Turbo Detection Aided Autoencoder for Multi-Carrier Wireless Systems: Integrating Deep Learning into Channel Coded Systems

Abstract

A variety of deep learning schemes have endeavoured to integrate deep neural networks (DNNs) into channel coded systems by jointly designing DNN and the channel coding scheme in specific channels. However, this leads to limitations concerning the choice of both the channel coding scheme and the channel paramters. We circumvent these impediments and conceive a turbo-style multi-carrier auto-encoder (MC-AE) for orthogonal frequency-division multiplexing (OFDM) systems, which is the first one that achieves the flexible integration of DNN into any given channel coded systems while achieving an iteration gain. More explicitly, first of all, we design the MC-AE independently of both the channel coding arrangement and of the channel model, where the output layer of the MC-AE decoder is designed for both accepting and producing reliable soft-bit decisions. Owing to the fact that bit-dependency is imposed by the MC-AE mapping, our bespoke MC-AE decoder becomes capable of achieving a beneficial iteration gain, when the extrinsic information is exchanged between the soft-decision MC-AE decoder and the soft-decision channel decoder. Secondly, in order to be able to interpret the performance advantages of our MC-AE over the conventional OFDM, we map the MC-AE’s input-output relationship to an equivalent model-based representation. The associated theoretical analysis verifies the fact that during the process of data-driven signal reconstruction across OFDM’s subcarriers, a beneficial frequency diversity gain is achieved by the proposed MC-AE design. Finally, our simulation results demonstrate that the MC-AE is capable of achieving substantial performance advantages over both conventional OFDM and OFDM based index modulation (OFDM-IM) in channel coded systems.