Shift-Invariant Structure-Imposed Convolutional Neural Networks for Direction of Arrival Estimation

This paper frames the estimation of directions of arrival of plane waves impinging on an array of sensors as a classification problem using convolutional neural networks (CNNs). We propose a methodology to impose the shift-invariant structure inherent in data to CNNs. We use several methods to pre-process the data collected from sensor arrays and feed the pre-processed data as inputs to CNNs. For all CNNs, data sets corresponding to different signal-to-noise ratio (SNR) levels are generated. The data sets associated with the lowest SNR level are used for training while the other data sets are used for validation. Comparison of the accuracy of the shift-invariant structure-imposed CNNs with those of CNNs that are based on raw data, sample covariance matrices, and principal eigenvectors is provided. The simulations show that shift-invariant structure can be efficiently and most accurately imposed using the optimal signal subspace basis estimates as CNN inputs.