BMSFormer: An efficient deep learning model for online state-of-health estimation of lithium-ion batteries under high-frequency early SOC data with strong correlated single health indicator

The efficient and accurate state-of-health (SOH) estimation is crucial for reducing risks and ensuring effective application in battery management systems (BMS) of resource-limited devices. However, many recent state-of-the-art SOH estimation approaches rely on resource-consuming structures to obtain good performance. In this paper, an efficient deep learning model for SOH estimation, namely BMSFormer, is constructed. BMSFormer mainly integrates a Local-Global Fusion Attention module to capture both long-term and short-term dependencies while reducing computational complexity compared to traditional Softmax-based attention. Additionally, two kinds of depthwise separable convolution are embedded to fuse multi-scale and multi-channel features, enhancing feature diversity with fewer parameters than standard convolution. Three widely used battery datasets, each with different chemistries and operating conditions, are employed to evaluate the performance of BMSFormer. The experiments results illustrate that the proposed model achieves higher accuracy, lower computational consumption, and stabler performance across various hyperparameter combinations compared to alternative models.