Enhanced Coherent DOA Estimation in Low SNR Environments Through Contrastive Learning

Conventional methods for coherent direction-of-arrival (DOA) estimation often encounter considerable errors in low signal-to-noise ratio (SNR) environments. Meanwhile, deep-learning (DL) approaches perform well but typically assume known signal or noise power levels for normalization - a premise not always practical in real scenarios. This study introduces a novel contrastive-learning approach to enhance the performance of the DL method for coherent DOA estimation in a low SNR environment without the assumption of a known signal or noise power scale. The methodology includes the contrastive-learning optimization objective and the two-step training strategy for coherent DOA estimation. The proposed optimization objective has been proved to significantly increase the mutual information lower bound of neural networks in a self-supervised manner without the need for labels. Simulations and experiments verify that our method substantially reduces estimation errors in low SNR and coherent environments.