Page 169 - Neural Network Modeling and Identification of Dynamical Systems

4.3 APPLICATION OF ANN MODELS TO ADAPTIVE CONTROL PROBLEMS UNDER UNCERTAINTY CONDITIONS 159
TABLE 4.1 Table of correspondence of flight regimes Thus, separate test signals in the form of a de-
used in the synthesis and testing of the neurocontroller flection of control surfaces (elevons in the case
(NC) for the Hypersonic Research Vehicle. of the hypersonic research vehicle) in the consid-
Flight regime, Flight regime, ered approach are not required. This fact, how-
for which the NC for which the NC ever, does not eliminate the need for disturb-
was tested was synthesized
ing external influences on the controlled object,
M = 5, H = 32 km M = 7, H = 30 km
through which we reveal the information on the
M = 5, H = 30 km M = 7, H = 28 km
reactions of a given object to control actions. This
M = 6, H = 30 km M = 5, H = 28 km
M = 7, H = 30 km M = 5, H = 32 km information is the basis for the adaptive adjust-
M = 7, H = 28 km M = 6, H = 32 km ment of the control laws used. We used the ref-
M = 7, H = 32 km M = 5, H = 28 km erence signal for the angle of attack as one of
M = 6, H = 32 km M = 7, H = 28 km the types of such influences. As shown above,
M = 6, H = 28 km M = 7, H = 32 km the primary requirement for this signal is that
M = 5, H = 28 km M = 7, H = 30 km it has to provide the complete coverage of the
state space of the system under study. Besides,
it is necessary to take into account the dynam-
ond approach is analyzed in experiments, the ics of the system by varying the rate of transi-
results of which are given below. tions between individual states. A large number
When considering the first of the two ana- of computational experiments demonstrated the
lyzed options, it is actually required to solve efficiency of this approach.
for the controlled object the task of real-time In all computational experiments, the results
identification in order to have reliable informa- ofwhichwegiveinthe nexttwosections for
tion about the nature of the influence of control the MRAC and MPC schemes, the control sys-
actions on the behavior of the controlled ob- tem operation was set in the same way. The
ject. first 20 sec some disturbing reference signal was
For the convergence of the real-time identifi- fed to the input of the control system, which
cation process, the presence of a test (nontarget) is necessary for solving the identification prob-
signal at the system input is required for some lem. For the next 20 seconds, we are testing
time (in such a case this is the system adaptation the system. For this testing, we use some se-
time). This signal can be some additional sig- quence of stepwise effects, which are separated
nal to the actuator to configure the ANN model. in time from one another, so that the transi-
However, for nonlinear systems, the condition tion process from one such disturbance can com-
of proper tuning is also the consideration of as plete before the moment of submitting the next
many states as possible from the region in the one.
state space in which the system operates. There-
fore, the test signal for the ANN model is intro- MODEL REFERENCE ADAPTIVE CONTROL
duced not by the control surface deflection, but In this subsection, in Figs. A.61–A.69 the data
by the reference signal determined precisely in of the computational experiments on the estima-
the state space and passed through the control tion of the influence of the source data accuracy
system. In the neurocontroller, the input signal on the control characteristics for the case of the
is directly the reference signal, that is, to set the MRAC system are given.
controller in the case under consideration, the As noted above, the results of the compu-
test signal is input as a varying reference signal tational experiments shown in Figs. A.61–A.69
concerning the angle of attack. demonstrate the operation of the control system