FAME-SOH: A Feature-Augmented Multi-Expert Ensemble Framework for Robust Online Estimation of Real-World EV Battery SOH Using Only Voltage and Current Data

To address the common challenges in practical applications of online State of Health (SOH) estimation for power batteries, such as randomness, limited dataset size, and performance constraints, this study proposes a Feature-Augmented Multi-Expert Ensemble framework for SOH estimation (FAME-SOH). The proposed method relies solely on voltage and current data, enabling highly accurate and robust SOH estimation for real-world electric vehicle (EV) batteries. FAME-SOH first leverages voltage sequence segment information to construct coverage indicators, thereby avoiding the limitations of relying on a single manually extracted feature. A segmentation and random-missing data augmentation strategy is then employed to expand the dataset and enhance generalization under limited-data conditions. Subsequently, mask-based Transformer expert models are trained for different charging segments to adapt to diverse operating conditions. Then, a Similarity-Aware Feature Augmentation (SAFA) module is introduced to generate augmented features based on the matching degree between the input data and each expert’s training set. Finally, a Multi-Layer Perceptron (MLP) aggregator is used to learn dynamic weight allocation rules among experts, mitigating the influence of unsuitable experts under certain scenarios. Multi-condition tests on real-world EV operational data demonstrate that FAME-SOH achieves an R² of 0.924 (RMSE: 0.83%) for SOH estimation under incomplete charging curve scenarios, outperforming benchmark methods in both accuracy and robustness, and offering a new technical pathway for the intelligent upgrading and online health management of EV battery management systems.