Complex-Valued Neural Network for Estimating the Number of Sources in Radar Systems

This article proposes a method for estimating the number of sources using the time-domain received signal in frequency-modulated continuous-wave radar systems. Conventional approaches typically estimate the number of sources in the frequency domain via the fast Fourier transform (FFT) on the received signals. However, these conventional methods face challenges in distinguishing targets that have identical distances and velocities, resulting in an underestimation of the source count on the range–Doppler (RD) map derived by applying the FFT to the received signal. Consequently, this underestimation can lead to inaccurate results in subspace-based direction-of-arrival (DOA) estimation algorithms. In contrast, our method directly uses the received signal in the time domain, which can distinguish between targets that have identical distance and velocity. We introduce a complex-valued neural network that estimates the number of sources using signal vectors extracted from analog-to-digital converter (ADC) samples of a received signal matrix. The proposed method achieved an average accuracy of 91% in various simulations and experiments, and its effectiveness was verified through comparison with conventional source estimation algorithms. In addition, we can also obtain accurate results in the subspace-based DOA estimation algorithms using the number of sources estimated by the proposed method.