Deep Learning-Aided Orthogonal Representation Approach to DOA Estimation of Wideband Signals

Wideband source signals occur in many applications of array signal processing. With the use of time-delay orthogonal representations of the incident uncorrelated wideband signals, the array output vector herein is divided into three parts. The first part is of the rank-1 form as if the incident signals are from the narrowband sources. The spatial signature of each signal is characterized by a direction of arrival (DOA) and a second-order statistics-dependent vector, which can be viewed as the generalization of the steering vector. The second part contains the so-called virtual interferences, which are uncorrelated with the signals. The third part is the sensor noise term. A deep-learning-based scheme is further developed for suppressing the virtual interference term in the covariance matrix of the array output. This enables a regime of time-domain wideband DOA estimation, using straightforwardly the subspace projection technique already developed under the narrowband assumption, with neither subband decomposition nor frequency focusing. The choice of the orthogonal representation time-delay parameter involved in the subspace projection procedure is also discussed. Simulation results are included to demonstrate the efficacy of the proposed approach.