Deep learning-based image-like channelization for broadband receiver

We propose a novel image-like channelization method that utilizes a convolutional recurrent neural network (CRNN) for channel synthesis to reduce the bandwidth requirements of the electrical hardware. In this study, the spectrum of a 30-GBaud QPSK signal is spectrally sliced and received by four low-speed coherent receivers based on a conventional coherent optical communication system. After the recovery of the trained CRNN, the average error vector magnitude (EVM) of the 30-GBaud baseband signal is improved from over 60% by uncorrected channel synthesis to around 15%.