DLMUSIC: The Improved MUSIC Algorithm Based on Deep Learning

Direction of arrival (DOA) estimation is an important problem in both wireless communication and radar applications. This paper proposes a new algorithm combining deep Learning with multiple signal classification (MUSIC), called DLMUSIC. We use multiple convolutional neural networks (CNNs) to estimate the covariance matrix, and then take the eigenvalue decomposition of the covariance matrix instead of the empirical covariance matrix, meanwhile we use the spatial spectrum approximated by Gaussian function as the label of training data instead of MUSIC spectrum, so our algorithm can overcome many limitations of the MUSIC algorithm which include the inability to localize coherent sources and the poor performance at low signal-to-noise ratio (SNR). Simulation results demonstrate that the DLMUSIC algorithm perform much better than MUSIC algorithm at low SNR with a few snapshots.