Spectrum Sensing and Prediction for 5G Radio

Abstract

In future wireless networks, it is crucial to find a way to precisely evaluate the degree of spectrum occupation and the exact parameters of free spectrum band at a given moment. This approach enables a secondary user (SU) to dynamically access the spectrum without interfering primary user's (PU) transmission. The known methods of signal detection or spectrum sensing (SS) enable making decision on spectrum occupancy by SU. The machine learning (ML), especially deep learning (DL) algorithms have already proved their ability to improve classic SS methods. However, SS can be insufficient to use the free spectrum efficiently. As an answer to this issue, the prediction of future spectrum state has been introduced. In this paper, three DL algorithms, namely NN, RNN and CNN have been proposed to accurately predict the 5G spectrum occupation in the time and frequency domain with the accuracy of a single resource block (RB). The results have been obtained for two different datasets: the 5G downlink signal with representation of daily traffic fluctuations and the sensor-network uplink signal characteristic for IoT. The obtained results prove DL algorithms usefulness for spectrum occupancy prediction and show significant improvement in detection and prediction for both low signal-to-noise ratio (SNR) and for high SNR compared with reference detection/prediction method discussed in the paper. Keywords: Spectrum sensing • Spectrum prediction • Machine learning • 5G • LTE • Convolutional neural network • Recurrent neural network • Neural network • Deep learning This work was supported by the DAINA project no. 2017/27/L/ST7/03166 "Cognitive Engine for Radio environmenT Awareness In Networks of the future" (CERTAIN) funded by the National Science Centre, Poland.

Federated Learning for 5G Radio Spectrum Sensing

“…The RNN algorithm has been used in [ 10 , 11 ], while [ 12 ] focused on support vector regression (SVR)-based algorithm for prediction of PU’s next spectrum state. The frequency patterns combined with time dependencies have been discussed in [ 13 , 14 , 15 ]. These papers also consider using DL algorithms, namely RNN and convolutional neural network (CNN), which is a great choice for two-dimensional data.…”

“…As a shared ML model, we consider the application of CNNs, the goal of which is sensing and prediction of the spectrum occupation (or availability), i.e., the actual creation of the mentioned common model of spectrum occupation in time, frequency, and location. We choose CNN ML because our previous research [ 2 , 13 , 16 ] proved that CNNs are a better choice for faded 5G signal detection than RNN, k-nearest neighbors, and Decision Trees algorithms, and do not need large and complex data sets, as they can find hidden patterns within data. The final ML models are created by the central FL node by intelligent averaging weights of the CNN and then shared for the use of SU.…”