A Physics-Informed iTransformer and Kolmogorov–Arnold Network Hybrid Architecture for Lithium-Ion Battery State-of-Health Estimation

Accurate estimation of state of health (SOH) is essential for the safe and reliable operation of lithium-ion batteries. However, existing SOH methods face significant challenges, including dependence on complete charge–discharge cycles and a lack of physical constraints, which constrain both estimation accuracy and generalization performance. To overcome these limitations, a physics-informed iTransformer-KAN hybrid architecture is proposed in this paper for accurate and generalizable SOH estimation. The variable-centric tokenization mechanism inherent to iTransformer is employed to effectively capture heterogeneous temporal dependencies among multivariate battery health indicators. Furthermore, the conventional MLP projection layer is substituted with Kolmogorov–Arnold Network (KAN), which delivers nonlinear mapping performance and enhanced parameter efficiency via spline-based learnable activation functions. In addition, the targeted features are extracted from the second peak region of the Incremental Capacity (IC) curve, as this region demonstrates stronger sensitivity to underlying degradation mechanisms. The proposed model is further refined through hyperparameter optimization performed by the Cuckoo Catfish Optimizer (CCO) to ensure robust performance across different operating conditions. Three battery aging data sets validate that the SOH can be accurately estimated by the proposed model, with mean absolute error ranging from 0.21% to 0.55%, 0.27% to 0.55%, and 0.94% to 1.75%, respectively, outperforming mainstream methods.