Direction of Arrival Estimation Using One-dimensional Convolutional Neural Network and Gated Recurrent Unit

This paper introduces a deep learning (DL) framework to address direction of arrival (DOA) estimation problem. Traditional signal processing methods such as multiple signal classification (MUSIC) highly rely on signal model and array geometry. However, DL methods, being data-driven, make analytical process of signal or array less important. In this paper, a neural network architecture combining one-dimensional convolutional neural network (1D CNN) and gated recurrent unit (GRU) is proposed to estimate DOA of multiple signals. The multi-signal DOA estimation is treated as a multi-class multi-label classification issue. First a dataset using the covariance matrix of target signals received by a circular antenna array is generated. The proposed 1D CNN-GRU model then learns the relationship between covariance matrix elements and DOAs through training. Experimental results show that our proposed method has higher accuracy than MUSIC and is able to deal with multi-path DOA estimation. Besides, 1D CNN-GRU is proved to have lower root mean squared error (RMSE) than other DL methods, because features over small local areas and time-sequence are both learnt by 1D CNN layers and GRU layers. In addition, 1D CNN-GRU exhibits effectiveness in experiments using real-world data.