Copyright (c) 2005 IFAC.

All rights reserved

16th Triennial World Congress, Prague, Czech Republic

ADAPTIVE CONTROL OF TWIN ROTOR MIMO SYSTEM: POLYNOMIAL

APPROACH

Marek Kubalk, Vladimr Bobl, Petr Chalupa

Department of Process Control, Faculty of Technology,

Tomas Bata University in Zlin, Nm. TGM 275, 762 72 Zln, Czech Republic,
Tel.: +420 57 603 3217, Fax: +420 57 603 3333, E-mail: kubalcik@ft.utb.cz

Abstract: This paper presents real time control of the Twin Rotor MIMO System
laboratory model. The objective laboratory model is a multivariable nonlinear unstable
system of high order. It is not possible to stabilize the system and satisfactorily track
reference signals using classical controllers with fixedly set parameters. Two methods
based on self tuning control utilizing adaptive approach are discussed. Both designed
methods are based on polynomial approach. In first case an algorithm taking into account
internal interactions among input and output variables was used. The second method
utilizes the principle of decentralized control with an additional logical supervisor for
switching of the recursive identification in particular loops. In both cases stability of
closed loop system was ensured and after an adaptation phase the asymptotic tracking of
reference signals was achieved. Quality of control achieved by particular methods is
compared and discussed. Copyright 2005 IFAC

Key words: Polynomial methods, Nonlinear system, Self-tuning control, Multivariable

control, Decentralized control, Real-time control.

1. INTRODUCTION disadvantage lies in influences of its faults to the

In this paper, a comparison of two approaches for
adaptive control of a multivariable laboratory model
(Twin Rotor MIMO System) is presented.
The classical approach to the control of MIMO
systems is based on a design of a matrix controller
used to control all system outputs at a time. The
computation of matrix controller is realized by one
central computer. Basic advantage of this approach is
possibility of reaching optimal control courses
because the controller can use all information known
about the controlled system. Disadvantage of usage
of the central matrix controller is its demands to
computer resources because the number of operations
and consumed memory depend on the square of the
number of controlled signals. Nowadays this problem
is reduced thanks to great progress in the
development of computer hardware and leads just to
the increased price of the control system. Another
controlled system. In case of failure of central
controller, all the controlled signals are afflicted and
thus the reliability of the controller is fundamental.
Reaching of required reliability can then be
unbearable from the financial point of view,
especially in critical applications.
Alternative solution for the control of MIMO systems
is usage of decentralized approach. In this case, the
system is considered as a set of interconnected
subsystems and the output of each subsystem is
influenced not only by the input to this subsystem but
also by the input to the other subsystems. Each
subsystem is controlled by stand - alone controller.
Thus, the decentralized control is based on
decomposition of the MIMO system to subsystems
and design of a controller for each subsystem (Cui
and Jacobson, 2002). Another advantage of the
decentralized approach is, that setting of controller
parameters (in this case a choice of poles of the

892
characteristic polynomial) is for SISO control loops
a lot more easier than for MIMO control loops. On
the other hand, the control courses of decentralized
control system are suboptimal because the controllers
do not use information from the other subsystems.
Disadvantageous is also a limited applicability of the
decentralized control only for symmetric systems
(systems with an equal number of inputs and
outputs).
2. TWIN ROTOR MIMO SYSTEM
The real-time laboratory model Twin Rotor MIMO
System (producer Feedback Instruments, LTD
United Kingdom) is shown in Fig 1. This system
provides a high-order, non-linear system with
significant cross-coupling. The main parts of the
system are the pedestal, the jib connected to pedestal
and two propellers at the ends of the jib. The system
jib can freely rotate around vertical axes by about
330 degrees (process output y1 ( t ) ) and horizontal
axis by about 100 degrees (process output y2 ( t ) .
The system inputs u1 ( t ) , u2 ( t ) are the voltages used
to drive motors of the propellers and outputs are
angular rotations with respect to horizontal and
vertical axes..
1 2 3 4 5
1 small propeller
2 jib
3 measurement of
rotation angles
4 pedestal
5 big propeller
Fig. 1. Twin rotor MIMO model (helicopter)
Despite the strong interactions in the system,
decomposition to subsystems is straightforward: the
first subsystem consists of the small propeller which
drives the angular rotation around vertical axis, the
second subsystem consists of the big propeller
driving the angular rotation around horizontal axis.
Before the control circuit was connected as a closed
loop, the experiments obtaining a static characteristic
of the systems had been performed. The influence of
the first system input to the second system output is
small but the course of the second output is a sign of
nonlinearity of the system. Another problem of the
control of this system is a big hysteresis which is
present in the system. The static characteristics of
first subsystem, which was measured for increasing
and decreasing input signal, are shown in Fig. 2. The
great influence of changes of second system input to
the first output was confirmed by this measurement.
893
300
250
200
150
y1 - vertical axis
100
50
0
-50 u2 = -0.25
-100 u2 = 0.05
-150
-200
-0.25 -0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25
u1 - small propeller
Fig. 2. Static characteristics of the first subsystem
showing hysteresis
3. DECENTRALIZED CONTROL USING LOGIC
SUPERVISOR
Each output of multivariable controlled system can
be affected by each system input. The measure of the
affection is determined not only by cross - coupling
of the MIMO system but also by the course of the
system input signals. When the decentralized
approach is used to control such a system then, from
the point of view of controller in particular
subsystem, the transfer function varies in time even if
the MIMO system is linear and stable.
The presence of subsystems interconnections is the
main reason for using self-tuning controllers in
decentralized approach to ensure required course of
controlled variables. Identification algorithms
suitable for usage in decentralized control must
include weighting of identification data where new
data affect estimation of model parameters more then
older data. This requirement is a consequence of the
time varying influences of the other subsystems to
the identified subsystem. The influence of control
variable (ui) to the corresponding controlled variable
(yi) decreases with increasing gain of subsystems
interconnections. This could lead to unstable process
of recursive parameter estimates. The stability of
recursive identification can be increased by ensuring
that just one of the controllers connected to the
multivariable systems works in adaptive regime in
particular time. Recursive identification parts of other
controllers are suspended and parameter model
estimates are constant for that time. The process of
switching on and off the recursive identification is
controlled by a new part of a control circuit the
supervisory system. Switching the identification on
and off can be described as a process of transferring
token among subsystems where only the controller,
which currently has a token, can perform recursive
identification. The token is move to other subsystem
when the selected criterion is fulfilled.
The supervisory system represents a second level of a
control and thus a control circuit with supervisory
system has a hierarchical structure of control (Bobl,
et al. 2004).
u_in(k)
The inclusion of supervisory logic into control circuit y(k) u(k)
w(k)
brings the problem of defining a strategy for id_cntrl R1 u1 y1
switching the identification of individual subsystems Reference adaptive
controller 1 y1
signal 1
on and off i.e. moving the token between subsystems.
Three basic approaches can be used in deciding when E1 R1
to suspend identification of particular subsystem and E2 R2
u2 y2
move token to the other one: Logic supervisor y2
y on base of elapsed time,
u_in(k) u(k)
y on base of values from the currently identified y(k) Controlled
w(k)
subsystem, id_cntrl R2 process
y on base of values from the other subsystems. Reference adaptive
signal 2 controller 2
It is also possible to combine these approaches. The
basic ideas of these approaches are discussed in
Fig. 3. Simulink control circuit with TITO controlled
(Chalupa, 2003).
system
A logic supervisor has been proposed to utilize and Y ( z) B( z 1 )
GW ( z ) = = (1)
simplify the design of supervisory logic. This W ( z ) A( z ) P( z ) + B( z 1 )[Q( z 1 ) + ]
1 1

approach is suitable for usage in real-time industrial

applications. The idea of logical supervisor is based where
on the following two principles:
A( z 1 ) P( z 1 ) + B( z 1 )[Q( z 1 ) + ] = D(z 1 )
y Assigning priorities of individual subsystems
(2)
y On-line evaluation of criterions for each
subsystem
The logic supervisor was tested in connection with is the characteristic polynomial of the closed loop.
the controllers from the Self-Tuning Controllers
Library (Bobl and Chalupa, 2002). The properties of Polynomials
controllers were tested in MATLAB/Simulink
environment using control scheme shown in Fig. 3. A( z 1 ) = 1 + a1 z 1 + a 2 z 2 , B( z 1 ) = 1 + b1 z 1 + b2 z 2 (3)
The model contains continuous TITO system
controlled by two self-tuning controllers and logical represent the controlled system and
supervisor which provides identification switching
between input-output pairs. Model also includes P( z 1 ) = (1 z 1 )(1 + z 1 ) (4)
saturation blocks applied to control values. 1 1 1
Q ( z ) = (1 z )(q0 q2 z )
The quality of the control process is affected by
many parameters e.g. sampling period, algorithm of are polynomials of the controller. From (4) it is
control law, presence algorithm of logical supervisor, obvious that the controller contains the integrator.
saturation, initial parameter estimates. The on-line
identification uses recursive least squares method The characteristic polynomial
with adaptive directional forgetting.
D( z 1 ) = 1 + d 1 z 1 + d 2 z 2 (5)
The same controller configuration as in case of the
MIMO control loop design has been chosen for the is described by the dominant poles for a second
control of the Twin Rotor MIMO System. The order continuous-time model
structure of the SISO control loop is shown in Fig. 4.
The transfer function of the closed control loop is
given by the relation D ( s ) = s 2 + 2n s + n2 (6)

w e 1 u ( )
B z 1 y where the dominant poles are given by the desired
( )
P z 1 A( z )
1 damping factor and the natural frequency n of the
closed-loop. Then the controller algorithm, so called
Q(z-1) PP2B-1, is given by following equation (Bobl and
Chalupa, 2002)

u ( k ) = ( q0 + ) y ( k ) + ( q0 + q2 ) y ( k 1)
(7)
Fig. 4. Control loop with SISO controller q2 y ( k 2 ) ( 1) u ( k 1) + u ( k 2 ) + w ( k )

894
 4. MULTIVARIABLE CONTROL y 1 (k ) = a1 y1 (k 1) a 2 y1 (k 2 ) a 3 y 2 (k 1)
a 4 y 2 (k 2 ) + b1 u 1 (k 1) + b2 u 1 (k 2 ) +
4.1. Mathematical Model of the Twin Rotor MIMO
System + b3 u 2 (k 1) + b 4 u 2 (k 2)

The examined twin motor MIMO system is a typical y 2 (k ) = a 5 y1 (k 1) a 6 y1 (k 2 ) a 7 y 2 (k 1)

example of a two inputs two outputs system with a 8 y 2 (k 2 ) + b5 u 1 (k 1) + b6 u 1 (k 2 ) + (13)
significant cross coupling. Internal structure of the
+ b7 u 2 (k 1) + b8 u 2 (k 2 )
kind of system is shown in Fig. 5
4.2. Design of 2DOF Controller

The structure of the closed loop, shown in Fig. 6, was

presented in (Ortega and Kelly, 1984).

Fig. 5. A two input two output system

The transfer matrix of the system is defined as Fig. 6. Block diagram of 2DOF configuration
Generally, the vector W (z ) of input reference signals
G (z ) G12 (z )
Y (z ) = G (z )U (z ) = 11 U (z ) (8) is specified as
G 21 (z ) G 22 (z )
W (z ) = F w
1
(z )h(z )
1 1
(14)
where
Here, the reference signals are considered as a class
U (z ) = [u 1 (z ), u 2 (z )]
T of step functions. In this case h(z 1 ) is a vector of
(9)
is the vector of manipulated variables and constants and Fw (z 1 ) is expressed as

Y (z ) = [ y1 (z ), y 2 (z )] 1 z 1 0
T
(10)
Fw (z 1 ) = (15)
0 1 z 1
is the output vector.
The compensator F (z 1 ) is a component formally
It is possible to assume that the dynamic behaviour of separated from the controller. It has to be included in
the system can be described in the neighbourhood of the controller to fulfil the requirement on the
steady state by the discrete linear model in the form asymptotic tracking. If the reference signals are of
of the matrix fraction (Kuera, 1991) the same class as the step functions are, then F (z 1 )
G (z ) = A 1 (z 1 )B(z 1 ) = B1 (z 1 )A11 (z 1 ) (11) is an integrator.

Where polynomial matrices AR22[z-1], BR22[z-1] It is possible to derive the following equation for the
are the left coprime factorization of matrix G(z-1) and system output (operator z-1 will be omitted from some
matrices A1R22[z-1], B1R22[z-1] are the right operations for the purpose of simplification)
coprime factorization of G(z-1). Y = A 1 BU = A 1 BF 1 P 1U 1 (16)

The algorithm was designed for the following model

Where
with second order polynomials
U 1 = (W Y ) QFY (17)
1 + a1 z 1 + a2 z 2 a3 z 1 + a4 z 2 The corresponding equation for the controllers
A(z 1 ) = (12) output, as shown in the block diagram in Fig. 6,
a5 z + a6 z 1 + a7 z 1 + a8 z 2
1 2

follows as
b z 1 + b z 2 b3 z 1 + b4 z 2
B (z 1 ) = 1 1 2 2 U = F 1 P 1U 1 (18)
b5 z + b6 z b7 z 1 + b8 z 2
The substitution of U1 and Y results in
This model proved to be effective. Expression (12)
can be transcribed to the difference equations of the U = F 1 P 1 [ (W A1BU ) QFA1BU ] (19)
model

895
The equation (19) can be modified using the right u 2 (k ) = 3 e1 (k ) + 4 e 2 (k ) q 5 y1 (k )
matrix fraction of the controlled system into the form (q 6 q 5 ) y1 (k 1) + q 6 y1 (k 2 ) q 7 y 2 (k )
U = A1 [PFA1 + ( + FQ)B1 ]W (20) (q 8 q 7 ) y 2 (k 1) + q 8 y 2 (k 2 )
The closed loop system is stable when the following ( p 4 1)u 2 (k 1) + p 4 u 2 (k 2 ) p 3 u1 (k 1) +
diophantine equation is satisfied + p 3 u1 (k 2 )
PFA 1 + ( + FQ )B1 = M (21)
5. EXPERIMENTAL RESULTS
Where R22 z [ ] is a stable diagonal polynomial
1
The Twin Rotor MIMO System is a nonlinear system
matrix
with variable parameters that is practically too
1 + m1 z 1 + m2 z 2 + complex to be controlled using controllers with fixed
0 parameters. The nonlinear dynamics was in the
+ m3 z 3 + m 4 z 4
M (z 1 ) = (22) neighbourhood of a steady state described by the
1 + m1 z + m2 z 2
1
+
0 linear models given in previous sections. Adaptive
+ m3 z 3 + m 4 z 4 control was performed using the designed controllers.
and the structure of the matrices P, Q and was The sampling period was chosen in both cases
chosen as follows T0 = 0.5s . The other parameters for the decentralized
1 + p1 z 1 p2 z 1 control - damping factor = 10 and natural
P (z 1 ) = (23)
p3 z
1
1 + p4 z 1 frequency n = 1 were chosen in virtue of several
q + q 2 z 1
q3 + q 4 z 1
realized experiments. The matrix M (the pole
Q (z 1 ) = 1 placement of the multivariable controller) was also
q 5 + q 6 z q 7 + q 8 z 1
1

obtained from a number of experiments in the form

2
(z 1 ) = 1 1 0,9 z 1 + 0,19 z 2
3 4 0
M ( z 1 ) =
0,009 z 3 0,002 z 4 (27)
The solution of the diophantine equation results in a 1 0,9 z + 0,19 z
1 2

0
set of sixteen algebraic equations with unknown 0,009 z 3 0,002 z 4
controller parameters. Using matrix notation the
algebraic equations are expressed in the following The initial parameter estimates were chosen without
form any previous information in both cases.
The control courses of twin rotor MIMO system
1 0 b9 0 b13 0 b9 b13 p1 m1 a 9 + 1

b14 p 2 m 2 + a 9 a10
using both controllers are presented in Fig. 7, Fig. 8,
a9 1 a13 b10 b9 b9 b14 b13 b13 b10
a10 a 9 a14 a13 b10 b10 b9 b14 b14 b13 0 0 q1 m3 + a10 Fig. 9 and Fig. 10. These figures demonstrate that

a10 a14 0 b10 0 b14 0 0 q2
=
m4 the strong nonlinear, unstable and high order system
0 1 b11 0 b15 0 b11 b15 q 3 a11

a11 a15 1 b12 b11 b11 b16 b15 b15 b12

b16 q 4 a11 a12
can be stabilized and also quite good asymptotic
a a
12 11 a16 a15 b12 b12 b11 b16 b16 b15 0 0 1

a12

tracking can be achieved by using adaptive control
a12 a16 0 b12 0 b16 0 0 2 0 without apriori information about model of the
process.
1 b13 p3 a13

0 b9 0 b13 0 b9


Table 1 contains the values of the control quality
a9 1 a13 b10 b9 b9 b14 b13 b13 b10 b14 p 4 a13 a14
a10 a9 a14 a13 b10 b10 b9 b14 b14 b13 0 0 q5 a14 criterions. A criterion is a sum of powers of tracking
errors and a sum of increments of manipulated
a10 a14 0 b10 0 b14 0 0 q6
=
0
0 1 b11 0 b15 0 b11 b15 q 7 m5 a15 + 1 variables. The table contains the values after 50 s

a11 a15 1 b12 b11 b11 b16 b15 b15 b12 b16 q8 m6 + a15 a16
a a when the identified parameters became steady.
a a b12 b12 b11 b16 b16 b15 0 0 3 m7 + a16
12 11 16 15
a12 a16 0 b12 0 b16 0 0 4 m8
(24)
The controller parameters are derived by solving
these equations. The control law apparent from the
block diagram is defined as
FPU = E FQY (25)

The control law in the form of difference equations is

defined by the following expression
u1 (k ) = 1 e1 (k ) + 2 e 2 (k ) q1 y1 (k )
(q 2 q1 ) y1 (k 1) + q 2 y1 (k 2 ) q 3 y 2 (k )
(q 4 q 3 ) y 2 (k 1) + q 4 y 2 (k 2 ) (26)
( p1 1)u1 (k 1) + p1u1 (k 2 ) p 2 u 2 (k 1) +
+ p 2 u 2 (k 2 ) Fig. 7. Control courses using the matrix controller

896
 6. CONCLUSIONS

In case of the decentralized control it was necessary

to choose a large value of the damping factor . It
means approximation of the dynamic behaviour of
the process by an over damped second order model.
It caused a slow approach of the controlled variables
to the reference signals. On the other hand,
oscillations of both controlled variables and the
manipulated variables were significantly damped.
This is quite important for this kind of process. The
multivariable controller ensured faster asymptotic
tracking of the reference signals. The courses of the
Fig. 8. Control courses using the matrix controller controlled variables and the manipulated variables
controllers output appeared to have more oscillatory characters. These
facts are also evident from the values in Table 1.
Sums of powers of tracking errors are greater in case
of the decentralized control (slow courses), whilst
sums of increments of manipulated variables are
greater for multivariable control (oscillations of
manipulated variables).

Despite the fact, that relatively simple control

algorithms without any sophisticated numerical
computation were used, quite good control results
were achieved by both approaches. An advantage of
the proposed strategies lies in their simplicity and
applicability in industrial practice.

Fig. 9. Control courses using the decentralized ACKNOWLEDGEMENTS

controller
This work was supported in part by the Grant Agency
of the Czech Republic under grants No.102/05/0271,
102/04/P244 and in part by the Ministry of Education
of the Czech Republic under grant MSM
7088352101.

REFERENCES

Bobl, V. and P. Chalupa (2002). Self-Tuning

Controllers Simulink Library. http:
//www.utb.cz/stctool/.
Bobl, V., P. Chalupa and P. Dostl (2004).
Decentralized control of nonlinear systems
using self-tuning controllers. In: Proc. of the
IFAC Workshop on ALCOSP 04, Yokohama,
Fig. 10. Control courses using the decentralized Japan, 385-390.
controller controllers output Chalupa, P. (2003). Discrete Decentralized Control
Systems. Ph.D. Thesis, Tomas Bata University
Table 1 Control quality criterions in Zlin, Faculty of Technology, Zlin.
Cui, H and E. W. Jacobsen (2002). Performance
limitations in decentralized control. Journal of
Controller
e e u u
2 2 2 2
1 2 1 2 Process Control, 12, 485-494.
Kuera, V. (1991). Analysis and Design of Discrete
Multivariable 136,75 694,45 0,0012 0,0074 Linear Control Systems. Academia, Prague.
Ortega, V., and Kelly R.: (1984). PID self tuners:
Decentralized 182,66 2784,2 0,0004 0,0003 Some theoretical and practical aspects. IEEE
Trans. Ind. Electron., IE - 3, 332 338.

897