A novel uncertainty-inclusive SOH estimation model for lithium-ion batteries based on KMCGAN

Uncertainties in lithium-ion batteries pose a challenge for accurate state of health (SOH) estimation. In this work, we propose a novel uncertainty-inclusive SOH estimation model including Kolmogorov-Arnold multi-head attention (KMHA) and constraints generative adversarial network (CGAN) to enhance the accuracy and reliability of estimation, which is characterized by conditional constraint, distinctive KMHA architecture, and uncertainty estimation. Voltage, current, and temperature are key factors affecting SOH, are regarded as constraints to guide the generation process, thereby improving the training stability. The KMHA architecture is deigned to capture the complex dependencies among voltage, current, and temperature, enhancing the model's understanding of the relationships among these influencing factors. Diverse results are generated through sampling to achieve uncertainty estimation. Where multilayer perceptron is replaced by KAN, which reduces computation and improves interpretability. Validations across diverse battery datasets from NASA and Oxford demonstrate that KMCGAN outperforms existing models in competitive performance. Additionally, ablation studies confirm the positive contribution of each component to the complete model. In this case, the average mean absolute error for all cells is 0.6031, the average coefficient of determination is 0.9906, and the average uncertainty is 0.8200, proving the potential of KMCGAN in terms of accuracy and uncertainty.