A novel SOH estimation method of sodium-ion batteries based on multi-channel threshold residual network

Sodium-ion batteries (SiBs) have been widely studied in the field of energy storage due to their abundant resources and high safety. However, their state-of-health (SOH) estimation is not straightforward, due to the complex aging mechanisms and dynamic working conditions. This study proposes an SOH estimation framework based on Multi-channel Threshold Residual Network (MTRN), which combines multi-modal feature selection and threshold selection techniques. The multi-modal feature selection framework is based on an optimization strategy which consists of 3 stages: mutual information filtering, principal component dimensionality reduction, and dynamic adaptive lasso regression. It allows to extract the high contributing health factors from 28 original features and reduces 85 % of the feature dimensions while retaining high correlation features, which solves the problems of feature redundancy and nonlinear correlation. The MTRN architecture incorporates a multi-channel attention mechanism to dynamically assign key information, applies KAN to learn univariate basis functions in order to fit nonlinear degradation, and establishes a threshold residual shrinkage module to distinguish between noise and real degradation trends. On the Dataset A/B, which is a self-built SiB dataset, the RMSE, MAE, and MAXE of MTRN are reduced by 40.28–60.56 % compared with those of the TCN and KAN models. Under extreme noise conditions of 150 mV, the increase of MAE is controlled within 0.85 %. On the Dataset C/D, the MAE values are respectively 0.62 % and 0.73 %, which verifies the high adaptability of the proposed model to the differences in chemical systems. This study provides a high-precision and high-robustness solution for the SOH estimation of SiBs.