Fast Direction-Of-Arrival Estimation Of Multiple Targets Using Deep Learning And Sparse Arrays

In this work, we focus on improving the Direction-of-Arrival (DoA) estimation of multiple targets/sources from a small number of snapshots. Estimation via the sample covariance matrix is known to perform poorly, since the true manifold structure is not revealed for a small number of samples. First, we explicitly model the sample covariance matrix that is used for the DoA estimation as a noisy version of the true one. Next, we employ a stacked denoising autoencoder (DAE) that predicts a statistically "richer" version of the sampled matrix that is subsequently used for the DoA estimation. Moreover, we consider a limited number of sensors (comparable to the number of sources) in a non-uniform linear configuration and introduce an end-to-end hybrid DoA prediction-estimation scheme. Results demonstrate significant improvement compared to the conventional approach.