When Does a Neural Receiver Help? Calibration-Drift Benchmarking and Detect-and-Rollback for 5G/6G NR

核心问题

Determine when a neural receiver helps for 5G/6G NR detection under calibration drift.

场景：5G/6G NR physical uplink shared channel detection with convolutional neural receivers, compared against MMSE, under in-distribution and drifted transmitter/channel conditions.

主要方法

Calibration-drift benchmarking evaluates neural receiver behavior when transmitter or channel parameters leave the training envelope. The comparison is framed against minimum mean-square error detection for 5G/6G NR PUSCH, separating in-distribution gains from drifted-condition reliability. Detect-and-rollback suggests a runtime or evaluation mechanism that detects unfavorable neural-receiver conditions and reverts to a safer receiver path, although the payload does not expose the decision rule.

关键贡献

Frames neural receiver utility as a robustness question for 5G/6G NR rather than only an in-distribution performance comparison. Focuses on calibration drift, where transmitter or channel parameters depart from the training envelope, as the key stressor for convolutional neural receivers. Positions DeepRx-like convolutional receivers against MMSE PUSCH detection to identify when learned detection helps and when it may become unsafe. Introduces or studies a detect-and-rollback approach intended to preserve practical receiver reliability under drift.

研究启发

What calibration-drift dimensions are benchmarked, such as transmitter impairments, channel model mismatch, waveform parameter changes, or hardware calibration errors? How does detect-and-rollback decide that the neural receiver should be abandoned, and what baseline does it roll back to? Are there quantitative BLER, throughput, or robustness results showing when DeepRx is better or worse than MMSE under drift? Does the benchmark include multiple SNR regimes, channel models, mobility settings, or NR numerologies?