On Spectrum Sensing, a Machine Learning Method for Cognitive Radio Systems

Abstract

Spectrum sensing plays an important role in enabling cognitive radio technology for the up-and-coming generation of wireless communication systems. Over the last decade, several sensing methods have been proposed, including energy detection, cyclostationary feature, and matched filter. However, these techniques present several limitations. Energy detection performs poorly under low signal-to-noise ratio, cyclostationary features are complex, and matched filter requires some prior knowledge about the primary user signal. In addition, all of these techniques require setting a threshold which needs the prior knowledge of the noise distribution. Thus, the reliability of spectrum sensing is still an open issue in wireless communication research. In this paper, we propose a spectrum sensing method based on a machine learning theory for cognitive radio networks. The spectrum sensing problem is rigorously modeled and out of which a large-scale comprehensive dataset is built. This dataset is then used to train, validate, and test several machine learning techniques, including random forest, support vector machine with different kernels, decision tree, Naïve Bayes, K-nearest neighbors, and logistic regression. The models were extensively tested and evaluated using metrics such as the probabilities of detection, false alarm, and miss-detection as well as the accuracy of the classification. The simulation results show that the random forest model outperforms all the other machine learning methods.