摘要
Edge intelligence empowered by 6G is a promising technique for realizing AIoT applications. However, making intelligent decisions through distributed training across various mobile devices, considering model accuracy, system heterogeneity, and training latency, is still an open challenge. In this paper, we propose a United Split Federated Learning (U-SFL), a two-level collaboration framework with a "first-sequential-then-parallel"principle that integrates sequential split learning within clusters and parallel training across clusters to achieve fine-grained and coarse-grained knowledge sharing jointly. Furthermore, we design an evolutionary game-based clustering approach that enables devices to identify the environmental fitness of the cluster and self-organize themselves into a suitable cluster with a theoretical evolutionary equilibrium proof. By integrating the U-SFL framework with evolutionary game-based clustering, we effectively mitigate the overhead of sequential training, achieving faster convergence with fewer clusters and reversing the conventional belief that only higher parallelism can reduce training latency. The experimental results show that the U-SFL framework outperforms the SOTA edge intelligence frameworks and preserves optimized training latency in a large-scale heterogeneous split federated learning system. © 2025 IEEE.