A Deep Learning Architecture for Broadband DOA Estimation

Abstract

An efficient neural network-based approach for broadband direction of arrival (DOA) estimation is presented in this paper. The received data of the uniform circle array (UCA) is transformed into direction image, which is used as the input of the neural network. The phase component of the spatial covariance matrix of the received signal is extracted to form the direction image. We establish a convolutional neural network (CNN) with five hidden layers to learn the inverse mapping from the space of possible antenna element excitations to the space of possible angular directions to the signal source. DOA estimation is formulated as a regression problem, where the each DOA label to the direction image is consisted of the sine and cosine values of the angle of arrival. Simulation results show that the trained CNN network can be successfully used for broadband DOA estimation. The performance of the developed CNN model is comparable to the performance of the conventional algorithms at the lower signal-to-noise ratio. Importantly, the proposed CNN estimator further reduces the computation time which makes it successful to apply to real-time applications.