Hansong Ma

To address the issues of subcarrier orthogonality loss and inter-carrier interference (ICI) caused by carrier frequency offset (CFO), this paper proposes and experimentally validates a frequency offset estimation (FOE)-assisted dual-domain Transformer equalizer within an advanced, high-capacity optical-heterodyne radio-over-fiber (RoF)–wireless orthogonal frequency division multiplexing (OFDM) transmission system. To rigorously test the algorithm’s robustness under extreme physical conditions, the experimental platform integrates offline 16-GBaud signal generation, optical I/Q modulation, dual-optical-tone transport over a single-mode-fiber RoF feeder, remote photonic heterodyne frequency conversion based on a uni-traveling-carrier photodiode (UTC-PD), 4.6 km free-space wireless transmission, and 160-GSa/s ultra-high-speed real-time sampling. In this system, the receiver front-end employs an FOE module to pre-compensate for the dominant global CFO-induced phase rotation; subsequently, a low-complexity, compact local-window Transformer is utilized to perform adaptive residual compensation for local data-dependent impairments—such as residual waveform distortion and residual ICI—in both the time and frequency domains (before and after the Fast Fourier Transform, or FFT). This synergistic architecture, combining a physical model-driven approach with a self-attention mechanism, effectively mitigates the adverse impact of global frequency offset on neural network convergence. Experimental results demonstrate that, under conditions of strictly aligned multiply accumulate (MAC) operation complexity, the dual-domain architecture achieves significantly superior performance—in terms of bit error rate (BER), error vector magnitude (EVM), and constellation quality—compared to traditional linear DSP methods and baseline networks such as DNNs, CNNs, and LSTMs. Operating in 16 GBaud QPSK mode with an input optical power of 0 dBm, the system achieves a BER of $1.89\times {10}^{-4}$, representing performance improvements of approximately 5.98-fold and 1.92-fold over the standalone Transformer and FOE-assisted DNN schemes, respectively.