面向边端协同推理的模型分割与无线资源分配方法研究OA

For multi-user device-edge collaborative inference in the 6G intelligent air interface,fixed model partitioning and static spectrum allocation are unable to effectively accommodate heterogeneous model loads and inter-user resource competition.In this paper,we first develop a device-edge collaborative inference model that accounts for spectrum overlapping and co-channel interference.Then model partitioning and wireless spectrum allocation are jointly formulated as an optimization problem aimed at minimizing the total system latency,and accordingly we propose an interference-aware action-decomposed proximal policy optimization(IAAD-PPO)algorithm.The proposed algorithm employs three conditional actions to determine the model partition point,the starting position of the allocated spectrum,and the spectrum occupancy length,respectively.Action masking is introduced to ensure the validity of the selected actions,while a system-level reward is designed to coordinate resource competition among multiple users.A device-edge collaborative inference testbed is implemented using AlexNet,MobileNetV2,ResNet50,and VGG16.Experimental results show that,in a four-user concurrent inference scenario,the average end-to-end latency of IAAD-PPO is 55.5 ms,representing reductions of 13.8％,67.8％,and 60.6％compared with DQN,local execution,and full edge execution,respectively.