Robust State of Health Estimation for Heterogeneous Batteries With Privacy Preserving

Abstract

The state-of-the-art approaches to state of health (SOH) estimation typically generate models tailored to specific battery datasets, requiring retraining for other battery types and failing to construct a universally robust model across diverse battery data. Challenges in achieving a robust model span statistical heterogeneity, adaptability, and resilience to noise and cyber-attacks. This study introduces a novel, privacy-preserving robust SOH estimation for heterogeneous batteries using customized federated learning (FL). The approach aggregates local SOH models into a global model while preserving data privacy at the source. It leverages statistical and system utility indicators for battery management system client selection, incorporating the Epsilon-Greedy method to dynamically include new clients and adjust the participant count based on global model efficacy. Additionally, a performance-discrepancy weighted aggregation mechanism is designed by using the L2 distance and loss indices as weights for local models. A discrepancy-based personalization method further refines the number of personalization layers, enhancing local model performance. Through detailed case analysis, the proposed algorithm is shown to surpass conventional FL, centralized, and local training methods across scenarios of varied battery types, data inaccuracies, communication errors, and operational risks, demonstrating superior adaptability and robustness.