A Gaussian mixture model method for eigenvalue-based spectrum sensing with uncalibrated multiple antennas

Abstract

In this article, a novel spectrum sensing technique for cognitive radio (CR) is proposed where eigenvalues of the covariance matrix of the received signal are used as features for detection with multiple uncalibrated antennas. In the proposed scheme, training decision vectors composed of eigenvalue components are approximated as Gaussian mixture model (GMM) and underlying distribution parameters are extracted using expectation-maximization (EM) algorithm. Using the obtained parameters, posterior probability of subsequent decision vectors are computed and the channel is classified as either occupied or unoccupied. This is different form existing spectrum sensing techniques where elements of the covariance matrix or its eigenvalues are reduced to a single decision statistic resulting in loss of useful discriminatory information. Proposed technique overcomes this limitation by forming decision vectors from the eigenvalue features and performing GMM based classification in multidimensional space. Simulation results also reveal that proposed method outperforms state-of-the-art techniques for detection of primary user (PU) signal using uncalibrated antennas.