Learning-Based Task-Centric Multi-User Semantic Communication Solution for Vehicle Networks

Abstract

With the application of 5G in in-vehicle network scenarios, the inevitable crisis of scarcity of communication resources is deepening. At the same time, considering the high demands for communication delay in vehicular network scenarios, it is essential to meet the requirements for reliable information transmission in high-speed mobility environments. To address these issues, we propose the Task-Centric Multi-User Semantic Communication (TCMSC) solution, designed to meet the energy consumption and transmission time delay requirements in line with the new paradigm of semantic communication. Our solution introduces a task-centric semantic processing model aimed at improving Semantic Spectral Efficiency (S-SE). The TCMSC solution is tailored for multi-service vehicle scenarios, optimizing power consumption and enhancing reliability, making it well-suited for challenging environments. Moreover, we propose a novel method using Stochastic Network Calculus (SNC) to accurately model semantic task delay and calculate the upper bound of Vehicle-to-Infrastructure (V2I) delay-bound violation probability. However, to tackle the increased optimization complexity from SNC while simultaneously enhancing feature extraction, we propose the Transformer Advantage Actor Critic (TR-A2C) algorithm. This algorithm leverages the Transformer to capture dynamic vehicle parameters across scenarios, accelerating the TCMSC solution. Experimental results demonstrate that, compared to traditional single-service vehicle dispatch, TCMSC improves delay violation probability by ${2.88\%}$ and reduces power consumption by ${31.1\%}$, all while effectively enhancing S-SE in complex traffic environments.