Researchers have developed an integrated framework for estimating battery state of health, or SOH, by combining incremental capacity analysis with image feature transformation and a hybrid machine-learning pipeline. The new approach is designed to improve both estimation accuracy and generalization across different operating conditions, a goal that is increasingly important as lithium-ion batteries are used in electric vehicles, stationary storage systems, and other applications where reliable health assessment directly affects safety, maintenance, and economic value.
SOH estimation is a central challenge in battery management because lithium-ion batteries do not age in a simple or linear way. As batteries cycle, their capacity, internal resistance, and electrochemical behavior evolve through complex and nonlinear processes. Battery operators need reliable health estimates to avoid safety problems, schedule maintenance, extend service life, and make better decisions about reuse or retirement. Yet estimating SOH accurately remains difficult, especially when models trained under one set of conditions must perform under another. That challenge becomes even more significant as batteries operate across different charging rates, temperatures, and usage profiles.
One established technique for assessing battery aging is incremental capacity, or IC, analysis. IC methods can reveal aging-sensitive electrochemical characteristics that are not obvious in raw voltage-capacity data alone. But extracting and using those features systematically is not always straightforward. In the new study, the researchers addressed this by transforming one-dimensional raw IC data into two-dimensional time-series images using the gramian angular field, or GAF, algorithm. This image-based representation allows the model to capture richer structural information than might be available from manually selected scalar features or simpler one-dimensional processing alone.
The framework then uses a CNN-LSTM model to extract intricate features from the transformed images before passing those features into an XGBoost estimator for final SOH prediction. In practical terms, this means the method is trying to combine the strengths of multiple machine-learning paradigms: convolutional neural networks for spatial feature extraction, long short-term memory networks for sequence-aware representation, and XGBoost for robust supervised estimation. The paper also compares three different CNN architectures, exploring how the design of the feature extractor influences performance. That comparative aspect is important because it moves beyond demonstrating one isolated model and instead helps clarify how architecture choice affects battery-health prediction quality.
According to the article, the addition of the XGBoost module substantially improved model performance. The results show a reduction in average RMSE of 68.2%, with the framework achieving an RMSE of 1.76% on the NASA dataset. On the CALCE dataset, the model achieved an RMSE below 1.5% when the training and test sets used consistent charge rates, and less than 2.7% when charge rates varied. Those numbers matter because they suggest the method is not only accurate under relatively matched conditions, but also retains useful robustness when the operating context shifts, which is often the harder and more practically relevant case.
The study's broader significance lies in how it handles feature representation and generalization together. Many SOH estimation methods struggle because either the extracted features are too shallow to represent complex aging patterns or the model becomes too specialized to one dataset or usage condition. By converting IC curves into structured images and then combining deep feature extraction with boosting-based estimation, the proposed framework tries to bridge that gap. The results suggest that this hybrid strategy can better capture the nonlinear degradation information embedded in battery behavior while remaining more adaptable across distinct datasets.
That has practical implications for battery monitoring in real systems. More accurate SOH estimation can improve battery management strategies, reduce uncertainty in operational planning, and support more reliable decisions about charging control, maintenance timing, and second-life use. In electric vehicles and grid-connected battery systems, even modest gains in estimation accuracy can translate into better safety margins and more efficient asset utilization. Methods that remain robust when charging conditions vary are particularly valuable, because real batteries rarely operate in the stable and repetitive conditions available in ideal laboratory experiments.
Further work will still be needed to assess how the framework performs under broader real-world conditions, including additional chemistries, temperature ranges, aging mechanisms, and noisy field data. Even so, the study offers a compelling example of how physically informative battery features can be combined with modern data-driven methods to improve health estimation. By turning IC curves into images and then learning from them through a layered machine-learning architecture, the framework points toward a more flexible and accurate path for battery SOH assessment in future energy systems.
Reference
Author:
Ping Ding a , Taotao Li a , Yajun Qiao b , Linfeng Zheng c , Hui Deng a , Weixiong Wu a
Title of original paper:
Integrated framework for battery SOH estimation using incremental capacity and image feature transformation
Article link:
https://www.sciencedirect.com/science/article/pii/S2773153725001161
Journal:
Green Energy and Intelligent Transportation
DOI:
10.1016/j.geits.2025.100366
Affiliations:
a Energy and Electricity Research Center, Jinan University, Zhuhai 519070, China
b Guangzhou Power Supply Bureau, Guangdong Power Grid Co., Ltd., CSG, Guangzhou 510620, China
c Sino-German College of Intelligent Manufacturing, Shenzhen Technology University, Shenzhen 518118, China
Experimental study
Not applicable
27-Feb-2026