Enhanced state of health prediction for lithium-ion batteries using a hybrid convolutional–Kolmogorov–Arnold network

Reliable estimation of the state of health (SOH) of lithium-ion batteries (LIBs) is essential to ensure the continuous and safe operation of energy-dependent systems. Traditional methods for LIB SOH estimation using neural networks often lacked the flexibility required to handle high-dimensional data, resulting in restricted accuracy. To address this issue, this study proposes a hybrid convolutional neural network (CNN)–Kolmogorov–Arnold network (KAN) model to provide a more accurate estimation of LIB SOH. With its learnable activation functions, KANs overcome the inflexibility challenges of traditional neural networks and have the potential to replace conventional multilayer perceptrons (MLP). In this study, five indirect health features (IHFs) were extracted from the charge–discharge voltage and capacity enhancement curves. To address the issue of weak informative content in IHFs, we try to combine the CNN’s powerful high-dimensional feature extraction capabilities to optimize the IHFs, thereby improving the prediction performance of the KAN. Convolutional layers are employed to extract local features from the IHFs, whereas pooling layers further reduce their dimensionality while preserving crucial information. This optimization enhances the feature set and significantly improves prediction accuracy. Finally, the optimized features are fed into the KAN model for SOH estimation. The SHAP technique was applied to calculate the contribution of each IHF to address the black-box nature and lack of interpretability inherent in data-driven models. This study uses publicly available datasets provided by CALCE, Oxford, as well as the Toyota Research Institute and the Massachusetts Institute of Technology (MIT) for algorithm validation. The experimental results demonstrate that the CNN–KAN model performs excellently, with an root mean squared error (RMSE) less than 0.007, mean absolute error (MAE) less than 0.005, and mean absolute percentage error (MAPE) below 0.006. Compared with the errors in other studies, the results highlight the strong predictive accuracy and reliability of the CNN-KAN model.