Cross-domain state of health estimation for lithium-ion battery based on latent space consistency using few-unlabeled data

Accurate state of health (SOH) estimation is crucial for battery safety and reliability. Traditional SOH estimation methods are designed based on extensive labeled data for specific chemical compositions or cycling conditions. However, batteries in real-world applications often operate with unlabeled degradation profiles that make SOH estimation more complex. To tackle this issue, we propose a novel domain adaptation framework for SOH estimation via an autoencoder combined with adversarial learning based on the aging data from only one cell. The autoencoder is self-supervised and used to extract the hidden health features and the adversarial learning is used to align these features across domains, enabling domain-invariant feature extraction. Further, a novel SOH estimation method is presented by the similarity comparison of hidden health features from the source domain and target domain, where the traditional regression is avoided. Tests on both self-collected and public datasets show our framework accurately estimates SOH across different compositions, cycling rates, and temperatures, with RMSE under 1.64% for various cycling rates and temperatures, and 2.03% for different chemical compositions. This study offers an efficient solution to cross-domain SOH estimation, significantly reducing the economic and time costs associated with labeled data collection.