Machine Learning-Based Fault Classification of Lithium-Ion Batteries Using Experimentally Measured Charging Features

Ensuring safe energy storage in lithium-ion batteries requires reliable fault diagnosis. However, classifying faults from small experimental datasets demands rigorous control over partition variability, hyperparameter bias, and measurement noise. This study evaluates five machine-learning classifiers using 110 experimentally measured charging observations (63 faulty, 47 non-faulty), each described by nine electrical and thermal features. The evaluation framework incorporated 30 repeated, stratified 80/10/10 partitions with an inner 5-fold cross-validation for hyperparameter tuning. LogitBoost showed the highest observed repeat-level mean performance, with an F1-score of 0.9930 ± 0.0214 (95% CI: 0.9841–1.0000) and accuracy of 0.9909 ± 0.0277 (95% CI: 0.9818–1.0000), closely followed by neural network and random forest models. The Friedman test revealed no statistically significant overall differences in performance among the models (chi-square = 7.6467, p = 0.1054). Under a computational zero-mean Gaussian perturbation test with amplitudes up to 10% of the development-set feature standard deviation, the mean F1-score drop remained below 0.0034, and prediction consistency exceeded 0.9881. Consensus feature rankings identified intermediate charging currents as most decisive, showing moderate stability across partitions (mean rank correlation = 0.6361). Although physical cell identifiers were not retained, these results demonstrate robust sample-level, within-dataset fault discrimination, providing a methodology for reliable evaluation of limited experimental battery data.