Robust Lithium-ion Battery SOH Estimation via Denoising Autoencoder-CNN-Transformer with Noisy Multi-Feature Input

Accurate estimation of the State of Health (SOH) of lithium-ion batteries is challenged by sensor noise in multi-source operational data. Existing recurrent neural network (RNN)-based approaches can also suffer from computational bottlenecks. To address these issues, we propose a DAE-CNN - Transformer framework. A Denoising Autoencoder (DAE) is employed to encode noisy input features into compact, noise-robust latent representations. These encoded features are then processed by a CNN-based attention module that adaptively emphasizes the most informative components. Finally, a Transformer network models long-range temporal dependencies to predict SOH accurately. Evaluated on NASA and CALCE datasets with added Gaussian noise, the proposed model significantly outperformed baseline Transformer and DAE- Transformer architectures in terms of RMSE, MAE, and R2, demonstrating superior accuracy and robustness” particularly under high noise conditions.