Direction of arrival estimation with convolutional neural networks and multiple signal classification

Recently, there has been a proliferation of applied machine learning (ML) research, including the use of convolutional neural networks (CNNs) for direction of arrival (DoA) estimation. With the large increase of research in this area, it is important to balance the performance and computational costs of CNNs with classical methods of DoA estimation such as Multiple Signal Classification (MUSIC). We outline the performance of both methods of DoA estimation for single source and two source cases and compare them to the Cramer-Rao lower bound (CRLB). For each source case, a CNN was trained for a perfect uniform line array (ULA), a perturbed ULA, and a ULA with missing sensors. The three cases are interesting studies as classical methods for DoA estimation such as MUSIC assume perfect array conditions, whereas the CNNs assume nothing about the structure of the data. The training data for each network was created by leveraging a signal model to create synthetic data at different SNRs. The results for each network are then compared to the results for MUSIC using the same signal case and array condition. The results indicate that for the single source case the CNN only performs significantly better than MUSIC for the perturbed array case. For the two-source case, the CNNs significantly outperform MUSIC for all ULA conditions, however, when compared to the CRLB it is shown that the CNN typically produces a biased estimate.