Reliable Modulation Classification at Low SNR Using Spectral Correlation

In this paper, we propose a novel signal classification method using cyclic spectral analysis and neural network for cognitive radio applications. In cogni- tive radio, it is desirable to have an accurate and reliable signal classification algorithm which can operate at low signal to noise ratio and without knowledge of the carrier frequency and bandwidth of the target signal. Cyclic spectral analysis has been proven to be a powerful tool for classifying signals. However, the amount of data introduced by spectral analysis is too large for any classifier to employ. Hence, a spectral analysis based feature extraction has to be performed to drastically reduce the data. Specifically, we propose to use both the α profile and the frequency profile of the Spectral Coherence Function (SOF) as the feature. Numerical results show significant performance improvement compared to those of using only the α profile feature.