Battery health monitoring and remaining useful life (RUL) estimation for electric vehicles face two significant challenges: battery data heterogeneity and sample imbalance. This study presents a novel approach based on Transformer architecture to specifically address these issues. We utilized the NASA lithium-ion battery cycling dataset, which contains charge-discharge and impedance measurement data under various temperature conditions. To tackle data heterogeneity, we developed a multimodal feature fusion strategy that effectively integrates battery sensor data from different sources and formats, including time-series charge-discharge sensor data and spectral impedance sensor measurements. To mitigate sample imbalance, we implemented an adaptive resampling technique and hierarchical attention mechanism, enhancing the model’s ability to recognize rare degradation patterns. Our Transformer-based model captures long-term dependencies in the battery degradation process through its self-attention mechanism. Experimental results demonstrate that the proposed solution significantly improves battery degradation prediction accuracy, achieving a 21.3% increase in accuracy when processing heterogeneous data and a 24.5% improvement in prediction capability for imbalanced samples compared to traditional methods. Additionally, through case studies, we validate the applicability of this method in actual electric vehicle battery management systems, providing reliable data support for battery preventive maintenance and replacement decisions. The findings have important implications for enhancing the reliability and economic efficiency of electric vehicle battery management systems.