Title: An Enhanced Deep Learning Framework for Massive MIMO CSI Feedback for B5G Applications
Authors: Kabiru Abubakar Yahya and Dr. Ibrahim Babale Gashua
Volume: 10
Issue: 7
Pages: 137-142
Publication Date: 2026/07/28
Abstract:
Massive multiple-input multiple-output (mMIMO) is a cornerstone of fifth-generation (5G) and beyond-5G (B5G) networks, promising exceptional spectral and energy efficiency. In frequency division duplex (FDD) mode, however, the downlink channel state information (CSI) must be fed back from the user equipment (UE) to the base station, and the feedback overhead scales with the number of antennas, rapidly consuming the scarce uplink resources. Deep learning (DL) based CSI feedback, pioneered by CsiNet, has demonstrated remarkable compression and reconstruction performance, yet existing frameworks exhibit critical limitations in generalization across varying channel distributions, robustness to feedback errors, and computational efficiency for real-time deployment. This study proposes an enhanced deep learning framework that integrates a multi-scale feature extraction encoder, a transformer based refinement module, and a lightweight hyper network for bit rate adaptation, jointly optimized via a novel perceptual loss function that penalizes both reconstruction error and precoding performance degradation. Extensive simulations under 3GPP 5G NR channel models demonstrate that the proposed framework achieves up to 4.5 dB improvement in reconstruction normalized mean squared error (NMSE) and 12% higher sum rate relative to state-of-the-art CsiNet variants, while reducing encoder complexity by 35%. The results confirm that the fusion of multi-scale encoding, attention mechanisms and adaptive quantization bridges the gap between theoretical optimality and practical, hardware-constrained deployment of DL-based CSI feedback in massive MIMO systems.