Spectrum Sensing Based on Blindly Learned Signal Feature (1102.2840v2)

Abstract: Spectrum sensing is the major challenge in the cognitive radio (CR). We propose to learn local feature and use it as the prior knowledge to improve the detection performance. We define the local feature as the leading eigenvector derived from the received signal samples. A feature learning algorithm (FLA) is proposed to learn the feature blindly. Then, with local feature as the prior knowledge, we propose the feature template matching algorithm (FTM) for spectrum sensing. We use the discrete Karhunen--Lo{`e}ve transform (DKLT) to show that such a feature is robust against noise and has maximum effective signal-to-noise ratio (SNR). Captured real-world data shows that the learned feature is very stable over time. It is almost unchanged in 25 seconds. Then, we test the detection performance of the FTM in very low SNR. Simulation results show that the FTM is about 2 dB better than the blind algorithms, and the FTM does not have the noise uncertainty problem.