DOA Estimation Based on CNNs Embedded With Mamba

The convolutional neural networks (CNN) have been proved to be more efficient in Direction of Arrival (DOA) estimation of underwater acoustic array signals. Most of CNNs use the covariance matrix of the array signal or the spectrum as features to estimate the angle of incidence directly or to remove the noise. However, feature processing of the original signal inevitably loses part of the information, and using the original signal directly makes the number of parameters and computation huge. We propose a DOA estimation method that does not require feature processing. Mamba’s ability to handle very long sequences is utilized instead of feature extraction. Upon completion of the modeling and simulation processes, the model that has been put forward demonstrates a superior level of precision across a spectrum of scenarios characterized by differing snapshot counts and signal-to-noise ratios when juxtaposed with the CNN. Notably, in conditions where the signal-to-noise ratio is diminished, the proposed model exhibits a generalization accuracy that surpasses that of CNN by a margin of 3%. The approach delineated in this research paper is capable of reliably discerning the direction of arrival of multiple signals whose characteristics are not previously known.