Revealing State-Feature Dependencies in Dynamic Degradation: An Exofeature-Aware Transformer for Battery State-of-Health Prediction
Ruyi Huang, Yan Chen, Cheng Liu
Abstract
Reliable prediction of lithium-ion battery state-of-health (SOH) is essential for battery management systems in Consumer Electronics (CE). To overcome the limitations of existing models, which often underutilize heterogeneous state–feature dependencies by treating multivariate inputs as a unified time series, an Exofeature-Aware Transformer (EFATransformer) is proposed for long-horizon SOH prediction. Within the EFATransformer, two tailored embedding strategies are designed: a channel-wise embedder for multivariate statistical features that bridges temporal dynamics and channel-specific characteristics, and a patch-wise embedder for the SOH sequence, augmented with a learnable global representation to bridge feature-to-SOH correlations. These representations are jointly modeled through a stack of Collaborative Attention Mechanism (CAM) blocks that jointly leverage Historical SOH Self-Attention (HSSA) to capture intra-temporal dynamics and global–local interactions, and Feature-to-SOH Cross-Attention (FSCA) to identify degradation-relevant features while enhancing model interpretability by highlighting channel contributions. Nested leave-one-out cross-validation is conducted on the public dataset from Xi’an Jiaotong University. The EFATransformer outperforms six state-of-the-art methods across distinct charging/discharging protocols. Visualizing FSCA attention weights provides interpretable insights into state-feature dependencies in battery degradation, transforming post-hoc interpretability into battery health management. These attributes align well with Machine Learning for CE topics and show strong potential for deployment in safety-critical consumer devices.