Proceedings of the 17th World Congress

The International Federation of Automatic Control

Seoul, Korea, July 6-11, 2008

Robust Real-Time Control of a Two-Rotor

Aerodynamic System
Petko H. Petkov Nicolai D. Christov
Mihail M. Konstantinov

Technical University of Sofia, 1000 Sofia, Bulgaria

(e-mail: php@tu-sofia.bg)

Universite des Sciences et Technologies de Lille, 59655 Villeneuve
dAscq, France (e-mail: Nicolai.Christov@univ-lille1.fr)

University of Architecture, Civil Engineering and Geodesy, 1046
Sofia, Bulgaria (e-mail: mmk fte@uacg.bg)

Abstract: This paper presents the design and experimental evaluation of a two-degree-of-
freedom discrete-time -controller for a laboratory two-rotor aerodynamic system with ten
uncertain parameters. The controller implemented is of 24th order and ensures robust stability
and robust performance of the closed-loop sampled-data system. This controller is realized on
a PC by using the Real Time Workshop of MATLAB R
with a sampling frequency of 100 Hz.
The experimental results are close to the results predicted by using the linearized model of the
system and highlight many of the difficulties associated with the practical implementation of
robust control laws.

Keywords: Robust control; aerodynamic system; -synthesis; real-time control.

1. INTRODUCTION

The real implementation of robust control laws is useful in

teaching the modern Robust Control Theory. Such imple-
mentation allows to attract the students attention to the
theoretical and practical difficulties associated with the
realization of high-order controllers in presence of uncer-
tainties, nonlinearities, disturbances and noises. The com-
bined usage of theoretical methods, software for computer-
aided design and sophisticated laboratory equipment helps
to build the necessary knowledge and engineering abilities
for control of complex systems.
This paper presents the design and experimental evalu-
ation of a two-degree-of-freedom discrete-time -controller
for a laboratory two-rotor aerodynamic system with ten
uncertain parameters. The controller implemented is of
24th order and ensures robust stability and robust per-
formance of the closed-loop sampled-data system. This Fig. 1. Two-rotor aerodynamic system
controller is realized on a PC by using the Real Time
position in the vertical plane, while the tail propeller
Workshop of MATLAB R
with sampling frequency of 100
controls the beam position in the horizontal plane. There
Hz. The experimental results are close to the results pre-
are two counter-weights fixed to the beam that determine
dicted by using the linearized model of the system and
the stable equilibrium position. The system is balanced in
highlight some of the difficulties associated with the im-
such a way, that when the motors are switched off, the
plementation of robust control laws. The controller design
main rotor end of beam is lowered. The control actions
and implementation are used in the laboratory exercises
are the motor supply voltages. The measured system
in a course on Robust and Optimal Control, taught in the
outputs are the two angles of beam deviation in the
Technical University of Sofia.
horizontal plane (azimuth angle) and in the vertical plane
(pitch angle). The motor control is realized in pulse-width
2. SYSTEM DESCRIPTION
modulation (PWM) mode.
The two-rotor aerodynamic system (TRAS) is shown in The TRAS control consists in stabilization of the beam in
Figure 1. There are two propellers at the both ends of arbitrary (in the practical limits) desired position (pitch
a beam, driven by DC motors, joined to the base with and azimuth) or tracking of a desired trajectory. In the
an articulation. The main propeller controls the beam design of robust control systems one usually implements

978-3-902661-00-5/08/$20.00 2008 IFAC 6422 10.3182/20080706-5-KR-1001.0187

17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

controllers obtained by H design or -synthesis, see Zhou
et al. (1995). In this paper we present the results from -
synthesis only, which has the advantage that it guarantees
the achievement of robust performance of the closed-loop
system.
on the system. The plant is two-channel and there is an
interaction between the two channels. In order to reveal
in full the dynamic behavior of the plant it should be
considered as multivariable, i.e. the two channels can not
be considered as independent.

3. DERIVATION OF THE UNCERTAINTY MODEL

Table 1. System parameters
In the design of robust control laws it is advisable to use
linearized models in which it is easy to set the uncertainty Symbol Description
in specific parameters. For this aim the models, obtained Ih moment of inertia of the tail rotor
by analytic linearization, are very appropriate. Jh moment of inertia with respect to the vertical axis
Iv moment of inertia of the main rotor
In the given case the plant nonlinear model is provided Jv moment of inertia with respect to the horizontal axis
by the manufacturer of the laboratory set-up in the form kHh velocity gain of the tail rotor
of nonlinear differential equations and algebraic relations kFh thrust coefficient of the tail rotor
as well as in the form of Simulink R
model, see Anon1 kHv velocity gain of the main rotor
kFv thrust coefficient of the main rotor
(2006). The model of the two-rotor aerodynamic system
kfh friction coefficient in the vertical axis
is linearized analytically under the usual assumption for kfv friction coefficient in the horizontal axis
small deviations of the variables, describing the system khv coefficient of the cross moment from tail rotor
behavior, around their trim values. to pitch
kvh coefficient of the cross moment from main rotor
to azimuth
Rv coefficient of the return torque
lm length of the main part of the beam
lt length of the tail part of the beam

The parameters of the linearized model are defined in

Table 1.
The coefficients kHh , kFh , kHv , kFv are determined
by linearization of the the corresponding rotor static
characteristics, obtained experimentally. The gain Rv is
the coefficient of the return torque corresponding to the
forces of gravity and depends nonlinearly from the pitch
angle v . After the linearization, this gain is obtained as
Rv = k1 sin(v,nom ) k2 cos(v,nom ) where k1 and k2
are constants. The inertial moment around the vertical
axis, Jh , is also a nonlinear function of the pitch angle
and is obtained as Jh = k3 cos2 (v,nom ) + k4 , where the
coefficients k3 and k4 are determined by the mass and
geometric sizes of the beam and devices mounted on it.

Table 2. Parameters of the linearized model

Parameter Value Units

Ih 1/37000 kgm2
Iv 1/6100 kgm2
Jv 3.00581 102 kgm2
kfh 5.88996 103 Nms/rad
kfv 1.27095 102 Nms/rad
khv 4.17495 103 Nm
kvh 1.78200 102 Nm
kHh 9.83891 103 rad/s
kFh 2.12932 105 Ns/rad
kHv 4.87457 103 rad/s
Fig. 2. Linearized model of the two-rotor aerodynamic kFv 3.07723 104 Ns/rad
lm 0.202 m
system
lt 0.216 m
k1 5.00576 102 Nm
The linearized TRAS model is shown in Figure 2 where k2 9.36008 102 Nm
the subindex nom denotes the nominal value of the corres- k3 2.37904 102 kgm2
ponding variable. The input variables are the voltages uh k4 3.00962 103 kgm2
and uv of the tail rotor and main rotor motors and output
variables are the azimuth angle h and pitch angle v . The nominal values of the model parameters are given in
The moments Mhd and Mvd are the disturbances acting Table 2.

6423
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

As uncertain parameters in the mathematical description Singular Value Plot of G

of the aerodynamical system we consider the moment of 60

inertia Jh in respect to the vertical axis, the thrust coeffi-

40
cients kFh , kFv of both rotors, the velocity gains kHh , kHv
of the two rotors, the friction coefficients kfh , kfv , the cross 20

moment coefficients kvh , khv , as well as the coefficient 0

Rv of the return torque, all together 10 parameters. The
uncertainties in the moment of inertia Jh and in the coef-

Magnitude (dB)
20

ficient Rv are due to their dependence on the pitch angle 40

v , the uncertainties in the coefficients kFh , kFv , kHh ,
and kHv are introduced as a result of the measuring and 60

approximation of the static characteristics of the rotors, 80

the uncertainties in the coefficients kfh and kfv are due to

the errors in determination of the friction moments, and 100

the uncertainties in the coefficients kvh and khv result from 120

simplification of the aerodynamic interaction between the

two channels. Further on we assume that the moment of 140
10
3 2
10
1
10 10
0 1
10
2
10
3
10

inertia Jh and the coefficients kFh , kFv , kHh and kHv are Frequency (rad/sec)

known with errors up to 10 % while the rest coefficients -

Fig. 4. Frequency response characteristics of the uncertain
with errors up to 5 %.
plant
The ten real uncertain parameters are set by using the
function ureal from Robust Control Toolbox, ver. 3, see
Balas et al. (2006). The uncertain TRAS model is obtained
on the basis of the block-diagram, shown in Figure 2,
implementing the function sysic from the same toolbox.
As a result one finds an uncertain state space object.
The uncertain TRAS system is described as a control plant
by the equation
Md
y=G ,
u
where
h uh Mhd
y= ,u = , Md = .
v uv Mvd

Fig. 5. Block-diagram of the closed-loop system with

performance requirements
The system has reference inputs (r), input disturbances
Fig. 3. Uncertain model of the TRAS system (d) and noise (n) introduced in measurement of the angles
h and v . (Here and further on the disturbance vector is
denoted for brevity by d.) The TRAS uncertain model is
The uncertain model of the two-rotor aerodynamic system
the state space object G.
is shown in Figure 3.
The system has two output signals (ey and eu ). The
Let us introduce the representation
block M is the ideal dynamics model that the designed
G = [Gd Gu ] closed-loop system should match to. The feedback of the
such that system is realized by the vector yc = y + Wn n, where the
y = Gd Md + Gu u. measurement noise n is a random vector with unit 2-norm
In the last expression Gd is the plant transfer function and Wn is the transfer matrix of the noise shaping filters.
in respect to disturbances and Gu is the plant transfer
The transfer function matrix M of the ideal matching
function with respect to control.
model is chosen as diagonal in order to suppress the
The frequency response plot of the uncertain plant singular interaction between the two
channels and
is taken as
values, obtained from the transfer function Gu , is shown wm1 0
in Figure 4. M (s) =
0 wm2
4. -SYNTHESIS OF A DISCRETE-TIME where
1
CONTROLLER wm1 = ,
1.5s2 + 1.2s + 1
The block-diagram of the closed-loop system that includes 1
wm2 = 2
.
the uncertain TRAS model, the feedback and the con- 2.0s + 1.6s + 1
troller, as well as the elements reflecting the performance In the choice of this model it was assumed that the azimuth
requirements, is shown in Figure 5. dynamics is faster than the pitch dynamics.

6424
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

The noise transfer function matrix

is taken as Inverse Performance Weighting Function

wn 0 30

Wn (s) = 1/wp1

0 wn 20
1/wp2

s
where wn = 102 s+1 is a high pass filter whose output is 10

significant above 10 rad/s.

0

To obtain good performance of the system responses we

Magnitude (dB)
shall implement a two-degree-of-freedom controller, see Gu 10

et al. (2005). The control actions are generated according 20

to the expression
30
r
u = [Kr Ky ] = Kr r Ky yc ,
yc 40

where Ky is the output feedback transfer function matrix 50

and Kr is the pre-filter transfer function matrix.

60
8 7 6 5 4 3 2 1 0 1 2
10 10 10 10 10 10 10 10 10 10 10
The control actions are realized by a computer in real time Frequency (rad/sec)

with sampling frequency fs = 100 Hz. For this reason the

-synthesis is done for this sampling frequency. Fig. 6. Inverse performance weighting functions
Let us denote by P (z) the transfer function matrix of
the discretized eighteen-input eighteen-output open-loop Control Action Inverse Weighting Function

system, that consists of the plant model plus the weighting 90

1/wu1
functions and let the block-structure P is defined as 1/wu2

80

0
P := : R1010 , F C64
0 F 70

The first block of the matrix P , the block , corresponds

60
to the parametric uncertainties, included in the model of
Magnitude (dB)

the aerodynamic system. The second block F is a ficti- 50

tious uncertainty block, used to include the performance

requirements into the framework of the -approach. The 40

inputs of this block are the weighted error signals ey and

eu , and the outputs are the exogenous signals r, d and n. 30

The aim of the -synthesis is to find a stabilizing controller 20

K, such that for each frequency [0, /Ts ], where

10
Ts = 2/fs , the structured singular value satisfies the 10
0 1
10 10
2

Frequency (rad/sec)
10
3 4
10
5
10

condition
P [FL (P, K)(j)] < 1, Fig. 7. Inverse control action weighting functions
where FL (P, K) is the closed-loop transfer function mat-
rix. The fulfilment of this condition guarantees the robust
performance of the closed-loop system, i.e. respectively. The control weighting functions are chosen as
high pass filters with appropriate bandwidth in order to

Wp (So Gu Kr M ) Wp So Gd Wp So Gu Ky Wn

W u S i Kr Wu Si Ky Gd Wu Si Ky Wn
impose constraints on the spectrum of the control actions.

<1 The experiments with the control laws designed shows that
for all uncertainties with kk < 1, where Si and So are the behavior of the real closed-loop system is very sensitive
the input and output sensitivity functions, respectively. to the weighting functions used. That is why the precise
tuning of the weighting functions requires a large volume
The -synthesis is done for several performance weighting of experiments.
functions that ensure a good balance between system per-
formance and robustness. On the basis of the experimental The -synthesis is performed by using the function dksyn
results, we choose the performance weighting function from Robust Control Toolbox. Five iterations are per-
formed that decrease the maximum value of to 0.997.
2 80s + 1

8.4 10 0.01 The final controller obtained is of 24th order. At first
Wp (s) =
80s + 103
1 500s + 1
glance, the order of this controller is very high, which
0.03 7.5 10 eventually makes difficult its real time implementation.
500s + 10 3

and the control weighting function In fact, the experiments show that with the relatively low
sampling frequency and fast processor, the implementation
5 0.05s + 1

4.8 10 0 of this controller does not create difficulties. That is why
Wu (s) =
104 s + 1 .

controller order reduction is not necessary.
4 0.1s + 1
0 2.304 10
10 s + 1
4 The frequency response of , corresponding to the robust
performance analysis, is shown in Figure 8. Obviously,
The frequency responses of the inverse performance and the closed-loop system achieves robust performance, which
control weighting functions are given in Figures 6 and 7, also guarantees robust stability of this system.

6425
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

1
Robust performance The control actions for different values of the uncertain
parameters are shown in Figure 10.
0.9
The experiments with the controller designed show a
0.8
strong influence of the control action constraints (the
0.7 actuators saturation) on the performance of the closed-
0.6
loop system. In some cases the usage of weighting functions
ensuring very good transient responses of the linear closed-
loop system leads to generation of auto-oscillations in the
mu

0.5

0.4 real system and loss of stability. A very serious influence

on the dynamics of the closed-loop system may have the
0.3
noises at the actuator inputs, whose level is very high in
0.2 the case under consideration. This justifies the inclusion
0.1
of the noises in the controller design.
upper bound
lower bound
Singular Value Plot of the Closedloop Transfer Function Matrix
0
3 2 1 0 1 2 3
10 10 10 10 10 10 10 20
Frequency (rad/sec)

Fig. 8. Closed-loop system robust performance 10

Closedloop Transient Responses 0

0.6

Singular Values (dB)

0.4
10
0.2

0
h

20
0.2

0.4
30
0.6

0.8
0 20 40 60 80 100 120
40
Time (secs)

0.2 50
1 0 1
10 10 10
Frequency (rad/sec)
0.1

Fig. 11. Singular value plot of the closed-loop system

v

0.1
The frequency responses of the singular values of the
0.2 closed-loop transfer function matrix for random values of
0 20 40 60 80 100 120
Time (secs) the plant uncertain parameters are shown in Figure 11. It
is seen that as a result of achieving robust performance,
Fig. 9. Transient responses of the uncertain linear closed- the closed-loop system frequency responses are close to
loop system these of the model (shown with dashed lines).
The transient responses of the sampled-data closed-loop Singular Value Plot of the Sensitivity Transfer Function Matrix

system, obtained for different random values of the uncer- 40

tain parameters, are shown in Figure 9.

20

Azimuth and pitch control actions

0.4
0

0.2
Singular Values (dB)

20
uh

40
0.2

0.4 60
0 20 40 60 80 100 120
Time (secs)

80
0.06

0.04
100
8 6 4 2 0 2
10 10 10 10 10 10
0.02 Frequency (rad/sec)
uv

0.02 Fig. 12. Singular value plot of the sensitivity function

0.04

0.06
0 20 40 60 80 100 120
The frequency response plot of the singular values of dis-
Time (secs) turbance transfer function matrix (output sensitivity func-
tion So ) is shown in Figure 12. (The inverse performance
Fig. 10. Control actions of the uncertain linear closed-loop functions are shown with dashed lines.) The disturbance
system attenuation is more than 100 times (40 dB).

6426
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

Singular Value Plot of the Noise to Input Transfer Function Matrix Azimuth Control
1
100

0.5
80

uh
0
60

0.5
Singular Values (dB)

40

1
0 20 40 60 80 100 120
20
Time (secs)

Pitch Control
0 1

20
0.5

uv
40

0
60
0 1 2
10 10 10
Frequency (rad/sec)
0.5
0 20 40 60 80 100 120
Time (secs)
Fig. 13. Frequency response of the noise-to-control contour
The frequency response of the noise-to-control contour Fig. 15. Experimental control actions (noisy data)
(Figure 13) shows that one may expect high level of the Azimuth control
noises at the actuator inputs. 0.4

The experiments with the controller designed are done by 0.2

using the laboratory set-up, shown in Figure 1 along with

a PC with MATLAB
uh

R 0
, ver. 7.1. The generation of the C
driving program is done by using the Real Time Workshop, 0.2

see Anon2 (2006). For this aim a Simulink R

model of the
closed-loop system is used with built-in driver for interface 0.4
0 20 40 60 80 100 120
Time (secs)
with the plant.
Pitch control
0.5
Experimental transient responses
0.6
0.4
0.4
uv

0.2 0.3
h

0.2
0.2

0.4 0.1
0 20 40 60 80 100 120
Time (secs)
0 20 40 60 80 100 120
Time (secs)

Fig. 16. Experimental control actions (after filtration)

0.1 corresponding results are shown in Figure 16 and are close
0.05 to the theoretical results shown in Figure 10.
0
v

0.05
ACKNOWLEDGEMENTS
0.1

0.15

0.2
The authors are grateful to Fadata Company whose con-
0 20 40 60
Time (secs)
80 100 120
tinuous support to the Department of Systems and Con-
trol of the Technical University of Sofia made possible to
Fig. 14. Experimental transient responses perform the experiments presented in this paper.

REFERENCES
The experimentally obtained transient responses of the
closed-loop system, controlled in real time with the - Anon1. Two Rotor Aerodynamical System. Users Manual.
controller designed, are shown in Figure 14. The compar- Inteco Ltd., 2006. http://www.inteco.com.pl
ison with the transient responses of the linear closed-loop Anon2. Real-Time Workshop Users Guide. The Math-
system given in Figure 9 shows a good coincidence between Works, Inc., Natick, MA, 2006.
the theoretical and experimental results. G. Balas, R. Chiang, A. Packard, M. Safonov. Robust
Control Toolbox. The MathWorks, Inc., Natick, MA,
The experimentally obtained control actions are shown in 2006.
Figure 15. Due to the high level of the noises at the actu- D.-W. Gu, P.Hr. Petkov, M.M. Konstantinov. Robust
ator inputs, the control actions are severely contaminated Control Design with MATLAB R
. Springer, London
by errors. 2005.
To extract the true actuator inputs, the control signals K. Zhou, J. C. Doyle, K. Glover. Robust and Optimal
are filtrated by using first order Butterworth filter. The Control. Prentice Hall, Upper Saddle River, NJ, 1995.

6427