DL-NA-SBD: An Unsupervised Online Deep Learning Approach for Blind Channel Equalization

In contrast to most of the existing equalization methods, blind equalization (BE) can eliminate the effect of multipath fading without any known sequences. As a result, BE is a promising technique for wireless intelligence and non-cooperative communication. However, the conventional BE methods require long sequences or large-scale training to recover the received signals. In this paper, a deep learning-based neighborhood-assisted symbol-based decision (DL-NA-SBD) method is proposed to tackle this problem. Specifically, we replace the commonly used linear approaches with neural network to optimize the traditional SBD function and the mean square error (MSE) loss function in two stages to generate the equalizer coefficients. To avoid collecting large numbers of signals, we train the network using the online training strategy. Simulation results demonstrate that the proposed method achieves better inter-symbol interference (ISI) elimination and bit error rate (BER) performance compared with the conventional methods while requiring only a short-length signal sequence.