X and ½ depends on the order of A, P and the fixed parts of ½. We
D
have
n *n #(n ) !n (28)
X A Y .*/ PD
where n , n , n and (n ) are the orders of A, X, P and the fixed parts of ½, respectively. For
A X PD Y min D
example, if we choose n "n and an integrator in ½ we will obtain n "n "1. In this case we
PD A X Y
have only to compute one parameter x . If with this parameter we do not succeed in minimizing
1
the criterion (which will be explained subsequently) and in satisfying the constraints, we can
increase the order of X and ½ to have more degrees of freedom in the optimization program.

6.2. Performance index

The unstable poles and zeros of the open loop, i.e., the poles and zeros outside the unit circle in
discrete time or with a positive real part in continuous time, play an important role in the
performance of the closed loop and consequently in the form of the sensitivity function. Therefore,
these unstable poles and zeros should be considered in the design of a desired sensitivity function.
A central result, which explains the relation between the unstable poles and zeros of the open loop
and the sensitivity function, is the Zames—Francis equality,5,6 which is based on the Poisson
integral formula from complex function theory. The result that follows is the discrete-time version
of this equality which was obtained by Sung and Hara.14
Suppose that b"rej/ is an unstable zero of the open loop. If the closed-loop system is stable,
then the following relation holds:

P
n
r2!1
n log DB~1 log DS ( ju)D ·
a (b)D" · du (29)
yp 1!2r cos(u!/)#r2
~n
where Ba (b) is the Blaschke product
z!a
Ba (z)"< i (30)
1!a6 z
i i
formed from the unstable poles of the open loop a (here the overbar denotes the complex
i
conjugate). Since this equality is satisfied by S , it should also be satisfied by the desired
yp
sensitivity function S (z~1, h). Therefore, we can choose this equality as a criterion for the
d
optimization program in order to find the polynomials X and ½. The criterion can be expressed
as a function of h (the parameter vector) as follow:

KP K
n
r2!1
J (h)" log DS ( ju, h)D · i · du!n log DB~1 a (b )D (31)
i d 1!2r cos(u!/ )#r2 i
~n i i i
where b "r ej/i are the unstable zeros of the open loop. In fact, only the unstable zeros of the
i i
plant are known so the number of criteria is equal to the number of the unstable zeros of the
plant. It should be mentioned that a time delay in the discrete-time system is in fact an unstable

Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.
 ROBUST DIGITAL CONTROL 201

zero at infinity, that means r goes to infinity in equation (29) and the criterion becomes:

KP K
n
log DS (ju, h)D du!n + log Da D
J (h)" (32)
d d i
~n i
The objective will be then to minimize the multi-objective function:
J(h)"[J (h), J (h), 2 , J (h), J (h)] (33)
1 2 nuz d
where n is the number of the unstable zeros of the plant. It is of course very difficult to minimize
uz
all these criteria simultaneously. An alternative might be to minimize the maximum value of J(h).

Remark
The second term in the criterion J (h) is related to the unstable poles of the open loop. Consider
i
that the plant is strongly stabilizable, then it is possible to compute a stable controller which will
stabilize the closed loop. Thus, it may be supposed that the unstable poles of the open loop are the
only poles of the plant, if there are any. With this hypothesis we search the stable controllers
among all of the stabilizing controllers which satisfy the constraints. Then this hypothesis will be
verified by the final controller if the H norm of the output sensitivity function filtered by the
=
inverse of S is one, i.e., ES~1 · S E "1.
d d yp =
6.3. Constraints
It was shown that the robustness margins, i.e., modulus and delay margin, and the model
uncertainty, as well as the performance of the closed-loop system for output disturbance rejection,
can be defined as a template in the frequency domain for the output sensitivity function. This
template then, may be used as the constraint in the optimization algorithm to compute the
desired sensitivity function. These constraints are expressed as follows:
C (u))DS ( ju, h)D)C (u) (34)
L d U
where C (u) and C (u) are the upper and lower bound for the output sensitivity function,
U L
respectively.

6.4. Optimization program

The optimization program can be summarized as follows:
d Definition of the structure of the desired sensitivity function and the vector of parameters
A(z~1) (1!x z~1)(1!x z~1) 2 (1!x z~1)
S (z~1)" · 1 2 nX (35)
d P (z~1) (1!y z~1)(1!y z~1) 2 (1!y z~1)
D 1 2 nY
h"[x , x ,, 2 , x , y , y ,, 2 , y ] (36)
1 2 nX 1 2 nY
d Definition of the performance index according to the Zames—Francis equality

KP K
n
r2!1
J (h)" log DS ( ju, h)D · i · du!n log DB~1
a (b )D (37)
i n
d 1!2r cos(u!/ )#r2 i
~ i i i

h"Arg min max J (h) (38)

h i
i
( 1998 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 8, 191—210 (1998)
202 I.D. LANDAU AND A. KARIMI

d Definition of the upper and lower bound for the desired sensitivity function

C (u))DS ( ju, h)D)C (u) (39)

L d U
If the results of the optimization program are not satisfactory in terms of the constraints or the
performance index, we can increase the number of parameters, i.e., the order of X and ½, or
eventually we may use a pair of complex poles and zeros in S . However, the experimental results
d
show that even for complex systems it is not necessary to use complex poles and zeros in S . This
d
optimization problem can be solved using the CONSTR instruction in the optimization toolbox
of MATLAB.

6.5. H optimization
=
After computing the desired sensitivity function S (z~1), the output sensitivity function
d
S (z~1) filtered by the inverse of S (z~1) will be minimized using the H optimization method.
yp d =
This problem, known as the model matching problem, is the simplest problem in the H domain.
=
This problem may be solved either iteratively by the standard H optimal control problem
=
using the k-synthesis toolbox in MATLAB or directly by Nevanlinna—Pick’s algorithm.9
The second method is more suitable for this technique since in the model matching problem
the optimal solution can be achieved without any iteration. Therefore, a discrete time version
of the Nevanlinna—Pick algorithm has been developed to solve the application problem of
Section 7.
It should be noticed that in this technique only the partial pole placement, i.e., the closed-loop
poles defined by P , is guaranteed, so it is possible to obtain the other closed-loop poles near to
D
the unit circle. To avoid this problem we can change the stability domain by choosing a circle
with a radius less than one (or axis shifting in continuous-time domain).
It was mentioned in Section 4 that we need usually to open the loop or to reduce drastically the
gain of the polynomial R at certain frequencies where the gain of the system is very low in order to
prevent the unwanted excitation in the control input. It means that it is necessary to shape the
input sensitivity function simultaneously with the output sensitivity function, which is called the
mixed sensitivity problem in the literature.10 An alternative way is to augment the order of the
terms of B (the numerator of the plant) by introducing terms of the form (1#az~1) with a'0·7
which places zeros at 0·5 f and of the form (1#a z~1#a z~2) which introduce complex zeros at
s 1 2
other desired frequencies. Choosing an appropriate damping factor for the complex zeros will
reduce significantly S at the corresponding frequencies. These terms correspond to roots which
up
are very close to the unit circle and will be moved back to the controller (polynomial R ) after
computing it. It should be noticed that the stable zeros of the plant will be cancelled by the
controller in the H approach. Then, in order to avoid the cancellation of these zeros, they should
=
appear as unstable zeros. This is achieved by defining a reduced stability domain, leaving these
roots outside. Using this technique we can shape two sensitivity functions by only one weighting
filter.

7. APPLICATIONS
In this section a flexible transmission system which has been the subject of two benchmarks on
adaptive control15 and on robust digital control16 and has been also used as an example by
Kwakernaak17 is considered. For this system three models of the plant and several specifications
are given.

Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.
 ROBUST DIGITAL CONTROL 203

7.1. System description

The flexible transmission consists of three horizontal pulleys connected by two elastic belts
(Figure 4). The first pulley is driven by a DC motor whose position is controlled by local feedback.
The third pulley may be loaded by small disks of different weights. The system input is the
reference for the axis position of the first pulley. The system output is the axis position of the third
pulley measured by a position sensor, sampled and digitized. A PC is used to control the system.
The main characteristics of this system are the low damped vibration modes (damping factor less
than 0·05) and the large variations of their frequencies with the load. The discrete-time identified
models (¹ "50 ms) of the plant for two extreme situations (no load and full load), as well as for
s
an intermediate situation are given.
Unloaded model:
A (q~1)"1!1·41833q~1#1·58939q~2!1·31608q~3#0·88642q~4
0
B (q~1)"0·28261q~1#0·50666q~2 ; d"2
0
Half load model:
A (q~1)"1!1·99185q~1#2·20265q~2!1·84083q~3#0·89413q~4
1
B (q~1)"0·10276q~1#0·18123q~2 ; d"2
1
Full load model:
A (q~1)"1!2·0967q~1#2·3196q~2!1·93353q~3#0·87129q~4
2
B (q~1)"0·06408q~1#0·10407q~2 ; d"2
2
Several time and frequency specifications are considered for this problem which should be
satisfied for all loads:
(a) A rise time (0—90% of the final value) of less than 1 s for a step change in the reference
input.

Figure 4. Schematic diagram of the flexible transmission

( 1998 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 8, 191—210 (1998)
204 I.D. LANDAU AND A. KARIMI

(b) Overshoot less than 10% for a step change in the reference input.
(c) Rejection of step output disturbances filtered by 1/A within 1·2 s (for 90% rejection of the
measured peak value).
(d) Perfect rejection of constant disturbances (using integral action).
(e) Disturbance attenuation in the low frequency band from 0 to 0·2 Hz.
(f ) A maximum value of less than 6 dB of the output sensitivity function (modulus margin
greater than 0·5).
(g) A delay margin of at least 40 ms (80% of sampling period).
(h) A maximum value of less than 10 dB of the input sensitivity function in the frequency range
8—10 Hz.

For more details see Reference 16.

7.2. Controller design

A complete solution to this flexible transmission problem by Method A has been given in
Reference 3. This problem will be redesigned by Method B in the following steps:

1. The unloaded model which seems to be the most difficult to control is chosen as the
nominal model.
2. Two pairs of complex poles, with the same frequencies as those of the vibration modes of
the nominal plant, but with a damping factor of 0·8 for the first mode (which is more
energetic) and of 0·2 for the second mode, are considered as the dominant closed-loop
poles.
3. The desired sensitivity function is defined as follows:

A (z~1) (1!x z~1)(1!x z~1)

S (z~1)" 0 · 1 2 (40)
d P (z~1) (1!y z~1)(1!y z~1)
D 1 2
4. In order to obtain an integrator in the controller, x should be equal to one. This means
1
that the weighting filter will have a pole on the unit circle, which is not allowed in
H optimization. To avoid this problem we can choose x "0·999999.
= 1
5. It remains to compute three parameters in S by the optimization program. Therefore, the
d
parameter vector h will be: h"[x , y , y ].
2 1 2
6. The nominal model has an unstable zero and a time delay, therefore the performance index
is defined as J(h)"max[J (h), J (h)] where J and J correspond to the unstable zero and
1 d 1 d
the time delay in equations (31) and (32), respectively.
7. A template, like the template in Figure 2, with a minimum modulus margin equal to 0·5
and a delay margin of one sampling period, is chosen for the desired output sensitivity
function.
8. A constrained optimization program is run to find the vector of parameters. The results of
the optimization program are x "!0·3986, y "0·95 and y "0·4907. This solution
2 1 2
leads to the following weighting filter:

P (1!0·95z~1)(1!0·4907z~1)
¼(z~1)"S~1 (z~1)" D (41)
d A (z~1)(1!0·999999z~1)(1#0·3986z~1)
0
where

P "(1!1·1396z~1#0·3734z~2)(1#0·0886z~1#0·5135z~2)
D
Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.
 ROBUST DIGITAL CONTROL 205

9. The stability domain is reduced to a circle with radius 0·85 centred at a"0·15. Then, the
plant model and weighting filter equations are transformed accordingly to this new
stability domain (an r is used for the parameter in the new stability domain, for example,
¼r(z~1) denotes the weighting filter in the reduced stability domain). It should be noticed
that the reduced stability domain must contain the point [1, 0j ] in order to keep the
integrator in the weighting filter.
10. The weighting filter in the new system, which may have unstable poles, is factorized into an
all-pass function and a stable filter which will be used in H optimization.
=
11. The infinity norm of ¼r(z~1)Sr (z~1) should be minimized. First of all, using the
yp
Q parametrization Sr is written as follows:
yp
Br (z~1)
Sr (z~1)"1! 0 Q(z~1) (42)
yp Ar (z~1)
0
then, this problem is converted to the following model matching problem:
min E¹ !¹ Q E "c (43)
1 2 =
where ¹ "¼r and ¹ "¼r · Br /Ar . The optimum value of c is computed using the
1 2 0 0
Nehari’s theorem.18 Next, the parameters of Q are calculated using coprime factorization
and Nevanlinna—Pick’s algorithm in discrete time. Finally, the controller parameters are
determined using the following relation:
Rr Q
" (44)
Sr 1!QBr /Ar
0 0
where Rr/Sr is the controller in the reduced stability domain.
12. The controller is transferred back to the real stability domain and its order is reduced by
zero/pole cancellation techniques.
Although this controller satisfies the robustness margins, places the closed-loop poles in the
desired places and stabilizes two other plant models, it is still far from some of the specifications of
the benchmark problem. In order to improve the sensitivity function for the other plants, we can
reduce the desired output sensitivity function in the frequency region where the maximum
parameter uncertainty occurs. This region corresponds to the maximum of additive uncertainty
and is located between the frequency of the first vibration mode of the full loaded model and that
of the unloaded one. Introducing two pairs of complex zeros in this frequency region with an
appropriate damping factor as a fixed part in the numerator of the desired sensitivity function
leads to reducing the output and input sensitivity function in the critical region. Figure 5 shows
the desired sensitivity function (inverse of the weighting filter) before (dashed line) and after
introducing the following fixed part:
H (z~1)"(1!1·7052z~1#0·7753z~2)(1!1·4976z~1#0·6337z~2) (45)
S
corresponding to (u "5·65, f"0·45) and (u "8·29, f"0·55).
0 0
Figures 6 and 7 show the output and input sensitivity functions, respectively, for the three
models using the following simplified controller:
R(z~1)"0·4075!1·2444z~1#0·8792z~1
S(z~1)"1!0·5386z~1!0·4614z~2
¹(z~1)"0·4572!0·3232z~1!0·0912z~2

( 1998 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 8, 191—210 (1998)
206 I.D. LANDAU AND A. KARIMI

Figure 5. Desired sensitivity function

Figure 6. The output sensitivity functions for the different loadings

Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.
 ROBUST DIGITAL CONTROL 207

Figure 7. The input sensitivity functions for the different loadings

A second order system with u "6·3 (which corresponds to the first vibration mode of the full
0
load model) and f"0·8 is chosen for the tracking reference model. The parameters of the prefilter
¹(z~1) are computed independently from the controller. In fact the parameters of ¹(z~1) are the
results of an optimization program which minimizes the overshoot and rise time for the three
models. Figures 8 and 9 show the step and disturbance (filtered by 1/A) responses for the three
models.
Tables 1 and 2 show the simulation and experimental results of the controller in the different
loadings, respectively.
This controller satisfies 85% of the specifications of the benchmark in simulation. A compari-
son, in terms of satisfaction of performances and complexity with the controller designed by
Method A, the H controllers in the benchmark designed by Jones and Limebeer19 and
=
Walker,20 and the controller designed by Kwakernaak,17 is presented in Table 3. For a real time
comparison, only the time characteristics, i.e., rise time, settling time and disturbance rejection
time, are available. Therefore the normalized sum (with respect to the desired values) of these
characteristics for the three cases is used as a performance criterion (smaller criterion means
better performance). This criterion is then penalized by the normalized complexity of the
controller (defined as the sum of the order of the polynomial R, S and ¹ divided by 10) in order to
obtain a joint performance/complexity criterion. In Table 4 the controllers are compared with
these criteria. It is observed that the controller designed by Method B has a good result compared
with the other H approaches.
=

8. CONCLUSIONS
Two different techniques, based on pole placement combined with sensitivity function shaping,
have been presented. The first technique combines the main ideas in advanced robust control

( 1998 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 8, 191—210 (1998)
208 I.D. LANDAU AND A. KARIMI

Figure 8. Step and disturbance response in simulation

Figure 9. Step response in real time

Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.
 ROBUST DIGITAL CONTROL 209

Table 1. Simulation results

Rise time Over shoot Rej. of dist. Delay Modulus Atten. band Maximum
(s) (%) (s) margin (s) margin (dB) band (Hz) of S (dB)
up
No load 1·135 0% 1·811 50 !6·45 0·118 6·7
Half load 0·643 1·6% 1·703 345 !5·29 0·119 7·2
Full load 0·717 9·4% 2·105 560 !5·95 0·116 7.3

Table 2. Experimental results

Rise time (s) Overshoot (%) Settling time (s) Rej. of dist. (s)

No load 1·671 0% 2·7 2·55

Half load 0·714 0% 3·1 1·55
Full load 0·742 8·1% 2·1 1·75

Table 3. Comparison with other approaches in simulation

Controller Performance (%) criterion Complexity (order of R#S#¹ )

Method A 97·12 12
Method B 85·09 6
Kwakernaak 88·80 18
Jones et al. 94·38 35
Walker 72·35 15

Table 4. Comparison with other approaches in real-time application

Controller Normalized sum of time Joint performance/complexity

characteristics criterion

Method A 10·86 12·06

Method B 13·64 14·24
Kwakernaak 12·71 14·51
Jones et al. 11·47 14.97
Walker 22·63 24·10

with the classical pole placement for SISO systems without using complex calculations. It adjusts
iteratively the sensitivity functions in the frequency domain where it is necessary. Therefore, it
seems that this technique is easy to understand for practising control engineers who are not
familiar with the advanced robust control theories. This technique has already been applied to
several real and industrial systems.
In the second technique, the idea of pole placement combined with sensitivity function shaping
has been transferred to H optimal control via a new interpretation of the weighting filter, as the
=
inverse of the desired sensitivity function. In this technique, weighting filter selection, which seems

( 1998 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control, 8, 191—210 (1998)
210 I.D. LANDAU AND A. KARIMI

to be the most important step in the H design, is carried out automatically by an optimization
=
program. Using this method a solution to a benchmark problem has been given, tested on a real
system and compared with other H type designs.
=
Research is in progress concerning the proof of the convexity of the optimization procedure for
building the desired sensitivity function via a different parametrization.

REFERENCES
1. Landau, I.D., F. Rolland, C. Cyrot and A. Voda, ‘Digital robust control. The combined pole placement/sensitivity
shaping method’, Presented at the Summer Control School of Grenoble ‘Robustness Analysis and Design of Robust
Controllers’. French version published in ¸a Robustesse, A. Oustaloup (Ed), Hermes, Paris, 1993.
2. Landau, I.D., C. Cyrot and D. Rey, ‘Robust control design using the combined pole placement/sensitivity function
shaping method’, European Control Conference, Gronigen, The Netherlands, 1993.
3. Landau, I.D., A. Karimi, A. Voda and D. Rey, ‘Robust digital control of flexible transmission using the combined pole
placement/sensitivity function shaping method’, European Journal of Control, 1(2), (1995).
4. Fenot, C., F. Rolland, G. Vigneron and I.D. Landau, ‘Open loop adaptive feedback control of deposited zinc in hot
dip galvanizing, Control Engineering Practice, 5 (1993).
5. Zames, G. and B.A. Francis, ‘Feedback, minimax sensitivity, and optimal robustness’, IEEE ¹rans. Automat. Control,
28 (1983).
6. Francis, B.A. and G. Zames, ‘On H optimal sensitivity theory for SISO feedback systems’, IEEE ¹rans. Automat.
=
Control, 29 (1984).
7. Kwakernaak, H., ‘A polynomial approach to minimax frequency domain optimization of multivariable systems’, Int.
J. Control, 44 (1986).
8. Postlethwaite, I., M.C. Tsai and G.D.W., ‘Weighting function selection in H design’, 11-th IFAC World Congress,
=
Tallian, 1990.
9. Doyle, C.J., B.A. Francis and A.R. Tannenbaum, Feedback Control ¹heory, McMillan, NY, 1992.
10. Kwakernaak, H., ‘Robust control and H optimization’, Automatica, 29 (1993).
=
11. Landau, I.D., ‘Robust digital control of systems with time delay (the Smith predictor revisited)’, Int. J. Control, 62(2),
(1995).
12. Landau, I.D., System Identification and Control Design, Prentice Hall, NJ, 1990.
13. Krause, J.M., ‘Comments on Grimble’s comments on Stein’s comments on roll off of H optimal controllers’, IEEE
=
¹rans. Automat. Control, 37 (1992).
14. Sung, H.K. and S. Hara, ‘Properties of sensitivity and complementary sensitivity functions in single-input single-
output digital control systems’, Int. J. Control, 48 (1988).
15. M’Saad, M., M. Duque and S. Hammad, ‘Flexible mechanical structures adaptive control, in Adaptive Control
Strategies for Industrial ºse, C. Shah and G. Dumont (Ed), Springer Verlag, Berlin, 1989.
16. Landau, I.D., D. Rey, A. Karimi, A. Voda and A. Franco, ‘A flexible transmission system as a benchmark for robust
digital control’, European Journal of Control, 1(2), (1995).
17. Kwakernaak, H., ‘Symmetries in control system design’, in ¹rends in Control A. Isidori (Ed), Springer Verlag, London,
1995.
18. Francis, B.A., ‘A course in H control theory’, in ¸ecture Notes in Control and Information Sciences, number 88,
=
Springer Verlag, Heidelberg, 1987.
19. Jones, N.W. and D.J.N. Limebeer, ‘A digital H controller for a flexible transmission system’, European Journal of
=
Control, 1(2), (1995).
20. Walker, D.J., ‘Control of a flexible transmission—a discrete time H approach’, European Journal of Control, 1(2),
=
(1995).

Int. J. Robust Nonlinear Control, 8, 191—210 (1998) ( 1998 John Wiley & Sons, Ltd.