Hybrid end-to-end battery modeling and SOH estimation via physics-data fusion and maximum mean discrepancy minimization

Lithium-ion batteries are pivotal to enabling energy transition and vehicle electrification. Accurate state-of-health (SOH) estimation is essential to ensure battery safety, reliability, and cost-effectiveness. However, existing physical model-based and data-driven SOH estimation methods encounter challenges such as difficulty in modeling complex multi-scale degradation phenomena and poor adaptability to variable usage conditions and cell chemistries. To overcome these barriers, this paper proposes an end-to-end hybrid modeling framework that integrates physical insights with data-driven learning through a physics-based and data-driven fusion network, self-supervised learning, and transfer learning. This approach bypasses explicit degradation modeling and dynamically updates aging-related parameters, preserving comprehensive aging features without the need for explicit SOH labels during training. Furthermore, domain adaptation based on maximum mean discrepancy minimization is employed to align aging representation distributions across different batteries and usage profiles. This consistency regularization approach enhances SOH estimation in the target domain, achieving high generalization across different battery chemistries and operating conditions.