Enhancing variational autoencoder for estimation of lithium-ion batteries State-of-Health using impedance data

State of health (SOH) estimation is critical for the safe and efficient operation of lithium-ion batteries (LIBs). Electrochemical impedance spectroscopy (EIS) encodes rich degradation information but is often affected by noise and exhibits limited generalization across battery types and operating conditions. To address these limitations, this study integrates variational mode decomposition (VMD) with a variational autoencoder (VAE) to construct the enhancing variational autoencoder (EVAE) model for effective feature extraction and subsequently maps the denoised latent features to SOH using a Transformer network. To further enhance adaptability, transfer learning is incorporated to enable the model to generalize across different chemistries, temperatures, and state of charge (SOC). Comprehensive experimental validation on both coin cells and 18650 commercial cells demonstrates that the proposed EVAE-Transformer consistently achieves MAE ≤3.65 % and RMSE ≤5.86 % across all tested conditions. At 25 °C, the method delivers particularly strong performance, with R2 ≥ 0.96, MAE as low as 1.31 %, and RMSE reduced to 1.61 %. Even at elevated temperatures of 45 °C, the best cases maintain high accuracy (MAE = 0.41 %, RMSE = 0.56 %), underscoring the robustness of the framework. In comparative studies, the EVAE-Transformer substantially outperforms a Transformer trained directly on raw EIS data (RMSE ≥7.24 %) and an EVAE–LSTM variant, confirming the effectiveness of the VMD-VAE feature extraction pipeline combined with the attention mechanism. By reducing dependence on extensive target-domain datasets and eliminating complex preprocessing, the EVAE-Transformer provides a practical, scalable, and generalizable solution for next-generation battery management systems.