A Double-CNN BP Decoder on Fast Fading Channels Using Correlation Information

Abstract

This paper develops a novel double convolutional neural network (CNN) based belief propagation (BP) decoder to improve the error correcting performance on fast fading channels with correlated channel gain and correlated noise. The proposed double-CNN BP decoder consists of two CNNs for predicting channel gain and noise samples, respectively, concatenated with a BP decoder. The input of the BP decoder is the log-likelihood-ratio (LLR) values obtained according to the predicted channel gain along with denoised signals based on predicted noises. We note that the residual noises no longer obey a Gaussian distribution when denoising the signals. Thus, we further derive a new method to obtain the LLR values accurately. Simulation results show that the proposed double-CNN BP decoder achieves a maximum of 7dB gain compared to the conventional BP decoder and the proposed algorithm compensates the performance loss of BP decoder caused by correlation information.