PREDICTION OF SPECTRUM OCCUPANCY USING MACHINE LEARNING ALGORITHMS: A CASE STUDY OF MINNA

Abstract

There is an alarming growth rate in spectrum usage, where some of the allocated spectra are fully engaged while others are sparsely utilized. The cognitive radio allows the primary users to use the available spectrum holes alongside the secondary users. The challenge of using cognitive radio technology is in the interference, which is a factor that causes a delay in the handoff time. This research developed a system that makes the cognitive radio operation more effective with little or no interference. Dataset were collected by scanning the spectrum between the frequency range of 80 MHz and 1 GHz using the Agilent N9342C Spectrum Analyzer (SA), which was connected to a personal computer and an antennae with a range of 47 MHZ to 1 GHz attached to the SA. The spectrum sensing exercise was carried out at Morris Fertilizer in Minna, Niger state, between 7:00 am-10:00 am (three hours). The method used in the sensing of the spectrum is Energy Detection. The dataset collected from the exercise was used to train and test different Machine Learning (ML) algorithms at a ratio of 7:3. The ML algorithms were used to predict the availability of the spectrum holes, that is, the frequency within the spectrum occupied or not occupied. The logistic Regression, Random Forest, Decision Tree, XGBoost and the K-Nearest Neighbour has training accuracy result of 94.84%, 99.93%, 99.93%, 99.86% and 98.19%, respectively and test accuracy result of 90.43%, 99.52%, 99.52%, 99.52%, and 97.61%, respectively. The test accuracy, precision, recall and F1-score are 90.43%, 90.40%, 93.39% and 91.43%, respectively was obtained with the application of logistic regression. Random forest results of accuracy, precision, recall and F1- score are 99.52%, 99.98%, 99.17% and 99.57%, respectively. For the Decision Tree, the test accuracy, precision, recall and f1- score are 99.52%, 99.99%, 99.17%, and 99.58%, respectively. The test accuracy, precision, recall and F1- score are 99.52%, 100.00%, 99.17% and 99.58%, respectively was obtained with the application of the XGBoost. Also, the test accuracy, precision, recall and f1-score are 97.61%, 100.00%, 95.87% and 97.89% respectively was obtained with the application of the KNN. From the result obtained, the XGBoost has the highest level of prediction accuracy. These results demonstrated the effectiveness of XGBoost when compared to other popular ML algorithms for spectrum occupancy prediction