Convolutional Neural Network-Based Regression for Direction of Arrival Estimation

This work utilizes convolutional neural networks (CNNs) to estimate the directions of arrival of plane waves impinging on an array of sensors. We propose a methodology to impose the shift-invariant structure inherent in the data to CNNs. We use several input formulations to structure the data collected from sensor arrays and feed the structured data as inputs to CNNs. For all CNNs, data sets corresponding to different signal-to-noise ratios (SNR) are generated. Several different CNNs are trained using different pairs of training and validation SNRs to investigate how root mean square error (RMSE) trends. RMSEs of the shift-invariant structure-imposed CNNs are compared with CNNs that are based on raw data, sample covariance matrices, and principal eigenvectors. The simulations show that shift-invariant structure can be efficiently imposed and has lower RMSE than the other input formulations; however, additional refinement is required to improve performance beyond that of classical subspace based DOA estimation methods.