Evaluation of a new spectrum sensing technique for Internet of Things: An AI approach

In the last decade we observed a great demand for wireless sensor applications as the connectivity of objects related to the Internet of Things concept increased. The growing number of wireless sensors leads to more spectrum demand and eventually to collisions due to overcrowding, causing a decrease in their performance level. Thus, to avoid collisions, detailed knowledge of the radio spectrum is required such as the degree of spectrum occupancy and the radio modulations used. This paper presents an analysis of the impact of different radio signal representations (I/Q coordinates, polar coordinates, and Fast Fourier Transform) on the performance level of machine learning algorithms in spectrum sensing classification. Our results shown that machine learning algorithms achieve a higher classification accuracy when the FFT representation of the radio signal is used, with a classification accuracy of 98.6%. When using the time series, the I/Q representation of the radio signal obtained an accuracy of 68.6% on the test dataset meanwhile the polar coordinates achieved an accuracy of 90%, respectively.