跳转至

GCNO:用于无线信道物理压缩的格拉姆切比雪夫神经算子

文章背景与核心概要

现代无线通信系统依赖大规模天线阵列来提升数据速率并服务更多用户,但这导致了沉重的信道反馈开销,要求接收设备不断向基站报告庞大且复杂的信道矩阵。传统的神经压缩方法通常将这些矩阵视为标准图像,将其编码为固定长度的码字,并依赖专门的神经解码器进行恢复。这种方法在面对多变的信道复杂度时适应性较差,且当阵列天线数量发生变化时,通常需要进行大量的重新训练。

为了解决这些局限性,作者提出了“格拉姆切比雪夫神经算子”(Gramian Chebyshev Neural Operator, GCNO),这是一种基于物理特性的变速率压缩方案。GCNO 不再将信道数据视为任意的像素网格,而是利用收发信道固有的物理结构,识别出与样本相关的潜在传播路径方向。该方法通过定位主导路径、利用一阶泰勒修正微调离散网格间的方向,并结合最小二乘法精确恢复复数路径强度,实现了高效压缩。

GCNO 在完全无需显式路径标签的情况下进行训练,允许基站通过传输的路径元组解析地重建信道,而非依赖于黑盒式的神经解码器。在三种不同的射线追踪环境中,GCNO 在同等负载下表现出更高的重建精度,或在同等精度下实现了更低的负载,并能无缝迁移至未见过的天线规模,无需重新训练。


📌 摘要 (Summary)

Modern wireless systems rely on large antenna arrays to boost data rates and serve more users, but this creates a heavy channel feedback overhead—requiring receiving devices to continually report large, complex-valued channel matrices to the base station. Conventional neural compressors treat these matrices like standard images, encoding them into fixed-length codes that require specialized neural decoders. Consequently, these systems fail to adapt to varying channel complexities and typically need extensive retraining when antenna counts change.

现代无线系统依赖大型天线阵列来提高数据速率并服务更多用户,但这产生了沉重的信道反馈开销——要求接收设备不断向基站报告庞大的复值信道矩阵。传统的神经压缩器将这些矩阵视为标准图像,将它们编码为需要专门神经解码器的固定长度代码。因此,这些系统无法适应变化的信道复杂度,并且在天线数量变化时通常需要大量的重新训练。

To address these limitations, the authors introduce the Gramian Chebyshev Neural Operator (GCNO), a physics-based, variable-rate compressor. Instead of treating channel data as arbitrary pixel grids, GCNO identifies a sample-dependent set of underlying propagation path directions by leveraging the inherent structure of receive-transmit channels: * Path Location: Utilizes the physical receive-transmit channel structure to locate dominant paths. * Direction Refinement: Employs a first-order Taylor correction to fine-tune directions that fall between discrete grid points. * Strength Recovery: Uses least squares to accurately recover complex path strengths.

为了解决这些限制,作者引入了格拉姆切比雪夫神经算子 (GCNO),这是一种基于物理的变速率压缩器。GCNO 不再将信道数据视为任意像素网格,而是通过利用收发信道固有的结构来识别一组依赖于样本的潜在传播路径方向: * 路径定位: 利用物理收发信道结构来定位主导路径。 * 方向细化: 采用一阶泰勒修正来微调落在离散网格点之间的方向。 * 强度恢复: 使用最小二乘法来精确恢复复数路径强度。

Trained entirely without explicit path labels, GCNO allows base stations to reconstruct channels analytically from transmitted path tuples rather than relying on a learned decoder. Across three distinct ray-traced environments, GCNO consistently outperforms neural feedback baselines by delivering higher reconstruction accuracy at equivalent payloads (or lower payloads at equal accuracy), while seamlessly transferring to unseen antenna counts without requiring retraining.

GCNO 完全在没有显式路径标签的情况下进行训练,允许基站从传输的路径元组中解析地重建信道,而不是依赖于学习到的解码器。在三种不同的射线追踪环境中,GCNO 通过在等效负载下提供更高的重建精度(或在相同精度下提供更低的负载),始终优于神经反馈基线,同时能够无缝迁移到未见过的天线数量,而无需重新训练。



🛠️ 相关媒体与资源 (Associated Media & Resources)