DEVELOPMENT OF A SPECTRUM OCCUPANCY PREDICTION MODEL FOR COGNITIVE RADIO SYSTEMS

Abstract

The two major users of a Cognitive Radio (CR) system are the Primary User (PU) and the Secondary User (SU). A secondary spectrum user cannot transmit in a channel before sensing and knowing the spectrum occupancy state as this may cause interference. This poses a major challenge because these operations ought to be performed in each time slot and thereby causing a substantial delay before the user gains access to the spectrum, leading to inefficient utilisation. Therefore, a channel predictive system will mitigate this problem. In this work, a machine learning model for spectrum occupancy prediction was developed. Power Spectrum Density (PSD) data were collected for 24 hours in Minna, Niger State and FCT Abuja both in Nigeria with 3 measurement sites per location within the VHF band (30-300 MHz). Exploratory Data Analysis (EDA) using power density plots was used to reduce the dimensionality of the dataset so that the data can be fit for machine learning. The power density plots reveal 12 distinct groupings or frequency sub-bands for the entire dataset. A Back-Propagation Neural Network (BPNN) model was developed to predict the spectrum occupancy using time-series data which was converted into a feature vector that was captured as time instances of the occupancy of all the frequency sub-bands. This serves as the input vector into the feedforward neural network. Twenty-four different input parameters, which capture hourly spectrum occupancy, were used with only one output that predicts the spectral occupancy. Comparison of the prediction results with the actual results obtained was done. The weight of the neural network initially generated randomly was improved using the Auto-Regressive (AR) model whose order is based on the time dimension of the feature vector. The coefficients of the AR model were obtained from the synaptic weights and adaptive coefficients of the nonlinear sigmoid activation function of one hidden layer with a ten-neuron Real-Valued Neural Network (RVNN). A linear activation function was used in the output layer. To obtain the AR coefficients, the training data and the corresponding expected occupancy (estimated from the raw data) are passed to the neural network alongside the number of neurons in the hidden layer. Based on the training data and the corresponding output data, the neural network model trains itself to come up with the best weights which can generally be used by the AR model for unseen data. After computing the weights, the performance is first tested on the entire training data, on the validation dataset, and the test dataset. Overall, the results for Minna dataset, in band 1 (30-47 MHz), reveal a highest actual spectral occupancy of 40.59% with a prediction accuracy of 99.06% while band 7 (137.05-144 MHz) has a lowest occupancy of 25.24% and a prediction accuracy of 99.31%. The corresponding results for the Abuja dataset, in band 1 (30-47 MHz), show a highest actual spectral occupancy of 39.11% with a prediction accuracy of 98.59% while band 11 (230.05-267 MHz) has a lowest occupancy of 22.13% in with a prediction accuracy of 99.40% were obtained for Abuja dataset. Clearly, band 1 had the highest spectral occupancy values in both locations and therefore should be avoided for Cognitive Radio (CR) deployment. The performance of the Neural Network prediction model reveals accuracy of 91.51% to an unseen test dataset, an accuracy of 99.02% on the training dataset and an accuracy 91.63% to the validation dataset.