Abstract:

With the development of perception and intelligent applications in connected vehicles, vehicle terminals continue to generate vast amounts of local data. Federated learning provides an effective approach for conducting multi-vehicle collaborative training whilst safeguarding data privacy, whilst the low-latency, high-reliability communication and edge computing capabilities of the 6th Generation Mobile Communication Technology (6G) networks further provide the network and computational power required for large-scale vehicle collaborative learning. However, traditional federated learning is typically designed for time-invariant data distributions and struggles to adapt to the dynamic environment characterized by vehicle movement and constantly changing data acquisition conditions. To address this, we propose A Data and Resource Co-sensing Federated Continual Learning Algorithm tailored for mobile connected vehicles. By optimizing terminal scheduling, the algorithm enhances the participation of data from new categories, whilst drift-aware adaptive aggregation dynamically adjusts the update step size of the global model. Experimental results demonstrate that the proposed algorithm accelerates the recovery of model performance following concept drift, achieving a final accuracy approximately 2.5 per cent higher than that of bandwidth-priority scheduling and 8.2 per cent higher than that of a fixed learning rate scheme.

Key words: 6G, federated continual learning, mobile connected vehicles, concept drift, device scheduling