A Multi-Encoder BHTP Autoencoder for Robust Lithium Battery SOH Prediction Under Small-Sample Scenarios

Accurate prediction of the state of health (SOH) in lithium batteries (LiBs) is essential for ensuring operational safety, extending battery lifespan, and enabling effective second-life applications. However, achieving precise SOH prediction under small-sample conditions remains a significant challenge due to inherent variability among battery cells and capacity recovery phenomena, which result in irregular degradation patterns and hinder effective feature extraction. To overcome these challenges, this study introduces an advanced autoencoder-based method specifically designed for SOH prediction in small-sample scenarios. This method employs a multi-encoder structure—comprising token, positional, and temporal encoders—to comprehensively capture the multi-dimensional characteristics of SOH sequences. Furthermore, a BHTP module is integrated to facilitate feature fusion and enhance the model’s stability and interpretability. By utilizing a pre-training and fine-tuning strategy, the proposed method effectively reduces computational complexity and the number of model parameters while maintaining high prediction accuracy. The validation of the NASA 18650 lithium cobalt oxide battery dataset under various discharge strategies shows that the proposed method achieves fast convergence and outperforms traditional prediction methods. Compared with other models, our method reduces the RMSE by 0.004 and the MAE by 0.003 on average. In addition, ablation experiments show that the addition of the multi-encoder structure and the BHTP module improves the RMSE and MAE by 0.008 and 0.007 on average, respectively. These results highlight the robustness and utility of the proposed method in real battery management systems, especially under data-scarce conditions.