JSCM and Modulation for CSI Feedback: Comprehensive Research Toward 6G Realization

CSI (Channel State Information) feedback is necessary for base stations to determine the direction of wireless signals and beamforming methods, but as the number of antennas increases, the amount of data to be transmitted also increases. Existing AI autoencoder methods can compress CSI, but they suffer from a cliff effect where performance drops sharply if bit errors occur during transmission.

Instead of processing compression, channel coding, and modulation as separate steps, Samsung researchers applied JSCM (Joint Source-Channel Coding and Modulation), which integrates these functions into a single AI model for joint learning. This approach ensures that CSI reconstruction performance degrades gradually even when channel conditions worsen, and verifies implementability in actual hardware and environments.

To facilitate practical application, the researchers addressed the following four issues:

• Handling real-world wireless environments: Residual fading, RF hardware defects, and various distortions were incorporated into the training process, and the PUCCH structure was redesigned to be compatible with existing 5G protocol stacks.
• Variable resource allocation: CSI information is organized into priority layers, ensuring that critical information is transmitted first when resources are scarce. A single model can handle various feedback resource sizes.
• Mitigating PAPR issues: Gaussian-shaped signals learned by AI can cause a high Peak-to-Average Power Ratio. The proposed annular-Gaussian modulation shaping balances between Gaussian distribution and a constant-power ring shape to strike a balance between accuracy and amplifier burden.
• CQI calculation: The terminal adds synthetic noise to the compressed representation and runs the base station decoder locally to estimate the quality of the CSI that will be reconstructed at the base station. Based on this, an appropriate CQI (Channel Quality Indicator) can be calculated.

To confirm the feasibility of JSCM, the researchers built a hardware PoC testbed and conducted verification in real-world environments.