Maximizing Uplink Potential: AI-Based Channel Selectivity Matching Using Compressed Channel State Information

In 5G-Advanced and 6G wireless networks, uplink throughput is a key performance indicator, but there are practical constraints to improving uplink spectral efficiency. Current 3GPP standard uplink transmissions rely on wideband codebook precoding, making it difficult to finely reflect subband-specific frequency-selective channel characteristics.

According to the 3GPP Rel-19 RAN workshop, subband precoding can provide throughput improvements of 13% in a 4-Tx antenna configuration and 23% in an 8-Tx configuration. However, subband non-codebook precoding causes a sharp increase in radio control signal overhead as the number of antenna ports and subbands increases, and the Power Amplifier (PA) performance and PAPR limitations of the terminal also make implementation difficult.

To address this, Samsung researchers proposed an AI-based high-resolution uplink precoding framework called SQNC-T. This approach combines a Transformer-based autoencoder with learnable Segment Vector Quantization (SVQ) to compress the subband non-codebook precoding matrix into small signal codewords. Simultaneously, Subband Power Equalization (SWE) is applied to satisfy the terminal's PA constraints.

The operation process is as follows:

• The base station estimates channel state information using SRS.
• The subband average channel covariance matrix is eigen-decomposed to calculate the optimal non-codebook precoding matrix.
• The AI encoder extracts spatial and frequency correlations of the precoding matrix and compresses it into a low-dimensional latent vector.
• SVQ divides the latent vector into multiple segments and quantizes them using learned codebook vectors.
• The compressed codeword indices are transmitted wirelessly, after which the terminal restores the precoding matrix with the decoder and performs power equalization.

The end-to-end AI model consists of an encoder, SVQ quantizer, dequantizer, and decoder. During the learning process, it jointly optimizes a reconstruction loss to improve the restoration quality of the precoding matrix, an encoder loss to make the encoder's latent vectors close to the codewords, and a clustering loss to ensure codewords converge to cluster centers. The goal is to implement high-resolution uplink precoding at the subband level while reducing control signal overhead.