Accurate estimation of the State of Health (SOH) of lithium-ion batteries (LiBs) is crucial for performance optimization and safety. However, traditional health features (HFs) extraction methods often fail to capture key HFs in the complex charging data of LiBs, resulting in low model accuracy. The primary contribution of this study is that we successfully derived the HFs of LiBs directly from the external information of the original charging data using a semi supervised architecture and achieved estimation of SOH. To address this issue, we propose a semi-supervised learning approach utilizing Variational Autoencoders (VAE) with Gated Recurrent Unit (GRU). The VAE encoder extracts HFs from unlabeled charging data, enhancing the understanding of data patterns. These HFs are then combined with labeled SOH data to train downstream regression networks for estimation. Experimental results demonstrate that our method performs excellently in practical electric vehicles (EVs), with a maximum Mean Absolute Percentage Error (MAPE) not exceeding 5%. This VAE based semi supervised technique provides a promising solution for SOH estimation of LiBs, demonstrating excellent accuracy and robustness.