Detection of Frequency-Hopping Signals With Deep Learning
Kyung-Gyu Lee, Seong‐Jun Oh
Abstract
Detection of the frequency-hopping (FH) signal is challenging when the hopping rate is unknown. Conventional spectrogram-based schemes can detect FH signals, but its performance is limited by the time-frequency resolution trade-off and spectral leakage. To alleviate this issue, we design convolutional neural network (CNN) and hybrid CNN/recurrent neural network (RNN)-based schemes. The CNN-based scheme alleviates spectral leakage by using feature maps. The hybrid CNN/RNN-based scheme mitigates the time-frequency resolution trade-off by using feature maps extracted from spectrograms of various window lengths. In simulations, the hybrid CNN/RNN-based scheme is shown to outperform the CNN-based and conventional detection schemes.