A BiGRU-MCNN-Attention receiver for uncoded SISO underwater acoustic OFDM: Real-water experiments, simulation, and cross-environment transfer learning

Underwater acoustic (UWA) orthogonal frequency-division multiplexing (OFDM) systems are subject to multipath propagation, frequency-selective fading, and environment-dependent distortion. Although conventional model-based receivers remain effective in many cases, their performance can degrade when channel estimation is inaccurate or deployment conditions differ from design assumptions. This paper investigates a data-driven receiver for uncoded single-input single-output (SISO) UWA-OFDM based on bidirectional gated recurrent units (BiGRU), multi-scale one-dimensional convolutional neural networks (MCNN), and lightweight feature attention. A formal waveform model, pilot-assisted least-squares (LS) baseline for real-water and linear minimum mean-square error (LMMSE) baselines for simulation, and a unified frequency-domain learning pipeline are presented for reproducible comparison. The receiver is evaluated in controlled tank experiments, more challenging fish pond experiments, and Watermark-based simulations. In the large-water-tank setting at 0.0015 W, the proposed receiver achieves 99.994 % test accuracy with a bit error rate (BER) of 3.3 x 10 − 5 . In the fish pond environment, it achieves 98.315 % test accuracy and a BER of 9.636 x 10 − 3 at 0.005 W, and 99.402 % test accuracy with a BER of 3.268 x 10 − 3 at 0.01 W. Simulation results show consistent BER reduction relative to LS and LMMSE across SNR conditions, supporting data-driven demodulation as a complementary receiver strategy under channel mismatch rather than as a replacement for coded practical links.