Semi-supervised battery state of health estimation for field applications

Abstract

Battery electric vehicles are exposed to highly diverse operating conditions and driving behaviors that strongly influence degradation pathways, yet these real-world complexities are only partially captured in laboratory aging tests. This study investigates a semi-supervised learning approach for robust estimation of battery state of health, defined as the ratio of remaining to nominal capacity. The method integrates a multi-view co-training algorithm with a rule-based pseudo labeling mechanism and is developed and validated using field data from 3000 BMW i3 vehicles with battery capacity of 60 Ah, collected since 2013 across 34 countries. The available data comprises standardized full charge capacity measurements, which serve as ground truth labels. The proposed training and validation pipeline is designed to address challenges inherent in real-world data generation and is particularly advantageous during early deployment of new battery technologies, when labeled data is scarce. By incrementally incorporating newly available labeled data into both evaluation and retraining, the model adapts to heterogeneous aging patterns observed in the field. Comparative analysis demonstrates that, relative to a supervised benchmark, the proposed method reduces estimation error by 28 % under limited-label conditions and by 6 % under optimally labeled scenarios, highlighting its robustness for field applications.