🕒 3 mins

 

CENTRIC @ EUCNC & 6G Summit 2025

Demonstrating Validation Setup for

AI-based CSI Compression with Hardware-in-the-loop

The fundamental vision of CENTRIC consists of 6G networks in which the air interface consisting of physical layer and the supporting protocol stack is devised and optimized in an automated manner for each use-case, application, or user-specific conditions using the power of cutting-edge AI and ML methods.

While it is expected that the 6G air interface will be designed following a native-AI framework, 3GPP initiated efforts towards introducing and specifying some AI features applicable to few candidates use cases in 5G-Advanced Release-19.  One of the considered use cases is the AI-based channel state information (CSI) feedback enhancements known as AI/ML for CSI compression, where an autoencoder-type neural network is deployed to compress and transmit the estimated downlink CSI at the UE  through an uplink transmission and decompress the compressed CSI at the gNB with minimum information loss.

This demo showcases a hardware-based validation setup for AI models performing CSI compression, developed in the CENTRIC project, including the following main elements:

1. Multi-port RF transceiver (Keysight MTRX E6464A)
2. RF Channel emulator (Keysight PROPSIM F8800A)
3. PC with GPU for AI processing
4. AI model under test (Nokia’s CSI Compression Model)

The CENTRIC project uses this setup to validate a low complexity transformer-based AI model for CSI compression developed by Nokia. The approach introduces a lightweight transformer-based neural network with an autoencoder architecture that achieves state-of-the-art performance with fewer parameters within the transformer-based architecture. As widely accepted in 3GPP, the system distributes processing by implementing the encoder on the User Equipment (UE) side and the decoder on the gNodeB side, enabling compression and accurate reconstruction of CSI across spatial and frequency dimensions. This architecture improves reconstruction accuracy while reducing CSI feedback overhead in uplink transmission. The model employs Normalised Mean Squared Error (NMSE) as a reconstruction metric for decompressed  CSI.

The CENTRIC project uses this setup to validate a low complexity transformer-based AI model for CSI compression developed by Nokia. The approach introduces a lightweight transformer-based neural network with an autoencoder architecture that achieves state-of-the-art performance with fewer parameters within the transformer-based architecture. As widely accepted in 3GPP, the system distributes processing by implementing the encoder on the User Equipment (UE) side and the decoder on the gNodeB side, enabling compression and accurate reconstruction of CSI across spatial and frequency dimensions. This architecture improves reconstruction accuracy while reducing CSI feedback overhead in uplink transmission. The model employs Normalised Mean Squared Error (NMSE) as a reconstruction metric for decompressed CSI.

After jointly training the encoder and decoder with quantisation using synthetic data based on  3GPP stochastic channel models, they are deployed in the PC (3) for processing of the testbed data. The trained models are validated by generating 5G-NR downlink signals with MTRX (1), which are then fed to the channel emulator (2) where convolution with a MIMO channel is emulated in RF, yielding the signal received at the UE side. This signal is fed back to MTRX (1), which digitises and feeds the IQ samples to the PC (3), where CSI estimation takes place based on the received samples. The AI models at UE and gNB are subsequently run to compress and decompress the estimated CSI, based on which a maximal ratio transmission (MRT) precoder is calculated. Finally, a precoded downlink transmission is emulated in the same way as the initial transmission, with the received IQ samples being fed again to the PC for link-level analysis.

With this setup, we can evaluate the performance of Nokia’s AI model for encoding and decoding of CSI vs two baselines: a) Perfect CSI knowledge at the gNB, b) estimated but uncompressed CSI at the gNB.