A Deep Learning-Based Automatic Modulation Recognition Scheme for Single-Channel OFDM QAM Signals in a Radio-over-Fiber System

To meet the requirements of ultra-high speed and large capacity for next-generation communication systems, radio-over-fiber (RoF) converged communication has emerged as a pivotal technical direction for broadband access. Meanwhile, automatic modulation recognition (AMR) in multi-modulation format scenarios is one of the key enabling technologies for realizing adaptive demodulation at the receiver. Therefore, aiming at the problem that existing AMR algorithms are insufficiently applicable in RoF systems, this paper proposes an integrated scheme for the generation, transmission and automatic modulation recognition of single-channel orthogonal frequency division multiplexing (OFDM) signals based on in-phase/quadrature (I/Q) modulation and heterodyne coherent detection. At the receiver end, a deep learning-based automatic modulation recognition model integrating multi-scale residual structures, the convolutional block attention module (CBAM) and bidirectional long short-term memory (Bi-LSTM) temporal modeling is introduced, which achieves high-precision modulation format classification for the received I/Q time-domain sequences. Simulation results demonstrate that the proposed model achieves stable and high-precision recognition of multiple modulation schemes including QPSK, 8QAM, 16QAM, 32QAM and 64QAM over a wide optical signal-to-noise ratio (OSNR) range. Meanwhile, after the completion of modulation recognition and demodulation, the bit error rates (BERs) of 4QAM, 8QAM, 16QAM and 32QAM signals can be reduced to below the hard-decision forward error correction (HD-FEC) threshold of 3.8×10³. Furthermore, the robustness of the proposed model is systematically evaluated by introducing typical physical layer impairments, including variations in laser linewidth, fiber length fluctuations, as well as phase random walk and chromatic dispersion. The results demonstrate that the proposed model can still maintain superior recognition performance under conditions of intense phase disturbance and low signal-to-noise ratio (SNR). Furthermore, ablation experiments have further verified the pivotal roles of the attention mechanism in resisting noise interference and the LSTM structure in suppressing phase rotation. The aforementioned research results demonstrate the feasibility, robustness and engineering applicability of the proposed AMR scheme in RoF systems, which provides a solid theoretical foundation and technical support for the introduction of adaptive modulation technologies in future broadband fiber-wireless converged access systems.