Robust Underwater Acoustic Pulse DOA Estimation Based on Saliency-Aware ConvLSTM

This paper presents a compact and efficient underwater locating device (ULD) localization system design tailored for deployment on compact underwater vehicles. The proposed system features a three-element hydrophone array and a low-power acquisition module, capable of operating in deep-sea environments with limited energy and space budgets. To address the challenge of direction-of-arrival (DOA) estimation under low signal-to-noise ratio (SNR) conditions, we introduce a deep learning framework that combines saliency prediction with a convolutional LSTM (ConvLSTM) architecture. Phase difference features extracted from the array signals are enhanced by a predicted saliency map, enabling the network to suppress background noise and focus on weak, transient ULD pulses. A total of 10,000 synthetic samples were generated using a realistic simulation model, and the system was further validated in lake trials using a commercial ULD. Experimental results demonstrate that our method achieves superior accuracy and robustness compared to traditional cross-correlation-based approaches, with an average DOA error of 4.97° under field testing conditions.