A lightweight model for modulation recognition of integrated underwater acoustic sensing and communication signals

Under the stringent resource constraints of underwater edge nodes and the severe impairments of underwater acoustic channels, automatic modulation recognition of integrated underwater acoustic sensing and communication signals (UISAC) remains challenging, especially when both lightweight deployment and robust performance are required. To address this issue, we propose Mixer, an ultra-lightweight network that combines signal enhancement with efficient feature representation. First, the impulsive-noise preprocessing and adaptive spectral block suppress impulsive noise and selectively enhance informative frequency subbands. Second, spatial density functions are coupled with complex-valued convolutions to capture local signal patterns and amplitude-phase coupling in the received signal. Third, the DWMG backbone integrates grouped depthwise-separable convolutions with an MLP-Mixer to efficiently fuse local and global features with low computational overhead. Experimental results under standard, generalization, and robustness evaluation settings show that Mixer remains competitive in the low-signal-to-noise ratio regime and provides a favorable balance between lightweight design and recognition performance. Compared with ULCNN, Mixer reduces the parameter count by 46.96%, and its single-sample CPU inference latency is 18.5 ms, demonstrating its potential for real-time deployment on UISAC edge nodes.