Distributed Learning Assisted Fronthaul Compression for Multi-Antenna C-RAN

Abstract

This paper investigates the uplink reception in the cloud radio access network (C-RAN), sometimes called centralized-RAN. C-RAN is an architecture for cellular networks which consists of many remote radio units (RRUs) connected to a central processor (CP). Due to the prohibitive complexity of computations, the most efficient uplink C-RAN schemes are challenging to be implemented in practical systems. Using deep neural networks (DNNs), we propose a new and low complex method for uplink C-RAN subject to some quantization rules. This is the first work that uses DNNs to mimic the C-RAN system to the best of our knowledge. Our architecture’s objective, called QDNet, is to jointly optimize the processing done at the RRUs, which considers the quantization constraints and the processing done at the CP side. Inspired by the projected gradient descent algorithm, QDNet is designed as a distributed DNN with sparse connections. The performance of QDNet is compared to current solutions such as the zero-forcing (ZF) equalizer and the sphere decoder (SD). In some scenarios, experiment results show that our scheme achieves 2 dB SNR gain over the linear ZF with the same or lower computational complexity. It also achieves near-optimal performance compared to the SD algorithm, especially for low-to-moderate fronthaul link capacity and many RRUs in the C-RAN system.

Keywords

• Cloud-RAN
• deep learning
• deep neural networks
• MIMO
• signal detection