Multimodal deep learning with time-frequency health features for battery SOH and RUL prediction

This study proposes a multimodal deep learning framework for joint prediction of the state of health (SOH) and remaining useful life (RUL) of lithium-ion batteries. Twelve representative impedance features—covering charge-transfer resistance, solid electrolyte interface (SEI) layer impedance, and ion diffusion—are extracted from electrochemical impedance spectroscopy (EIS) and combined with short voltage/current segments to form a compact, interpretable feature set. A residual multi-layer perceptron (ResMLP) is employed for SOH regression, and a temporal convolutional network with attention (TCN-Attention) is used for RUL estimation. Lifetime experiments on two battery types with different chemistries and form factors, evaluated through three rounds of paired cross-validation, validate the approach. Results show that the proposed features significantly reduce dimensionality and computational cost while substantially lowering SOH error, achieving an average normalized root mean square error of 2.3 %. The RUL prediction reaches an average error of 14.8 %. Overall, the framework balances interpretability, robustness, and feasibility, providing a practical solution for battery management systems (BMS) monitoring and life prediction.