Channel Replay Aided Siamese Network for Automatic Modulation Recognition of Underwater Acoustic Communication

In non-cooperative scenarios, automatic modulation recognition (AMR) of underwater acoustic communication (UAC) faces significant challenges due to the lack of prior information and the impact of fluctuating channel conditions. To address the low recognition rate caused by the above issues, we propose a novel signal modeling approach based on co-originating sample pairs (CSPs), which deconstructs signals into common-mode and differential-mode. The common-mode showcases the modulation properties of the communication signal, whereas the differential-mode reveals the characteristics of the channel. Using the common-mode of received signals for AMR can improve its adaptability and robustness in variable channels and low signal-to-noise ratio (SNR) environments. Building on this, we introduce a framework supported by a channel replay aided dual-branch Siamese network (CRaSN). The channel replay technique is employed to obtain CSPs of standard communication signals, with the dual-branch Siamese framework processing a sample from each pair individually. To empower this framework with the capability to extract common-mode, we devised a common-mode consistency constraint (CMCC) loss function, inspired by information theory principles, and merged it with a contrastive-like strategy to direct the model's training. Experimental results demonstrate that our proposed CRaSN method outperforms baseline algorithms at all SNR levels, with a recognition rate exceeding 90% at SNR >= -2 dB and surpassing 96% at SNR >= 0 dB.