State of Health Estimation for Lithium-ion Batteries via Spatial-Bidirectional Cross-Attention and Top-k Sparse Transformer

Accurate esimation of the State of Health (SOH) for lithium-ion batteries is critical for their safe operation. Existing data-driven methods struggle with multi-parameter coupling, capacity fluctuations, and noise interference in real-world scenarios, urgently requiring models with the ability to perceive localand global information collaboratively. To address these challenges, this paper proposes a novel SOH estimation framework integrating Spatial Bidirectional Cross-Attention (SBCA) and a Topk Sparse Transformer. Firstly, we reconstruct time-series data into a 2D feature matrix to enhance interactions among multiple variables. Secondly, the SBCA module is innovatively employed to enable the model track how each parameter changes over time on height direction and spot how different parameters affect each other on width direction. Finally, we use the Topk Sparse Transformer to make model only focuse on the most important parts of the data. It can dynamically select attention weights that dominate capacity degradation, effectively reducing redundant information. Experiments on the NASA and MIT datasets show that, even training with only 30 % of the data, the method demonstrates a Root Mean Square Error as low as 0.0042 Ah., significantly outperforming existing approaches. This makes it great for real-world battery health estimation where data is hard to get.