Remaining-capacity estimation of lithium-ion batteries using heatmap-based denoising autoencoder and vision transformer algorithm

Lithium-ion batteries, which are valued for their high energy density and reusability, gradually lose capacity as charge–discharge cycles accumulate. Remaining capacity—the usable charge expressed in ampere-hours—is a direct, real-time indicator of state of health (SOH) and is critical for safe, eff...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:JOURNAL OF POWER ELECTRONICS Jg. 25; H. 9; S. 1745 - 1760
Hauptverfasser: Jang, Yunseo, Park, Joung-Hu
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Singapore Springer Nature Singapore 01.09.2025
Springer Nature B.V
전력전자학회
Schlagworte:
Aging
Back propagation
Batteries
Datasets
Deep learning
Electrical Machines and Networks
Engineering
Kalman filters
Lithium
Lithium-ion batteries
Machine learning
Maintenance costs
Noise reduction
Original Article
Parameter identification
Power Electronics
Power management
Real time
Recurrent neural networks
Support vector machines
전기공학
Lithium-ion battery
Vision transformer
Denoising autoencoder
Correlation heatmap
Battery dataset
Remaining capacity
ISSN:1598-2092, 2093-4718
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Lithium-ion batteries, which are valued for their high energy density and reusability, gradually lose capacity as charge–discharge cycles accumulate. Remaining capacity—the usable charge expressed in ampere-hours—is a direct, real-time indicator of state of health (SOH) and is critical for safe, efficient battery management. We propose a deep learning framework that transforms multichannel time-series measurements into correlation heatmaps, which are then fed into a denoising autoencoder (DAE) to suppress sensor noise while preserving electrochemical patterns. These image-based representations are subsequently processed by a vision transformer (ViT) to estimate the remaining capacity. The ViT architecture captures complex spatial–temporal dependencies through self-attention mechanisms. Experiments on the NASA lithium-ion benchmark dataset show that the proposed heatmap-based DAE-ViT model achieves lower root mean square error and mean absolute error compared with recurrent neural networks, one-dimensional transformer encoders trained on DAE features, support vector regression, and XGBoost. To further validate the proposed model, we apply it to the CALCE CS2 dataset, where it also demonstrates robust performance under different degradation scenarios. Results demonstrate that image-based self-attention can deliver scalable, high-precision remaining-capacity estimation for lithium-ion battery management systems.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
https://link.springer.com/article/10.1007/s43236-025-01134-x
ISSN:1598-2092
2093-4718
DOI:10.1007/s43236-025-01134-x