LFM based Wideband DOA Estimation using Deep Neural Network at Low SNR

Abstract

This work focuses on deep learning-based wideband direction-of-arrival (DoA) estimation for a wideband in particular LFM in case of extreme noise. We propose a convolutional neural network (CNN) that utilizes the correlation matrix to estimate and trained using multi-channel data in low SNR conditions. By using a systematic approach and treating the problem as a way to identify multiple possible DoAs, the CNN is trained to predict DoAs under different SNR conditions. This allows the CNN to accurately estimate the directions from which signals are coming, regardless of the level of noise in the environment. The architecture proposed exhibits robustness to noise, works effectively with a small number of snapshots, and achieves high resolution in angle estimation. Experimental findings demonstrate notable enhancements in performance under low SNR conditions when compared to existing methods, without the need for parameter tuning for correlated and uncorrelated sources. The enhanced robustness of our solution has broad applications in various fields, including wireless array sensors, acoustic microphones, and sonars.