PrecoderNet: A Feedback Framework for Massive MIMO-OFDM Systems

Abstract

Efficient channel state information (CSI) feedback is critical for enhancing spectral efficiency in frequency division duplex (FDD) massive MIMO systems. This paper introduces PrecoderNet, a deep learning (DL)-based framework that avoids full CSI reconstruction by directly estimating the singular value decomposition (SVD)-based precoding matrix. This framework consists of a DL-based encoder and decoder. The encoder employs multi-scale feature extraction, multi-head self-attention, and an unpatched convolution block to generate codewords. The decoder integrates a multi-head self-attention mechanism, an unpatched convolution block, and an EnhanceNet block to reconstruct the dominant right singular vectors for two transmission layers from the received codewords. This design inherently reduces feedback overhead as the number of receive antennas increases. Moreover, it mitigates scalability issues by removing the input’s dependency on the number of receive antennas. Despite its higher computational complexity, PrecoderNet achieves substantial overhead reduction and delivers significant reconstruction accuracy for different compression ratios (CRs), making it well-suited for 5G-advanced and beyond.