Peer-to-Peer Personalized Federated Transfer Learning for Battery State of Health Estimation of Vehicles

The enhancement of battery reliability and safety during operational usage for electric vehicles has been a primary focus of state-of-the-art research, leading to the development of various data-driven technologies for battery state-of-health (SOH) estimation. Nevertheless, these approaches, rooted in traditional centralized computing paradigm, grapple with conflicts related to data needs versus privacy protection, robustness versus personalization, and transferability versus accuracy. To address these challenges, this study proposes a novel decentralized federated transfer learning (FTL) method, named P2P-PerFTL, which aggregates local SOH estimation models into a global model using peer-to-peer communication while preserving battery data locally, combining federated learning and transfer learning to realize personalization and transferability within various working conditions and batteries. This algorithm utilizes a domain-shift-based weighted aggregation mechanism for global model formulation, and constructs a personalized and transferable neural network architecture by assigning part layers as personalization layers and incorporating domain shift loss. Through a comprehensive case study with four FTL scenarios, it is demonstrated that the proposed P2P-PerFTL can train a highly proficient SOH estimation model with a limited volume of data from local clients across diverse operational conditions and battery types, and outperforms alternative training frameworks.