Robust BICM Design for the LDPC Coded DCO-OFDM: A Deep Learning Approach

Abstract

In this paper, a deep learning (DL) approach for enhancing the bit-interleaved coded modulation (BICM) receiver is designed to mitigate the clipping distortion in the low-density parity-check (LDPC) coded direct current-biased optical orthogonal frequency division multiplexing (DCO-OFDM) systems. This work aims to combine the neural network (NN) with the physical layer communications by using a model-driven DL architecture. We first develop a non-iterative NN-aided BICM (NN-BICM) receiver, where the NN is trained with the loss function of cross-entropy to output the modified conditional probability through the softmax activation function, thereby assisting in a log-likelihood ratio (LLR) improvement. Then, we propose two iterative NN-BICM receivers for iterative demapping and decoding. The single iterative design feeds the soft decisions from the LDPC decoder back to the demapper only, while the joint iterative design feeds the soft decisions back to the demapper and NN jointly. By adopting the iteration-wise pre-training strategy, the joint iterative design has been improved by representing the intractable relationship between the conditional probability and the a priori probability with a deeper NN architecture. We further investigate an efficient bit loading algorithm for DCO-OFDM systems employing the NN-BICM receiver. Both NN-BICM receivers and iterative schemes can obtain remarkable performance gains over the existing benchmarks.