A model‐driven robust deep learning wireless transceiver

Abstract

Abstract Recently, deep learning (DL) has been successfully applied in computer vision and natural language processing. The communication physical layer based on deep learning has received widespread attention. Introducing domain‐knowledge into neural networks (NNs), autoencoder based end‐to‐end communication system, incorporating radio transformer networks (RTNs) (RTNs‐AE) has achieved desirable performance with a channel model in the middle layer. The advent of RTNs underscores the power of expertise at DL. However, a tap‐length in the design of RTNs network must be assumed, which requires some channel information. To address this issue, a new deep learning wireless transceiver named pilot‐aided autoencoder (PA‐AE) is proposed. It can decode on a multipath fading channel without knowing the channel information and the equalization module design. The proposed scheme introduces a well‐designed auxiliary pilot, which carries the learned channel information into decoding with the transmitted signal. The decoding part recovers the sent information from the collected signal without specially designed modules for parameter estimation and equalization.

Keywords

• Modulation and coding methods
• Communication channel equalisation and identification
• Mobile radio systems
• Communications computing
• Neural nets