面向边缘智能的通信计算一体化研究

摘要：为了进一步提高无线数据聚合效率，空中计算技术通过利用无线信道波形叠加特性允许模型更新信息在空中“一次性”完成聚合，实现通信网与算力网的“网媒融合”。然而在这过程中，信道衰落和噪声可能会带来聚合失真。此外，更新数据的质量以及边缘设备的传输能耗也可能影响模型聚合以及收敛效率。为此提出了基于空中计算的联邦学习系统，并针对其存在的信道干扰、高效数据传输和数据失真问题建立动态设备调度机制，在满足接收端信噪比条件下选择适当数量质量较高的设备参与模型训练。该机制利用梯度重要性、信道条件和传输能耗衡量设备质量并保留累积未被选择设备的梯度以加速收敛。基于李雅普诺夫优化理论进行问题建模和求解，仿真结果表明该机制具有较高训练精度和较快收敛速度，同时针对不同噪声功率具有一定鲁棒性。

关键词：空中计算；联邦学习；设备调度；设备质量；鲁棒性

Abstract: Over-the-air computation (AirComp) technology leverages the waveform superposition characteristics of wireless channels to further enhance the efficiency of wireless data aggregation, enabling model update information to be aggregated “in one shot”. This achieves a convergence of communication networks and computational power networks, exemplifying the concept of “network and computation fusion”. However, channel fading and noise may introduce aggregation distortion during this process. Additionally, the quality of update information and the transmission energy consumption of edge devices can impact model aggregation and convergence efficiency. Therefore, we establish an AirComp enabled federated learning system and propose a dynamic device scheduling mechanism to address issues related to channel interference, efficient data transmission, and data distortion. Specifically, an appropriate number of higher-quality devices are selected to participate in model training while satisfying receiving signal-to-noise ratio conditions. It utilizes gradient importance, channel conditions, and transmission energy consumption to assess device quality and retains and accumulates gradients from unselected devices to accelerate convergence. The problem is modeled and solved based on the Lyapunov optimization theory. Simulation results demonstrate that this mechanism achieves higher training accuracy, faster convergence speed, and a certain level of robustness against varying noise power levels.

Keywords: over-the-air computation; federated learning; device scheduling; device quality; robustness