DFL-RUL: Decentralised Federated Learning for Battery Remaining Useful Life Estimation on Heterogeneous Edge-to-cloud
jaber pournazari, Mo’ath El-Dalahmeh, Dong-Hwan Park, James Marco, Truong Quang Dinh, Jung-Hoon Ahn, Mona Faraji Niri
Abstract
Accurate remaining useful life (RUL) prediction of lithium-ion batteries is essential for reliable and cost-effective electric vehicle operation, yet existing approaches largely rely on centralised training or overlook deployment constraints and data heterogeneity. This paper introduces DFL-RUL, a decentralised federated learning framework specifically designed to address feature-space inconsistency, temporal generalisation, and edge-level feasibility in real-world battery prognostics. Unlike prior federated RUL methods that assume aligned feature representations across clients, DFL-RUL integrates unsupervised, client-side PCA to automatically align heterogeneous sensor features before model aggregation. Local battery degradation is modelled using lightweight LSTM networks, while global knowledge is learned through FedAvg-based aggregation without sharing raw data. To reflect practical forecasting conditions, the framework is evaluated under a forward-in-time validation protocol, where only early-life cycles are available during training. Extensive experiments demonstrate that DFL-RUL achieves accuracy comparable to or exceeding local and centralised baselines, while significantly reducing communication cost and training latency. Moreover, runtime profiling on EV-class edge hardware confirms low inference latency and low energy consumption, validating the framework’s suitability for on-device deployment. These results show that reliable battery RUL estimation can be achieved in a privacy-preserving, hardware-aware, and temporally robust federated setting. • A decentralised federated learning framework (DFL-RUL) is proposed for accurate and privacy-preserving prediction of battery remaining useful life (RUL). • The framework combines PCA-based feature reduction, LSTM-based local modelling, and FedAvg aggregation within an edge-to-cloud setup. • Computational efficiency and runtime complexity are analysed, confirming real-time feasibility with sub-200 ms inference latency on EV-class hardware. • DFL-RUL achieves comparable or superior accuracy to centralised models while significantly reducing communication cost and training time. • The approach offers a scalable, hardware-aware foundation for intelligent battery health management in large EV fleets.