GRU-Based Sequence Detection for Faster-than-Nyquist Signaling

This paper presents a deep learning-based detector for faster-than-Nyquist (FTN) signaling that leverages a Gated Recurrent Unit (GRU) architecture optimized using the Nesterov-accelerated Adaptive Moment Estimation (NADAM) algorithm. Compared with Long Short-Term Memory (LSTM) networks commonly employed in similar detection tasks, GRUs offer improved computational efficiency, while NADAM contributes to stable and effective convergence in non-convex optimization settings. Rather than relying on generic neural models, the proposed design explicitly aligns the GRU input structure with the one-sided inter-symbol interference (ISI) span of FTN signaling, enabling the network to learn interference patterns efficiently without incurring unnecessary complexity. This structured integration results in reduced computational burden and enhanced convergence behavior. Simulation results demonstrate that the NADAM-optimized GRU achieves bit error rate (BER) performance close to the optimal BCJR algorithm for$\tau \geq 0.7$, while offering superior computational efficiency compared with conventional deep learning-based detectors. A detailed complexity comparison with the M-BCJR algorithm shows that the proposed approach reduces hardware resource usage—measured in Look-Up Tables (LUTs)– -by up to 76% while maintaining comparable BER performance in the same$\tau$regime. Additional evaluations further highlight its robustness, demonstrating reliable performance in quasi-static multipath Rayleigh fading channels and strong compatibility with LDPC-coded FTN transmission. These results collectively underscore the practicality and efficiency of the proposed GRU-based FTN detector.