Unlocking minute-level battery incremental capacity analysis construction using deep learning and multi-sequence alignment

Abstract

Incremental capacity analysis (ICA) is crucial for accurate, non-destructive lithium-ion battery degradation diagnosis, particularly for loss-sensitive electric vehicle (EV) applications. However, conventional ICA requires low-current charging over several hours, making it impractical under the EVs' multi-stage fast-charging conditions. Thus, this work unlocks a minute-level ICA construction framework for non-destructive mechanism diagnosis using stochastic charging segments. The multi-sequence alignment technique establishes the equivalent match between partial voltage segments and the ICA curve to eliminate conventional ICA data collection constraints. A residual-based convolutional neural network (R-CNN) is developed to achieve rapid and accurate ICA curve construction through feature fusion. Results demonstrate that 30 points collected within 5 min (starting from an arbitrary initial capacity) are sufficient for reliable ICA curve construction with the average mean absolute error (MAE) less than 0.061 Ah/V, and the average absolute percentage error (APE) less than 7.734 % for ICA peak estimation. The robustness of the proposed method under different working conditions has been verified. Through transfer learning, it is possible to adapt the pre-trained model to multiple fast-charging policies. Furthermore, the quantitative degradation mechanism from the rapidly constructed ICA curves facilitates practical electrode-level non-destructive battery diagnostics. This work can provide new perspectives for the characterization of battery degradation under fast-charging conditions.