TFLM: A time-frequency domain learning model for underwater acoustic signal reconstruction

Abstract

The reconstruction of underwater acoustic signals affected by seasonal variations is of great significance to improve the accuracy and stability of oceanic communication systems. Seasonal changes in water temperature, salinity, and density cause severe fluctuations and distortions to the underwater acoustic signals, which degrade the signal transmission quality. Existing reconstruction methods are often inadequate in adapting to these complex and dynamic environments and remain susceptible to noise interference. To tackle these challenges, we propose a novel Time-Frequency domain Learning Model (TFLM) for underwater acoustic signal reconstruction. TFLM decomposes the distorted signal into trend and seasonal components for reconstruction. The trend components are reconstructed using an enhanced Long Short-Term Memory (En- LSTM) that effectively captures long-term temporal features. For the seasonal components, a multi-layer encoder–decoder architecture is utilized to extract local features and address seasonal fluctuations. Extensive experimental evaluations demonstrate that TFLM outperforms existing methods in terms of reconstruction accuracy, providing a robust solution for underwater acoustic signal reconstruction under seasonal variability.