2D Spectral Analysis of OFDM Radar Data using Deep Learning

Spectral estimation is often encountered in a variety of signal processing tasks. One of the most common methods to estimate the spectrum of sampled data, is to use the Discrete Fourier Transform (DFT). However, using sampled data to estimate the spectrum with the DFT results in a limited accuracy and sidelobes. The limited accuracy is due to grid-oriented nature of periodogram and though this can be mitigated by interpolation, but it can lead to misinterpretation if the paths are too close to each other. These effects can cause issues in practical applications, e.g., sidelobes can mask weaker spectral components in the proximity of stronger components. To alleviate the problem of sidelobes, windowing techniques can be used on the cost of Signal-to-Noise ratio (SNR) loss. In this paper we present a novel approach which uses a Convolutional Neural Network (CNN) trained on synthetic data to estimate multidimensional spectra and predict the model-order of noisy input data. We compare our approach to conventional DFT methods using a harmonic retrieval data model.