Cognitive Radio Spectrum Classification using FLA-SVM

Spectrum sensing is an important component of cognitive radio. Spectrum sensing involves classification of a part of spectrum or a frequency band as either “occupied” or “unoccupied”. A secondary user is permitted to transmit in this frequency band only if it is “unoccupied”. Conventional method for spectrum sensing involves checking the energy of received signal against a threshold. Such a method for classification of spectrum sensing may not be efficient when the decision region between “occupied” class and “unoccupied” class is nonlinear. In this paper, we propose to implement such a nonlinear classifier using support vector machine (SVM). Since, cognitive radio measurements involve large dataset, application of SVM is difficult for spectrum sensing. To overcome this difficulty, we apply a new fast learning algorithm (FLA-SVM) proposed in the literature to this problem. Application of FLA-SVM results in sample points reduced to $1/4^{th} \sim 1/5^{th}$, even to $1/10^{th}$ of initial training samples. Using these final samples, the training time gets reduced considerably and training speed increases to a remarkable extent. The most significant aspect is that the accuracy of classification can be kept similar as when a large set of training samples is applied to train the SVM. The simulation result shows the FLA to be extremely effective for spectrum sensing.