Meta-Learning Driven Voltage Reconstruction and Spatiotemporal Fusion for Battery SOH Estimation Under Incomplete Data

Accurate estimation of state of health (SOH) is a critical challenge in battery management systems. However, in practical operation, charging data is often incomplete due to abnormal operating conditions or limited data access, which reduces the reliability of SOH estimation. To address this issue, an integrated method is proposed for battery SOH estimation by combining data reconstruction with image feature learning. Battery aging data is obtained through aging experiments. A model-agnostic meta-learning-based least squares generative adversarial network (MAML-LSGAN) is developed to reconstruct charging voltage data under different levels of data missingness. The proposed MAML-LSGAN improves the adaptability of voltage reconstruction under different battery aging conditions through meta-learning optimization. Subsequently, the Gramian angular field method is employed to transform the reconstructed 1-D voltage time series into 2-D image representations, enhancing the extraction of aging-related features. On this basis, a Swin transformer-LSTM model integrated with a cross-attention mechanism is constructed to capture spatial–temporal degradation features for accurate battery SOH estimation. The proposed method is validated on two different battery datasets, and experimental results demonstrate that it achieves high SOH estimation accuracy and stability under different training data ratios.