Lightweight modulation recognition of underwater acoustic communication signals

Automatic modulation recognition (AMR) of underwater acoustic (UWA) communication signals is crucial in signal demodulation, signal monitoring, interference identification, and underwater electronic countermeasures. However, the time-varying characteristics and multipath interference of the UWA channel pose great challenges to the reliability of the modulation recognition. With the development of machine learning technology, machine learning based modulation recognition schemes are applied to improve the modulation recognition accuracy and robustness. Due to the limited computing resources on underwater platforms, the AMR of the UWA signal demands applications with small storage footprint, low processing and memory requirements, and high energy efficiency. In this paper, we propose a lightweight modulation recognition scheme for the UWA communication signals. Firstly, the power spectrum, the square spectrum, and the correlation analysis are applied to the lightweight feature extraction of the common UWA communication signals. Secondly, to enhance the classification performance of the model, we propose the residual spatial-channel attention MobileNet1D (RSCA-MobileNet1D) classifier by devising a hybrid attention mechanism RSCA-Net, and incorporating it into the lightweight network MobileNet1D model through a designed RSCA-Bottleneck. Finally, based on a real world measured dataset and a simulated dataset with different signal-to-noise ratio (SNR), experiment results show that compared to the baseline model, RSCA-MobileNet1D achieves the highest recognition accuracy with relatively low computational burden.