Accurate monitoring of the State of Health (SOH) in lithium-ion batteries is essential for the safety and reliability of battery-operated systems. While deep learning-based SOH prediction models have shown high accuracy, their performance is heavily dependent on the quantity of experimental data. However, acquiring high-quality degradation data is time-consuming and costly, which severely limits dataset scale and consequently constrains model generalization and robustness. This presentation proposes a systematic methodology for autonomous experimental design using a newly developed 'Need Score' to identify the regions where additional data would most effectively improve model performance.

The proposed Need Score integrates three complementary indicators: prediction error (MAE), model uncertainty, and feature-space data sparsity, where neighborhood similarity and global distribution distances are quantified using K-Nearest Neighbors (KNN) and Mahalanobis distance. To ensure robustness against temperature-dependent behavior and cell-to-cell variability, a domain adaptation technique is introduced to normalize feature distributions across operating conditions. Experimental validation was conducted using the Cambridge University open dataset of LIR2032 coin cells. Results show that conventional random sampling and equal-interval data acquisition do not guarantee performance improvement, whereas adding only two strategically selected degradation intervals guided by the Need Score reduces prediction error by 19.7%. These results demonstrate the superior cost-effectiveness of targeted data acquisition and highlight the potential of autonomous experimental design for optimizing battery aging experiments under limited resources.