Logo: to the web site of Uppsala University

Federated learning enables decentralized model training across multiple clients while preserving data privacy. However, standard federated learning algorithms often perform poorly when client data is non-identically distributed, as is common in real-world applications such as electric vehicle battery monitoring. Clustered federated learning methods, such as the Adaptive Iterative Clustered Federated Learning (AICFL) scheme, aim to improve performance in such settings by grouping similar clients and training separate models per cluster. This thesis presents a proof-of-concept implementation of AICFL scheme within the FEDn framework. The solution includes custom logic for state progression, cluster identity estimation, and dynamic aggregation, all adapted to FEDn’s strict one-round-per-phase communication constraint. Clients are simulated in isolated Docker containers and operate on disjoint subsets of real-world lithium-ion battery data from NASA and the system is evaluated on the task of battery State-of-Health estimation. Results show that meaningful client clusters emerge under data heterogeneity and that the FEDn-based implementation can carry out multi-phase AICFL scheme rounds effectively. While runtime efficiency and full cluster adaptation are not implemented, the work demonstrates the feasibility of deploying clustered federated learning logic FEDn.