Page 129 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

better.
In addition, experimental tests have to be conducted in order
to evaluate the actual performance of the algorithm. In future
studies, the algorithm will be tested with a real signal received
on orbit.
By accurately estimating the potential success in decodifying
the received messages, a considerable amount of
computational resources, and therefore power consumption
is expected to be economized. Further studies should be
carried out in this field, in order to confirm the impact of the
use of these algorithms in a satellite’s lifetime.
Furthermore, international recommendations can be
developed including all the available information, simulations
and experimental data already obtained. Those standards
can be developed by organisms like the International
Figure 9 – SVM Classification results.
Telecommunication Union and can aid in the design and
Table 1 – kNN vs. SVM implementation of spatial ADS-B receivers.
Moreover, the International Civil Aviation Organisation
Method should regulate the mandatory use of ADS-B equipment and
kNN SVM
Indicator
also make mandatory the use of tamper proof devices.
P e 0.059 0.049
Classification Slow Very fast
5.2 Potential Impact
Training time No training time Time consuming
The previous study presents a different way to deal with the,
5. CONCLUSION
already known, problem of receiving ADS-B messages in
congested airspaces. Using machine learning and pattern
It is shown that, under the hypothesis stated, both methods
recognition methods is a novel analysis that can increase the
perform with little difference. It is clear that SVM is about 1%
amount of messages that a receiver could decode. This new
better than kNN. Under that circumstance, other indicators
technique can contribute to International Recommendations
have to be analyzed in order to define which method will be
and Standards to improve them, not only in a particular
better.
assumption, but also in the way that parameters are chosen.
One of the most important indicators is the time that takes to
If the addition of this method makes the system more
classify a new sample. In that case, SVM performs better.
efficient, the lifespan of the satellites will be improved due to
Nevertheless, if kNN is manipulated in order to obtain a single
reduction in energy consumption. Consuming less energy
and simpler boundary, it can be approximated by a function,
not only impacts on the battery depth of discharge, but
reducing computation complexity. But, in many cases that
also makes the satellite cheaper due to smaller electronic
method is not applicable, especially when the dimension of
parts. Therefore, using machine learning techniques could
the samples (i.e. the amount of features used) is increased.
potentially reduce the overall cost of satellite missions
Despite that the training time for SVM is important, this
carrying ADS-B receivers.
phase can be done offline, and once the system is trained, the
Making ADS-B a standard real-time global solution for civil
classification process itself is fast.
flight tracking enables safer flights and thus a potentially
Nonetheless, only one kernel was used to test SVM. The
increase the aircraft density.
results show that a simpler kernel can be used, improving the
performance of the method.
REFERENCES
Finally, after some fine tuning, the SVM method proof to be
the best choice for this problem. [1] R. Cochetti, “Low Earth Orbit (LEO) Mobile Satellite
Communications Systems,” Wiley Telecom, pp. 264,
5.1 Future Work 2015.
[2] ITU-R WP5B, “Reception of automatic dependent
It is planned to further improve the signal modeling. This
surveillance broadcast via satellite and compatibility
will be done by including phenomena such as doppler shift
studies with incumbent systems in the frequency
and phase shift. However, these additions to the model would
band 1 087.7-1 092.3 MHz,” Recommendation
only give more information and the performance would only
M.2413-0, International Telecommunication Union
increase. Thus, using the actual model, worst case in this
Radiocommunication Sector", 2017.
sense is taken into account. Also, different scenarios can be
modeled such as different aircraft densities and altitudes. [3] ITU BR Director, “Results of the first session of
Furthermore, different strategies of pattern recognition and the Conference Preparatory Meeting for WRC-19
feature extraction could be considered to be certain that there (CPM19-1),” International Telecommunication Union
is no other available method to this problem that performs Radiocommunication Sector", 2015.
– 113 –