Capacity estimation of lithium-ion battery through interpretation of electrochemical impedance spectroscopy combined with machine learning

Abstract

Lithium-ion batteries inevitably undergo degradation over extended use, making precise capacity estimation essential for reliable state monitoring and health prognostics. However, conventional capacity estimation methods, which rely predominantly on time-domain information, offer limited insights due to the complex degradation mechanisms intrinsic to lithium-ion batteries. Accordingly, a capacity estimation approach leveraging partial frequency electrochemical impedance spectroscopy data integrated with machine learning is proposed. A reconstructed simplified fractional-order model characterized by a minimal set of parameters and superior fitting performance is introduced to extra health indicators from EIS measurement. Detailed equivalent circuit parameter extraction is performed, and subsequently incorporated into an enhanced Gaussian process regression model for capacity estimation. The proposed approach is validated using datasets derived from real-driving profiles including comprehensive comparisons, demonstrating its feasibility and effectiveness. Experimental results confirm that the framework achieves high estimation accuracy, strong generalization, and robust performance, even in scenarios with limited data availability and training noise presence, with a root mean square error not exceeding 0.81 %, a mean absolute error below 0.62 %, and a maximum absolute error of no more than 0.0653 Ah.