Convolutional recurrent neural network-based channel equalization: An experimental study

In this paper, we revisit the idea of using deep neural network for channel equalization to account for nonlinear channel distortions as well as temporal variations of radio signals. Our insight is leveraging the the shift-invariant properties of the convolutional neural network (CNN) to learn matched filters analogous to the tap weights of conventional equalizer. Then we feed the learned filters into a subsequent recurrent neural network (RNN) with long-short-term-memory (LSTM) cells for temporal modeling of the channel. We train our proposed CNN-RNN (CRNN) equalizer based on real testbed collected data and enlarge the generalization ability of the learned network model as much as we can to adapt to different channel conditions. Experimental results show that the SER performance for our designated single-input single-output (SISO) system which utilises quadrature phase shift keying (QPSK) modulation scheme with the proposed CRNN-based channel equalizer outperforms that of other equalizers by average 2 to 5 dB at low signal-to-noise ratio (SNR).