Federated Learning for Connected and Automated Vehicles

The subject of this dissertation is the development of Machine Learning (ML) algorithms and their applications in Connected and Automated Vehicles (CAV). Decentralized machine learning algorithms have been proposed for CAV with noisy communication channels. Decentralized ML algorithms, referred to as Federated Learning (FL), are developed for multiple vehicles to collaboratively train models, thus enhancing performance while ensuring data privacy and security. Applications of FL for CAV (FL4CAV) are analyzed. Both centralized and decentralized FL frameworks are considered, along with various data sources, models, and data security techniques relevant to FL in CAVs. Three innovative algorithms for Decentralized Federated Learning (DFL) that effectively handle noisy communication channels are proposed. Theoretical and experimental results demonstrate that the proposed algorithms that share gradients through noisy channels instead of parameters are more robust under noisy conditions compared to parameter-mixing algorithms. Building on the exploration of decentralized federated learning, a novel decentralized noisy model update tracking algorithm is proposed to further enhance robustness and efficiency while addressing the challenge of data heterogeneity impact. The proposed algorithm performs better than the existing algorithms in handling imperfect information sharing. Expanding on these findings, applications of FL are proposed for heavy-duty diesel engines, which remain crucial due to their fuel efficiency and emissions characteristics. Finally, an FL algorithm to better predict and control aftertreatment temperature overshoots in real-time is demonstrated.