IEEE - Aerospace and Electronic Systems - April 2023 - 38

Neural Network-Based Controller for Terminal Guidance Applied in Short-Range Rockets
Table 6.
Monte Carlo Simulation Parameters
Parameter
Mean
Standard deviation
Parameter
Mean
Standard deviation
Parameter
Mean
Standard deviation
Parameter
Mean
Standard deviation
Parameter
Mean
Standard deviation
Parameter
Mean
Standard deviation
Initial f
Initial pitch
Nominal
20
CD0ðMÞ
Nominal
10% of nominal
CMa ðMÞ
Nominal
10% of nominal
Wind direction
20
CLa3 ðMÞ
Nominal
5% of nominal
CNdðMÞ
Nominal
10% of nominal
conditions, thrust characteristics, and aerodynamic coefficients.
Uncertainty in coefficients has been always taken into
consideration. Aerodynamics is often simulated using analytical
and modeling methods with little experimental, and accuracy
may differ from reality. Table 6 contains the specifics of
the uncertainty model.Note that, the sensor uncertainty model
is defined in the " Sensors " section. Ballistic shots, traditionally
controlled shots, and NN controlled shots are all subjected
to a set of 10,000 shots. After then, a total of 30,000
simulations are run. This simulation campaign is not the same
as the one used for NN in order to prove the applicability of
the proposed algorithms.
DISCUSSION
Results for the ballistic trajectories, NN controller for the nine
cases previously defined, and conventional controller
approaches are shown in Table 7. The CEP is the circle that
can accommodate halfof the impact sites, and it will be used
as a quality check parameter at the end ofthe simulation since
it is a valid reference for any method used. The better the
global GNC device is, the smaller the CEP is.
As it can be seen in this table, the results for BR training
algorithm are far better than results for the other training
38
0.01
CLa ðMÞ
Nominal
10% nominal
CMq ðMÞ
Nominal
10% of nominal
Initial azimuth
Nominal
0.01
CmfðMÞ
Nominal
10% of nominal
CspinðMÞ
Nominal
10% of nominal
Wind speed
10 m/s
5 m/s
CDa2 ðMÞ
Nominal
5% of nominal
CMa3 ðMÞ
Nominal
5% of nominal
Thrust
T(t)
10 N
CNqðMÞ
Nominal
10% of nominal
CmmðMÞ
Nominal
10% of nominal
algorithms for the three strategies. Inside BR training algorithm,
strategy one results the most accurate, which is positive,
as the complexity of the NN is lower. This fact
conducts us to conclude that for the currently studied application,
the BR training method with Strategy 1, provides the
the most accurate results. We can also conclude that SCG
training method is not suitable for this application, and LM
may be usable but its results are not as good as BRones.
Following, we present in Figures 9 and 10, a visual
comparison between the best, BR and Strategy 1, and the
best with LM training algorithm (Strategy 2) obtained
CEP, according to Table 7. Note that, none of the SCG
training algorithm strategies are represented, as their
results are similar to ballistic noncontrolled flights and the
SCG training method is concluded to be not applicable for
this kind of vehicle.
The left column of subfigures, in each figure, compares
ballistic flights against NN controller-assisted flights for each
of the three proposed shots (750, 850, and 950 m, respectively).
The second one compares NN controller results
against conventional modern theory controller results.
Controlled flights, on average, have tighter impact
groupings, with identical consequences for BR Strategy
1-based and conventional controllers, featuring lower performance
for LM Strategy 2-based controllers. Improvements or
IEEE A&E SYSTEMS MAGAZINE
APRIL 2023
IEEE - Aerospace and Electronic Systems - April 2023

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - April 2023
