Logo image
Back
A lithium battery state of health estimation approach based on SHapley Additive exPlanations-guided feature fusion and a diffusion model
Journal article   Peer reviewed

A lithium battery state of health estimation approach based on SHapley Additive exPlanations-guided feature fusion and a diffusion model

Ziheng Huang, Yu Zhang, Jiale Hou, Jie Hu, Cheng Chen, Weiheng Shih and Donghai Lin
Journal of energy storage, v 180, 124347
01 Dec 2026
DOI: https://doi.org/10.1016/j.est.2026.124347
Featured in Collection :   Drexel's Newest Publications

Abstract

Diffusion probabilistic model iTransformer-BiGRU Lithium-ion batteries Small-sample data State of health (SOH) estimation
Accurate and interpretable estimation of lithium-ion battery state of health (SOH) is essential for ensuring the safety and reliability of energy storage systems. However, existing data-driven approaches often suffer from degraded performance under small-sample conditions and provide limited insight into how predictions are made when processing raw sensor signals. To address these challenges, this paper proposes a novel end-to-end SOH estimation framework that synergistically combines a hybrid iTransformer–BiGRU (Bidirectional Gated Recurrent Unit) architecture, diffusion-based data augmentation, and SHAP (SHapley Additive exPlanations)-guided interpretability analysis. The model takes raw voltage, current, and temperature time-series sequences as direct input, bypassing manual feature engineering. The iTransformer module captures long-range temporal correlations across cycles, while the BiGRU network models local degradation dynamics within each cycle, enabling comprehensive representation learning. To mitigate severe data scarcity—especially in batteries with few aging cycles—a denoising diffusion probabilistic model is employed to synthesize realistic multivariate time-series trajectories, effectively expanding the training set while preserving underlying degradation patterns. The Kepler Optimization Algorithm (KOA) is further integrated to automatically tune key hyperparameters of the prediction network for optimal performance. To enhance transparency, SHAP are applied post-hoc to the trained model, quantifying the relative contribution of each extracted health feature to the final SOH prediction. This provides actionable insights into which degradation-indicative features (e.g., charging time, voltage increment, temperature rise slope) are most influential for health assessment. Experimental validation on the NASA battery dataset demonstrates that the proposed method achieves a mean absolute error (MAE) of 0.0014 (i.e., 0.14% of nominal capacity) and a root mean square error (RMSE) of 0.0018, outperforming several state-of-the-art benchmarks. Ablation studies confirm the effectiveness of both diffusion-based augmentation and the iTransformer–BiGRU backbone, while post-hoc SHAP analysis reveals consistent attention to physically plausible degradation signatures. This work delivers a high-accuracy, data-efficient solution equipped with post-hoc explainability for battery health monitoring under controlled small-sample conditions, as validated on the publicly available NASA 18650 LiCoO₂ dataset. •A KOA-iTransformer-BiGRU hybrid model is proposed for battery SOH estimation.•Diffusion-based data augmentation effectively addresses small-sample limitations.•SHAP analysis quantifies feature contributions and enhances model interpretability.•The proposed method achieves MAE as low as 0.14% on the NASA dataset.•The framework delivers accurate, robust, and interpretable SOH estimation.

Metrics

1 Record Views

Details

Logo image