Abstract:

An direction finding framework based on unsupervised deep neural networks is proposed for the uniform circular array super-resolution direction finding problem. First, a two-dimensional direction of arrival （DOA） estimation problem is transformed into sparse power spectrum reconstruction, where a deep neural network establishes an end-to-end mapping from spatial covariance vectors to sparse spectrum. Furthermore, a loss function with sparse feature constraints is constructed to learn two-dimensional spatial structural features of sparse spectrum in an unsupervised manner. Upon the completion of network training, accurate two-dimensional angle estimates are obtained through clustering post-processing of network outputs. Simulation results demonstrate that the proposed method achieves accurate angle estimation in two dimensions, excellent generalization and robustness in unknown scenarios, achieving high estimation accuracy in low signal-to-noise ratio and limited snapshots.

Key words: uniform circular array, super-resolution two-dimensional direction finding, deep neural network, unsupervised learning