Title: ReFLEX: Achieving Length-Generalizable MIMO-OFDM CSI Denoising Through Relative-Frequency Bias

Abstract:

This study investigates Channel State Information (CSI) denoising techniques for MIMO-OFDM systems, specifically addressing scenarios with variable New Radio (NR) resource block (RB) allocations. We introduce ReFLEX, a Transformer-based model capable of generalizing across different lengths. Its frequency attention mechanism incorporates a relative-frequency position bias (RFPB), which is derived from subcarrier offsets. Notably, a single trained checkpoint can manage unseen RB lengths and is applicable to sparse Demodulation Reference Signal (DM-RS) observations within a tested RB5/RB10 PUSCH configuration, eliminating the need for retraining.

Performance evaluations in a 3GPP TR 38.901 Urban Micro (UMa) Non-Line-of-Sight (NLOS) channel environment demonstrate that ReFLEX achieves a Normalized Mean Square Error (NMSE) of approximately $-9.6$ dB when processing unseen RB lengths. Furthermore, simulations of NR PUSCH/UL-SCH indicate that employing ReFLEX for denoising prior to time-frequency interpolation lowers the 10% Block Error Rate (BLER) threshold by roughly 2–3 dB.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC