IEEE - Aerospace and Electronic Systems - April 2023 - 29

Image licensed by Ingram Publishing
missiles based on traditional proportional navigation (PN). In
Lechevin and Rabbath [14], proportional-derivative navigation
guidance regulations for the terminal phase are proposed.
A full design for the autopilot and guidance devices
for a class of spin-stabilized and fin-controlled projectiles is
shown in Theodoulis et al. [21].
When GNSS signals are available, costly inertial navigation
systems can be substituted with less precise technologies
in order to save money while still retaining a
satisfactory degree ofprecision. This would make it possible
to update the inertial system at a cheap cost, limiting the
increase of faults. In addition, combining the signals of
numerous low-cost sensors to increase overall accuracy can
be a cost-effective and collateral-damage-free technique.
In the last decades, elementary inertial sensors have been
used to provide a rough position estimate. GNSS-aided inertial
navigation systems have been widely studied to improve
the accuracy ofthe inertial navigation system, which is used
as the primary navigation source [23]. Several studies have
also been conducted to improve the performance of the
GNSS positioning system [25]. These studies have been conducted
for a wide range ofapplications, including high-precision
positioning and autonomous navigation.
Integrated data fusion's advantages have been proved in
Bryne et al. [1], where it is described a collection ofnonlinear
observers in addition to INS/GNSS hybridization. Additional
sensors may be used to contribute to a filter, such as the Kalman
filter, if they are available [8]. The Kalman filter, however,
is computationally intensive and requires costly
integration to effectively function. This has led to the development
ofdifferent filter types that use simpler integration without
sacrificing too much accuracy, including the unscented
Kalman filter [11] and other similar techniques [2].The
unscented Kalman filter is a nonlinear, recursive algorithm
used to estimate the location and orientation of a moving
object in the presence of external disturbances. Unlike the
Kalman filter, the unscented Kalman filter uses unscented
transforms to find an optimal estimate ofthe current state estimate.
To put this simply, this means that the features in the
current state estimate are not perturbed, unlike the case with
the Kalman filter. This makes the unscented Kalman filter
much easier to integrate. The filter could be easily implemented
in many other air and space autonomous vehicles [17].
APRIL 2023
Going further with this fact, the use of artificial intelligence,
machine learning (ML), and automatic learning techniques, in
substitution of these previously mentioned classic filtering
algorithms, with the aim of reducing the computational cost
on board the aircraft, and therefore the weight and cost, show
a promising present and future [3], [7], [9].
Even with GNSS/IMU integrated systems, however,
many scenarios have significant uncertainty, unknown disturbances,
and anomalous readings, which can be particularly
noticeable during terminal guiding for low-cost
devices. As a result, finding new robust algorithms that
can reach the requisite accuracy levels at a cheap cost during
guiding is an important part of projectile research,
especially if GNSS signals are unavailable. For example,
modern laser guided ballistic rockets combine IMU, GPS,
and laser guidance capabilities to provide high accuracy
and all-weather assault capability [6].
New techniques and algorithms are now needed to
provide effective and resilient systems with a high level of
autonomy and accuracy at a reasonable cost. ML, for
example, is a promising approach in this respect. It offers
a wealth of options and ground-breaking solutions that are
especially intriguing for GNC applications, where its presence
is currently limited but promising. Moreover, using
ML methodologies to estimate the parameters, based on
aerospace vehicles' flight dynamics has the benefit that
when the algorithm is calibrated or trained, it is the algorithm
that returns the information, independently of physical-mathematical
foundations that govern dynamics.
These outputs can later be used within the GNC algorithm
for the input data [7], [9]. The applicability of these tactics,
however, is highly dependent on the representativity
and amount of real input and output data used in training.
This means that the training mission envelope is the only
place where desirable performance stability and convergence
can be found. Other control algorithms might be
used for this sort of application to assure convergence and
stability parameters under the given unpredictable situations.
For example, Hardier et al. [10] presented adaptive
control that employs adaptation rules to predict unknown
system parameter changes for multiple mission envelopes
live, and Wang et al. [22] presented a method for the
ascent trajectory based on feed-forward neural networks
IEEE A&E SYSTEMS MAGAZINE
29
IEEE - Aerospace and Electronic Systems - April 2023

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