New AI Method Predicts When Batteries Will Fail—Even With Bad Data
Researchers have developed a machine learning system that can accurately estimate battery lifespan and degradation despite missing or corrupted sensor data—a critical capability for electric vehicles and grid storage systems. The method maintains accuracy even when half the data is missing, potentially reducing costly premature battery replacements and improving the economics of EV manufacturing and energy infrastructure.
Originaltitel: A Robust Physics-Informed Data-Driven Method for Li-Ion Battery Degradation Estimation
<p>The estimations of state of health and remaining useful life (RUL) are essential for battery health management, closely linked to current and future capacity estimations. However, limited research on high-accuracy capacity estimation across varying horizons constrains RUL prediction accuracy. Additionally, models are vulnerable to disturbances like missing data and noise, further reducing reliability. To address these challenges, this paper proposes a robust physics-informed battery degradation estimation method. The feature-wise neural controlled differential equations (CDE) generate attention-weighted features to capture critical battery characteristics, while the learning-wise neural CDE models differential relationships. The empirical model further refines the degradation trend. Experimental results demonstrate the model's superior performance across different estimation horizons, missing data, and noise scenarios compared to existing methods. Notably, even with 50 % missing data and a 100 -cycle horizon, the proposed physics-informed data-driven method can achieve an estimation error of 0.0156.</p>