Deep Learning Based Parameter Estimation of Frequency Shift Keying LPI Radars

LPI radars aim to both perform their own functions and not be detected by their targets by transmitting at low power. For this purpose, they usually perform frequency modulation (linear incremental, triangular, frequency shift keying, etc.) within the signal. Due to the development of cognitive electronic warfare systems, it is essential to characterize the behavior of these radars automatically. We propose three different deep learning models to solve the less studied automatic parameter estimation problem compared to the LPI radar frequency modulation classification problem, which has been extensively researched in the literature. We designed these models as convolutional, recursive, and hybrid to predict frequency jump points and frequency values in the signal. In the numerical analysis, we evaluated the network performances in terms of correct detection, false alarm, and mean square error metrics. As a result, we saw that the hybrid model, which uses convolutional and recursive networks together, achieved the highest performance and we determined that the main contribution to this success came from the convolutional network.