Abstract
Ensuring reliable and efficient transportation in Cyber-Physical Systems (CPS) requires effective model training across distributed Consumer Internet of Vehicles (CIOV)-based Transportation CPS (TCPS). However, the high mobility of transportation terminals and frequent domain switching during training degrade global model accuracy, while malicious terminals uploading erroneous data further compromise system reliability. To address these challenges, this paper proposes Fed-ECC, a two-tier federated learning (FL)-based edge collaborative computing mechanism for dynamic CIOV-based TCPS. The first tier employs a deep reinforcement learning (DRL)-based clustering algorithm to form edge collaborative computing domains, optimizing terminal selection based on mobility, computational capability, and reliability. The second tier integrates a semi-asynchronous local aggregation mechanism with adaptive aggregation factors and an asynchronous regional aggregation mechanism based on data volume, improving aggregation efficiency and model convergence. Simulation results demonstrate that Fed-ECC enhances global model accuracy by 58.7%, accelerates convergence speed by 57.6%, and achieves 95% accuracy in traffic safety tasks, significantly improving obstacle detection and service reliability. These findings underscore the effectiveness, scalability, and robustness of Fed-ECC in addressing the challenges of high-mobility, large-scale CIOV-based TCPS.
Funding Information
  • National Research Foundation of Korea (NRF) grant funded by the Korea government (RS-2024-00423362)
  • National Research Council of Science and Technology (NST) grant by the Korea government (GTL24031-800)

This publication has 37 references indexed in Scilit: