Spectrum sensing using feature vectors

Signal detection techniques such as eigenvalue based detection (EBD), enable a cognitive radio to sense the presence of a Licensed User's (LU) signal without prior knowledge of the signal characteristics, channel and noise power. It has been shown very recently that in addition to eigenvalues, leading eigenvectors or features, can also be used for signal detection. Recently, a feature template matching (FTM) algorithm was proposed in this regard, that used a feature learning algorithm to learn features from real valued signals and then compared the learned feature with the subsequently extracted features to determine the presence of a signal. However feature learning requires prior knowledge of the LU signal feature. This paper presents a Multiple Feature Matching (MFM) algorithm that obviates the requirement for feature learning thus enabling blind signal detection. The MFM algorithm extract features from the I and Q components of the complex signals received at multiple RF front-ends and across multiple sensing instants. This ensures maximum exploitation of signal correlation across spatial and temporal domains. The highest similarity index amongst all the extracted features serves as the test statistic for signal detection. A correlated reception system comprising of software defined radios (SDRs) is used to test and compare the FTM and MFM algorithms using over-the-air wireless signals. The results show that some of the configurations of the MFM algorithm perform better than the FTM algorithm while also proving to be more robust.