Radar Signal Modulation Recognition Based on Bispectrum Features and Deep learning

Signal bispectral transformation can not only suppress the influence of Gaussian white noise on signal modulation recognition, but also retain the signal amplitude and phase information. It is also used to extract the non-linear characteristics. Compared with other high-order spectra, bispectrum has a simple processing flow. However, the direct use of all bispectrum as signal features will lead to two-dimensional template matching, causing lots of calculations. Converting two-dimensional bispectrum into one-dimensional sequence, for example, extracting slice information of bispectrum, or using integral bispectrum apparently reduce the amount of data to be processed while retaining part of the bispectrum information. We input the extracted bispectral transformation of radar signals into the neural network to realize modulation recognition. The simulations validate our conclusions that our proposed methods still have a high recognition probability while SNR is low.