Learning Based Spectrum Sensing in OFDM Cognitive Radios

Abstract

Cognitive Radio (CR) is an innovative technology introduced to efficiently utilize the spectrum. It allows secondary users to access the licensed portion of the spectrum when it is not occupied by the primary licensed user. Most of the CR applications are expected to operate in channels occupied by OFDM based systems since OFDM is the preferred modulation scheme of most recent wireless technologies. To be able to efficiently utilize the spectrum, CR must be able to properly sense the spectrum. This thesis models the spectrum sensing problem in a Cooperative CR system as a two class pattern recognition problem: signal present or signal absent. The signals from both classes have different characteristics which are learned by a linear classifier during the training phase. Once fully trained, the classifier utilizes this learning to classify any unseen data into one of the classes. The characteristics which differentiate the signals from both classes are called features and are acquired by the linear classifier through the process of feature extraction. In a cooperative CR network, each CR extracts features from its received signal and sends it to a fusion center. At the fusion center, a universal decision is made on spectrum occupancy based on features received from all the CRs. In this thesis, energy, correlation and entropy are used as features to distinguish between the primary OFDM signal and noise. The performance of the spectrum sensing schemes is evaluated in terms of the detection and false alarm probabilities. It is shown that energy and correlation detectors outperform the entropy detector in AWGN channels. However, in a fading channel, the correlation detector outperforms both the energy and entropy detectors due to the degradation of their performance caused by the deep fades in the channel. The performance can be improved by increasing the observation window size and by changing the number of users in the Cooperative CR network.