High-SNR phase unwrap method based on RMN-LSTM for distributed hydro-acoustic detection

Distributed acoustic sensing (DAS) exhibits great potential for underwater target detection. The line-spectrum component serves as a distinctive acoustic fingerprint for ship detection, and its accurate reconstruction relies on precise DAS phase demodulation. However, complex marine ambient noises severely deteriorate the signal-to-noise ratio (SNR) of the line-spectrum component, introducing additional phase jumps and even distortion during the unwrap process. To address these challenges, this work proposes a residual module nested long short-term memory (RMN-LSTM) neural network for high-SNR phase unwrap and line-spectrum target detection, combining the deep spatial feature extraction of residual module and long-range temporal dependency of the LSTM with symmetric left-right encoder-decoder network structure. The RMN-LSTM directly learns the mapping between distorted wrapped and true phase signals, enabling effective recovery of high-fidelity phase information. Experimental results demonstrate that this method accurately rebuilds line-spectrum signals and significantly outperforms conventional algorithms. The root mean square error (RMSE) of RMN-LSTM is reduced to 0.661, representing decreases of 95.5% and 96.3% against numerical optimization methods, 19.3% reduction against classical convolutional neural network (CNN)-based method, and 4.4% reduction against LSTM-based methods, achieving a comprehensive SNR improvement of about 27 dB. This work provides a robust technical approach for high-SNR underwater target detection without hardware upgrades.