Transfer-learning-adapted CNN-BiLSTM nonlinear equalizer for DSCM-enabled microwave-band fiber-wireless integration...

Fiber-wireless integration (FWI) provides a promising architecture for flexible and high-capacity fronthaul transmission, while digital subcarrier multiplexing (DSCM) enables subchannel-wise resource allocation with different modulation formats. However, in practical DSCM-enabled FWI systems, cascaded wireless and optical impairments lead to subchannel-dependent residual distortions after classical linear equalization, thereby limiting the uniform recovery of all transmitted subchannels. In this paper, we propose a DSCM-enabled microwave-band FWI transmission scheme integrated with a transfer learning (TL)-adapted convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) nonlinear equalizer. The proposed equalizer operates on the linear-equalized symbol sequence of each subchannel after classical digital signal processing (DSP) and predicts a residual correction for the target center symbol. A progressive TL adaptation strategy is further developed to reduce the training overhead of subchannel-specific equalization, in which a source model trained on a selected subchannel is adapted to impaired target subchannels via fully connected (FC) warm-up and full-network fine-tuning. The proposed scheme is experimentally validated in a 17.5-GHz microwave-band FWI system over a 2-m wireless link and a 25-km single-mode fiber-28 (SMF-28) link. Experimental results show that the CNN-BiLSTM nonlinear equalizer consistently reduces the average error vector magnitude (EVM) from 34.1% to 6.3% for quadrature phase shift keying (QPSK), from 21.7% to 8.7% for 8-ary quadrature amplitude modulation (8QAM), and from 19.6% to 15.1% for 16-ary quadrature amplitude modulation (16QAM). Moreover, with only 10% target-domain data, the TL-adapted equalizer reduces the EVM of QPSK-modulated target subchannel from about 41% to approximately 22%, approaching the full-training result of about 20%. These results verify the effectiveness and data efficiency of the proposed TL-adapted CNN-BiLSTM nonlinear equalizer for DSCM-enabled FWI transmission.