COMMUNICATIONS THEORIES & SYSTEMS
Liu Zhao, Wang Meng, Lin Kai, Li Shuang, Yang Xinghai, Wang Jingjing
The complexity of underwater environments renders underwater acoustic signals vulnerable to various forms of noise during transmission, creating significant challenges for signal demodulation tasks. This paper presents a novel demodulation method for multi-class single-carrier underwater acoustic signals. Our approach employs two innovative structures for modeling in the time and frequency domains, integrating these features for comprehensive discrimination. Specifically, we introduce a High-Efficiency Convolution (HEC) Block to extract time-domain waveform features and a Local-Global Attention (LGA) structure for time-frequency features, utilizing cross-attention to fuse these features. This method enables the network to learn hidden frequency, phase, and amplitude characteristics within high-dimensional features, effectively capturing both fine-grained local and long-distance global features. A classifier is then constructed to categorize multi-class modulation signals, completing the demodulation process. Simulation results highlight the method's exceptional performance: in a Gaussian channel with a signal-to-noise ratio (SNR) of 0 dB, the demodulation error rates for 2PSK, 2FSK, 2ASK, 4PSK, 4FSK, and 8PSK signals are all below 0.01, while the error rate for 16QAM modulated signals is less than 0.1. Additionally, validation using BELLHOP simulation data and real-world data collected from the Yellow Sea further demonstrates the proposed method's remarkable noise resistance and demodulation capabilities.