A Novel OFDM Equalizer for Large Doppler Shift Channel through Deep Learning

In this paper, we propose a practical deep neural network for OFDM symbol equalization and demonstrate its advantages in combating large Doppler Shift. In particular, a novel zero-forcing initialized neural architecture named Cascaded Net (CN) is proposed for equalization, where deep trainable network is cascaded behind a zero-forcing preprocessor to prevent the network getting stuck in a saddle point or a local minimum point. In addition, we propose a sliding equalization approach to detect those OFDM symbols with large number of subcarriers. We also evaluate this novel equalizer, as well as the sliding algorithm, using Rayleigh fading channel with large Doppler shift. The numerical results show this novel equalizer can achieve better performance than zero-forcing equalizer and classical ICI cancellation methods in SISO scenario. Thanks to proper training methods, this equalizer is relatively robust to traditional methods when channel estimation is inaccurate or Doppler shift changes.