Page 150 - Neural Network Modeling and Identification of Dynamical Systems

140 4. NEURAL NETWORK BLACK BOX MODELING OF AIRCRAFT CONTROLLED MOTION

equipment and damage to its structure, leading to the approximate nature of the ANN model,
to a change in the dynamic properties of the the real values of the variables describing the
aircraft and/or impeding its piloting. An effec- motion of the object inevitably differ from those
tive way to solve this problem is to use adaptive obtained as the outputs of the ANN model, an
control laws that fill in insufficient and/or inac- error appears that decreases the quality of con-
curate data about an object in the course of its trol. We propose one possible approach to de-
operation. crease this error in Section 4.3.2.3. This approach
All these adaptive control schemes require treats the inaccuracy of the ANN model as a dis-
the presence of a controlled object model. As turbance effect on the system that leads to a de-
was shown in Chapter 2, neural network imple- viation of the trajectory of the real object from
mentation of these schemes, which have high the reference trajectory. We attempt to reduce the
computational efficiency, requires the model of impact of this effect by introducing a compensat-
the controlled object to be represented as a neu- ing circuit into the system.
ral network as well.
Adaptive control schemes are often based on
the use of some reference model that specifies 4.3.2 Model Reference Adaptive Control
the desired behavior of the system under con- 4.3.2.1 General Scheme of Model Reference
sideration. The studies of which the results are Adaptive Control
described in this chapter are based on this vari-
In the case of MRAC problems, we implement
ant. Adaptive control systems that follow this
the controller as a neural network (neurocon-
approach perform the modification of parame-
troller) using ANN of the NARX type. Training
ters θ c (t) for the regulator according to the algo-
rithm implemented by the adaptation law. This of the neurocontroller is carried out in such a
modification is based directly on the tracking er- way that the output of the controlled system
ror value ε(t) = y m (t) − y(t), where y m (t) is the follows the reference model output as close as
output of the reference model and y(t) is output possible. A neural network model of the object
of the plant (controlled object). is required to implement the learning process of
Hands-on experience in the application of the neurocontroller.
the abovementioned adaptive control schemes Neural network implementation of the MRAC
demonstrates that the choice of reference model scheme (Fig. 4.6) involves two neural network
parameters has a fundamental influence on the modules: the controller network (neuro-
nature of the results obtained. An incorrect controller) and the plant model (ANN model).
choice of these parameters can lead to the con- As the first step, we solve the identification
trol system becoming inoperative. At the same problem for the controlled object, and then we
time, if the reference model parameters are se- use the obtained ANN model to train the neu-
lected adequately, it is possible to get a control rocontroller, which should provide for the most
system that solves the tasks assigned to it well. accurate possible tracking of the reference model
We present the results of the analysis of the in- output.
fluence of the reference model parameters on the The neurocontroller is a two-layer network
efficiency of the synthesized control system in fed with the reference input signal r(t), the con-
Section 4.3.2.4. trolled object output y p (t),and,insomecases,
The studied adaptive control schemes are es- the neurocontroller output u(t) at previous time
sentially based on the use of the ANN model steps (this connection is not shown in the di-
of the controlled object as a source of informa- agram) through the time delay lines (TDL ele-
tion about the behavior of this object. Since, due ments).