Model Reference Control of Nonlinear MIMO Systems by Dynamic
Output Feedback Linearization of ANARX Models
Juri Belikov and Eduard Petlenkov
Abstract— A dynamic output feedback linearization algo-
rithm for a model reference control of nonlinear multi-input
multi-output (MIMO) systems identified by an Additive Nonlin-
ear Autoregressive eXogenous (ANARX) model is introduced.
ANARX structure of the model can be obtained by training a
neural network with a specific restricted connectivity structure.
Linear discrete-time reference models are given in the form of a
transfer matrix defining desired zeros and poles of a closed loop
system. ANARX or NN-based ANARX models can be linearized
by the proposed linearization algorithm in a such way that the
transfer matrix of the linear closed loop system corresponds to
the given matrix of the reference models.
I. INTRODUCTION
The model matching problem is a popular problem in the
control systems theory. It has to be solved to design a model
reference controller to predefine the desired behavior of the
control system. This problem was considered and deeply
analyzed by different authors within the state space [5], [4],
[9] and input-output [8] approaches. Each of the algorithms
can be applied to a certain class of systems. In this paper
we propose the approach which can be applied to systems
identified by Additive Nonlinear Autoregressive Exogenous
(ANARX) model. In this case there exists an analytical
compensator, which makes possible the linearization of a
nonlinear controlled system by dynamic output feedback and
matching a linear reference model, what allows to apply
classical pole placement approach to design of the nonlinear
control system.
Class of neural networks based Additive Nonlinear Au-
toregressive Exogenous (NN-ANARX) models [6] proved
itself to be a perfect bridge allowing to combine NN-based
identification with classical control techniques [17]. In spite
of restrictions imposed on neural network’s connectivity, the
accuracy of identified model does not differ much compared
to the model obtained by means of fully connected neural
network. The lower number of synaptic weights in ANARX
model results in lower computational complexity which
can be important in real time applications. Applicability
of ANARX structures ranges from controlling a backwards
motion of a truck trailer [2] to modeling motions of surgeon’s
hand during medical surgery [12].
This work was partially supported by the Estonian Science Foundation
grant #6837 and by Nations Support Program for the ICT in Higher
Education ”Tiger University”.
J. Belikov is with Institute of Cybernetics, Tallinn University
of Technology, Akadeemia tee 21, 12618, Tallinn, Estonia
jbelikov@cc.ioc.ee
E. Petlenkov is with Department of Computer Control, Tallinn
University of Technology Ehitajate tee 5, 19086, Tallinn, Estonia
petlenkov@dcc.ttu.ee
Application of the algorithm proposed in [7] to lineariza-
tion of ANARX model leads to the finite discrete-time linear
closed loop system y(k + n)=v(k), where v(k) is the
reference signal, y(k + n) is the output of the model and n
is the order of the model.
In practice, when applying this algorithm to control of
nonlinear systems identified by ANARX or NN-ANARX
models [17], we also expect to obtain y(k+n)=v(k), where
y(k) is the output of the controlled system. However, this
may result in an undesirable behavior of the control system
(for example, significant overshootings, high control signals,
etc.). The latter means that the dynamics of the closed loop
system significantly depend on the identified model of the
controlled system. However, without loss of generality, we
can assume in our calculations that the identified model is
perfect, what means that the error between model and the
system equals zero and therefore ˆy(k)=y(k).
Thus, an algorithm for model reference control of non-
linear single-input single-output systems based on dynamic
output feedback linearization of ANARX models was pro-
posed in [14]. It results in the closed loop system, represented
by the following equation
y(k)+a
1
y(k − 1) + ...+ a
n
y(k − n)=
b
1
v(k − 1) + ...+ b
n
v(k − n), (1)
where a
1
,...,a
n
and b
1
,...,b
n
are parameters of the refer-
ence model or equation
Y (z)=G(z) · V (z),
where G(z) is a linear discrete-time reference model defining
the dynamics of the closed loop control system. Thus, desired
zeros and poles of the closed loop system can be predefined
on the stage of designing the control system.
The main contribution of this paper is devoted to the exten-
sion of the model reference control of nonlinear single-input
single-output systems based on dynamic output feedback
linearization of ANARX model [14] to the case of multi-
input multi-output systems.
A. Outline of the paper
The next sections of the paper are organized as follows.
Section II contains the concise description of ANARX and
NN-based ANARX structures. Besides that, it provides a
brief overview of the previous research on the problem of
control signal calculation. In Section III the model reference
control of nonlinear multi-input multi-output systems is
presented. Concluding remarks are drawn in the last section.
2009 IEEE International Conference on Control and Automation
Christchurch, New Zealand, December 9-11, 2009
WePT2.3
978-1-4244-4707-7/09/$25.00 ©2009 IEEE 536