Auto-Regressive RF Synchronization Using Deep-Learning

This work presents a novel pilot-less Deep-Learning-based synchronization mechanism that seamlessly integrates within state-of-the-art auto-encoder-based end-to-end communication systems. By re-using the idea of Radio Transformer Networks, an auto-regressive strategy is designed that learns to estimate and mitigate synchronization-related perturbations for arbitrarily modulated continuous communication, i.e., sample time offset (STO) and carrier frequency offset (CFO). A performance gain of 0.6 dB in the high-SNR regime compared to classic synchronization techniques is demonstrated. The strength of this approach is a shift from sample-by-sample to batch-wise processing according to the ML paradigm, which enables efficient and fast computation required for practical deployment scenarios using hardware-accelerated ML inference engines.