车辆边缘计算下的联邦学习设备选择及聚合优化方法OA

Due to the high mobility of vehicles and their limited computational resources,it is challenging to parallelize the updating and aggregation of federated learning models across all vehicles in vehicle edge computing scenarios,which directly impacts the convergence and training speed of federated learning.To improve the convergence rate and training efficiency of federated learning in vehicle edge computing,a multi-objective distributed federated learning device selection and aggregation optimization method is proposed that dynamically adapts to high-speed vehicle mobility.Firstly,by evaluating vehicle communication link stability and the residence time in the coverage areas,a mobility-aware prescreening algorithm is designed to select qualified vehicle candidates,ensuring the stability of participating devices.Subsequently,a device selection algorithm is designed by combining deep reinforcement learning dual-deep Q-network to identify and select optimal clients from the pool of candidates to participate in federated learning.In addition,a momentum clustering optimization algorithm based on a three-tier federated learning architecture is proposed to address model instability and slow convergence caused by data heterogeneity.Experimental results show that the proposed algorithm outperforms traditional federated learning methods in terms of model accuracy,processing time,and communication overhead.It effectively leverages the computational resources of vehicle edge devices to enhance both the training speed and accuracy of federated learning.