面向5G NR URLLC的神经CRC预测
Neural CRC Prediction for 5G NR URLLC
浏览论文内容
中文总结 AI 辅助
该研究针对5G NR URLLC提出一种混合散射神经CRC预测器，绕过传统均衡解码链，通过MCS索引提升可靠性，可部署于实时基带流水线，还能量化认知不确定性。
中文摘要 AI 辅助
我们针对5G新空口（5G NR）物理上行共享信道（PUSCH）提出了一种神经循环冗余校验（CRC）预测器，可为超可靠低延迟通信（URLLC）提供早期链路自适应决策。该预测器将轻量级卷积神经网络（CNN）与固定前端相结合，从接收信号和最小二乘信道估计中提取多尺度时频能量特征。我们研究了两种互补的前端实现：一种是由固定Gabor滤波器构建的小波散射前端，另一种是基于快速傅里叶变换（FFT）的散射前端，该前端在频域中应用具有几何尺度间隔的带通掩码。基于神经接收机设计原理，该预测器直接从接收资源网格估计解码后CRC结果，绕过了传统的均衡和解码链。我们进一步引入调制和编码方案（MCS）索引作为辅助条件输入，使决策边界适应工作码率。在多MCS的5G NR PUSCH数据集上的实验表明，混合散射预测器的性能显著优于纯CNN基线，而MCS条件输入进一步提高了不同信道条件下的可靠性。两种前端均兼容GPU加速推理，且足够轻量，可部署在实时基带流水线中。我们还演示了一种证据深度学习扩展，该扩展使用保守决策规则在单次前向传播中量化认知不确定性。
英文摘要
We propose a neural cyclic redundancy check (CRC) predictor for the 5G New Radio (5G NR) physical uplink shared channel (PUSCH) that enables early link-adaptation decisions for Ultra-Reliable Low-Latency Communications (URLLC). The predictor combines a lightweight convolutional neural network (CNN) with a fixed front-end that extracts multi-scale time-frequency energy features from the received signal and least-squares channel estimates. We investigate two complementary front-end realizations - a wavelet scattering front-end built from fixed Gabor filters, and an FFT-based scattering front-end that applies bandpass masks in the frequency domain with geometric scale spacing. Drawing on neural-receiver design principles, the predictor estimates the post-decoding CRC outcome directly from the received resource grid, bypassing the conventional equalization and decoding chain. We further introduce the modulation and coding scheme (MCS) index as an auxiliary conditioning input that adapts the decision boundary to the operating code rate. Experiments on a multi-MCS 5G NR PUSCH dataset show that the hybrid scattering predictors substantially outperform a pure CNN baseline, with MCS conditioning further improving reliability across varying channel conditions. Both front-ends are compatible with GPU-accelerated inference and are lightweight enough to be deployed within a real-time baseband pipeline. We further demonstrate an evidential deep learning extension that quantifies epistemic uncertainty in a single forward pass using a conservative decision rule.
