Deep Learning Based CSI Compression for Fronthaul Overhead Reduction

Modern base stations exploit distributed architectures defined by the O-RAN standard for efficient operation. In ORAN, the base station functionality is split between a radio unit (RU) and a distributed unit (DU). Further, based on 3GPP split option 7.2, O-RAN specifies two categories of RU: Cat-A (or 7.2A) wherein precoding is performed at the DU, and Cat-B (or 7.2B) wherein precoding operation is offloaded to the RU. Hence, in 7.2B, the precoding information is transferred from DU to RU, but, after application of beamspace compression (BSC). This involves conversion of the precoders to beamspace, which is followed by quantization of its real and imaginary coefficients. However, the BSC approach is suboptimal as it does not fully exploit the underlying correlation in the precoding vectors. To address this limitation, we propose a deep learning (DL) based beamspace precoder compression approach for O-RAN. Using 3GPP compliant simulations, we show that the proposed approach decreases the precoder transfer overhead by up to 85%, while also achieving a spectral efficiency gain of up to 25%. Furthermore, we show that by compressing the precoders in the beamspace, the operational complexity can be drastically reduced, specifically due to reduction in auto-encoder (AE) input dimensionality.