Direction-of-Arrival Estimation for Time-Varying Arrays Using FCN-Based Deep Learning

This paper presents a novel deep learning framework designed to tackle the challenging direction-of-arrival (DOA) estimation problem for time-varying arrays. Measurements from various array geometries are formulated using distinct and spatially overcomplete observation systems. Subsequently, a mapping relationship is established between the proxy spectrum and the shared spatial spectrum for deep learning. We introduce a fully convolutional network (FCN) that learns this mapping from a large training dataset. The problem is modeled as a multi-label classification task and the FCN is trained to predict DOAs. Numerical simulations show that the proposed architecture can better resolve closely spaced directions in low signal-to-noise ratio (SNR) compared to subspace-based and sparse reconstruction-based methods, the proposed FCN-based DOA estimator also exhibits high estimation accuracy over the entire spatial domain, and its ability to accurately identify the number of sources is significantly improved with the introduction of another training approach for cases where the number of sources is unknown.