Ultra-fast battery health estimation from random 100 mV charging fragments using conditional generative diffusion

Abstract

State-of-health (SOH) estimation is essential for lithium-ion battery management. However, real-world disruptions often result in fragmented battery data, making reliable SOH estimation challenging. Generative artificial intelligence offers a promising paradigm to address these limitations. Here, we present a conditional diffusion framework that generates complete battery data from random charging fragments spanning 100 mV, including fragments as short as 20 mV, to estimate SOH. Validated on 89 batteries across three cathode chemistries and nine protocols, the framework delivers stable, high-fidelity generation that outperforms generative adversarial network (GAN) and variational autoencoder (VAE) baselines. It achieves SOH estimation errors as low as 0.18% mean absolute error (MAE) and demonstrates higher estimation accuracy and robustness than the evaluated baseline methods under diverse missing-data conditions. In most fast-charging scenarios, the framework requires only charging fragments shorter than 30 s to estimate SOH. These results demonstrate that conditional diffusion provides a practical solution for battery health estimation under data-limited conditions.