Exploratory Analysis of Strain-Derived Features for Machine-Learning Estimation of State of Health in 18650 Lithium-Ion Cells

Accurate monitoring of lithium-ion battery state of health (SOH) is essential for ensuring safety and long-term performance. This study explores the use of strain sensing as a non-invasive diagnostic tool for SOH estimation in 18650 lithium-ion cells. Strain data were collected using external strain gauges during standard charge-discharge cycling, revealing mechanical deformation patterns that evolve with cell aging. To extract meaningful indicators from the strain response, we introduce a decomposition method that separates cyclic strain signals into three components: offset, slope, and residual. The offset captures irreversible deformation and structural degradation, the slope reflects reversible strain correlated with the state of charge, and the residual component captures nonlinear behaviors associated with lithium intercalation dynamics. Time-domain features were extracted from the residual strain signals to quantify changes in mechanical response. Features used include mean strain, maximum strain, root mean square strain, strain skewness, strain kurtosis, and maximum temperature. These features were found to correlate with capacity fade and serve as effective proxies for degradation. A machine learning framework was developed to predict battery SOH using a random forest model trained on the extracted strain-based features. This approach enables interpretable and scalable predictions by identifying the most informative features and mitigating overfitting. Results demonstrate the feasibility of strain-informed feature analysis coupled with data-driven modeling for real-time SOH monitoring in lithium-ion batteries. This study provides foundational evidence for integrating mechanical strain sensing into battery management systems to enhance reliability and safety in energy storage technologies.