Suggested Citation:
Zhao, Jingyuan, Zhilong Lv, Yunhong Che, Yuqi Li, Andrew Burke (2025)

End-To-End Millisecond-Level Battery Aging Diagnostics for Electric Vehicles

Battery health diagnostics are essential yet challenging due to varying operational conditions and diverse chemistries. This study proposes a hybrid deep transfer learning framework that combines two-dimensional convolutional neural network (CNN) with multi-head self-attention mechanisms. The model leverages standard battery metrics—voltage, charge, and their temporal variations across aging cycles—for feature extraction and predictive modeling in an end-to-end manner. Local convolution layers capture fine-grained dynamics, while global self-attention integrates long-range dependencies, enhancing generalization across chemistries and usage patterns. The framework was pre-trained on 77 lithium iron phosphate (LFP) cells and evaluated on 50 nickel manganese cobalt (NCM) and 17 nickel cobalt aluminum (NCA) cells. It achieved RMSEs of 2.96, 23, and 10.6 mAh, and R2 scores of 0.99, 0.928, and 0.927 for LFP, NCM, and NCA batteries, respectively. On a dedicated efficiency benchmark comprising 22 cells and 43,238 aging cycles, the proposed lightweight framework performs full life-cycle health diagnosis and aging-trajectory reconstruction for each cell with an average inference time of 31.3 ms, demonstrating that accurate capacity tracking can be achieved at millisecond-level latency. Real-world validation using a dataset from a fleet of 300 electric vehicles further demonstrated the robustness of the fusion model under noisy operational conditions, achieving an RMSE of 0.42 Ah, an R2 of 0.94, and a MAPE of 0.21 %. These results confirm that the proposed framework can effectively transfer aging knowledge across different chemistries and deployment scenarios, providing a scalable and accurate solution for battery capacity diagnostics through the strategic integration of local and global deep learning techniques.