Abstract |
The abundance of wireless devices raises the issue of their coexistence in
spectrum. The current static allocation scheme of the available frequency
bands, is proven inadequate to satisfy the ever increasing demands for bandwidth
and higher data rates. The ISM bands are overcrowded, while on the
contrary, the majority of licensed spectrum regions are already allocated and
at the same time underutilized. Hence, the need of adoption a Dynamic
Spectrum Allocation (DSA) scheme is emerging. This new spectrum sharing
policy allows unlicensed users to exploit licensed frequency bands, with the
assumption that their operation does not harm the licensed transmissions.
The unlicensed users should be able to identify idle frequency bands or
avoid interference with licensed users. Therefore, users endowed with cognitive
capabilities should be able to monitor the spectrum and adapt their
behavior appropriately. Cognitive Radios (CRs), implemented on Software
Defined Radios (SDRs), arise as a promising solution to the aforementioned
requirement, through their ability to dynamically reconfigure the operation
parameters.
Furthermore, in the effort of efficiently detecting the transmission opportunities
and facilitate the competition, the unlicensed users pursue to recognize
the identity of transmitted signals. For instance, the knowledge about
the modulation used by a received signal, could potentially reveal whether it
belongs to a licensed user or not. This is the primary concern of Automatic
Modulation Classification (AMC) methods, where various characteristics are
examined in order to identify the modulation of a received signal, without
any a-priori knowledge.
Machine Learning is a useful tool, that could potentially enhance Cognitive
Radios and modulation recognition by applying artificial intelligence. In
the most widely investigated form, supervised classification learning, known
signals are trained to export a predictive model that is able to recognize transmissions
out of a predefined modulation scheme dictionary. In this work, we
aspire to facilitate the adoption of Machine Learning algorithms in cognitive
systems and as an extension to achieve modulation identification.
To our knowledge, there is a lack of real-world systems that automatically
monitor a spectrum region to detect, extract and classify existing transmissions.
In our online approach, we apply efficient signal analysis techniques
to online detect and separate signals out of the observed frequency band on
the fly and we make use of the Random Forest algorithm to train our predictive
model. Additionally, our implementation is designed based on the
GNU Radio platform, so as to be easily configurable and extendable to support
alternative Machine Learning algorithms. Finally, we elaborate on the
assumptions made during the development of our system and we evaluate the achieved performance, in terms of classification accuracy and decision delay
|