A hybrid CNN-transformer based automatic modulation recognition system for underwater acoustic communication: Method and implementation

Abstract

Underwater acoustic (UWA) communication plays an essential role in marine exploration and oceanic resource utilization. Automatic modulation recognition (AMR) is a key technique for identifying the modulation of received signals without relying on prior transmitter information, and it is crucial for ensuring reliable and adaptive UWA communication. However, the AMR of UWA signals remains a challenging task due to severe multipath propagation, Doppler effects, and dynamic channel characteristics. To address these issues, this paper proposes a hybrid CNN-Transformer architecture, for efficient and accurate AMR. The CNN module with Multi-Scale Complex-Valued convolutions extracts robust local features under multipath and Doppler distortions, while the Transformer with a window self-attention mechanism captures long-range dependencies with reduced computational complexity. This design effectively balances recognition accuracy and computational efficiency, enabling deployment on edge devices. Experimental results across three field-collected datasets covering lake, open-sea and gulf environments demonstrate that the proposed model consistently delivers superior AMR performance, achieving accuracies of 98.83%, 95.75%, and 92.29% under increasingly challenging channel conditions. Additionally, a complete UAC-AMR system is designed and implemented in this study, where the proposed model achieves an NPU inference latency of 0.96 ms on the edge platform, fully meeting the real-time requirements of UWA communication.