Artificial intelligence estimates the state of health of cylindrical lithium iron phosphate batteries via data-driven model

Accurate state of health (SOH) estimation for lithium iron phosphate (LFP) batteries remains challenging under long-cycle conditions. Conventional data-driven approaches typically demand extensive aging data and are often validated over only a few hundred cycles. This paper proposes an SOH estimation method for long-life LFP batteries, integrating electrochemical insight with statistical feature engineering to improve interpretability. Using cycling data from two types of cylindrical 26650 LFP batteries with different capacities and lifespans, eight statistical metrics are extracted from the first peak of the incremental capacity curve during early charging. These features are denoised via Gaussian-weighted moving average filtering and evaluated through Spearman correlation analysis, confirming strong correlation with capacity degradation. After comparing three feature selection strategies, the six most relevant metrics are used as health indicators to train an LSTM model. Validation across two independent datasets shows SOH estimation with mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE) as low as 0.258%, 0.298%, and 0.235%, respectively. Even with significantly reduced training samples, the maximum errors remain below 0.528%, 0.604%, and 0.472%. The results demonstrate the practicality of the approach for accurate, data-efficient SOH estimation of long-life LFP batteries.