Health evaluation for in-service electric vehicle battery pack with recursive Gaussian process

Accurate and reliable state-of-health (SOH) estimation of lithium-ion batteries plays a vital role in the operations and maintenance of electric vehicles. Despite numerous SOH estimation methods developed under lab conditions, implementing these methods in real-world applications can be significantly challenging. In particular, the wide range of usage scenarios and lower data quality bring challenges to the feature engineering, model development, and online updating. The large-scale, continuously updated on-road electric vehicle (EV) data demands estimation models with stronger generalization capabilities and higher update efficiency. To address these issues, this article proposes a novel SOH estimation framework for in-service EVs based on feature extraction and recursive Gaussian process regression. First, capacity labels and corresponding features are extracted from field datasets covering two different battery chemistries. An interpretable selection method is used to select those key features with high contribution while reducing feature redundancy. Subsequently, a recursive Gaussian Process framework combining offline modeling with online updating is proposed for reliable and accurate SOH estimation. Two different kernel functions are used to model stationary and nonstationary dynamical processes. The Matérn kernel is employed to capture locally stationary degradation dynamics with adjustable smoothness, while the Wiener kernel models non-stationary cumulative aging trends, achieving superior accuracy on field data from EVs. The proposed framework is validated against field data with two types of batteries, achieving MAPEs of only 1.76 % and 2.71 % on independent validation sets, demonstrating its superior accuracy and effectiveness compared with conventional Gaussian Process and neural network-based transfer learning approaches, which highlights its potential for large-scale deployment in on-road EVs.