Performance analysis of ML models on 5G sub-6 GHz bands: An experimental study

Abstract

Massive Machine Type Communication (mMTC) uses sub-6 GHz bands frequency bands for their applications in the 5G network. The exponential growth of mMTC wireless networks has made these bands overcrowded. Due to the increase in wireless traffic, spectrum scarcity is a significant constraint at sub-6 GHz bands for the 5G and beyond networks. Cognitive radio technology uses spectrum sensing (SS) techniques to access the spectrum opportunistically to resolve this issue where signal processing techniques (SPTs) are considered to design SS. However, the adaptiveness of SPTs is not feasible in the real-time environment due to the random spectrum access behaviour of the primary user and the fading environment. To minimize this issue, machine learning (ML) models are adapted. Different ML models are examined, and their performance is analyzed to find a better accuracy model in spectrum hole identification at 5G sub-6 GHz bands. The dataset of large-scale frequency samples is built from a universal software radio peripheral (USRP—2953R) at sub-6 GHz bands over \(\eta - \mu\) fading environmental conditions. A highly imbalanced dataset issue is reduced and compared with varying resampling techniques, and random oversampling is best to resolve the anomalies in the dataset. Random forest, naive Bayes, logistic regression, K-nearest neighbor, and decision trees are primary classifiers to train and detect the spectrum holes at 5G sub-6 GHz bands. Random forest outperforms the remaining ML models in spectrum hole identification in terms of probability of detection and accuracy.