Cross-Validated Machine Learning Models for Smartphone Battery State-of-Health Classification

Abstract

This study investigates the application of Machine Learning (ML) classification models for assessing smartphone battery State-of-Health (SoH) using temperature, usage, environment, and device-related parameters. A dataset comprising 675 observations and 15 attributes was obtained from ten smartphones, with battery SoH categorized as Poor, Fair, or Good. Eight ML algorithms: Naïve Bayes (NB), Decision Tree (DT), Logistic Regression (LR), Gradient Boosting (GB), Support Vector Machine (SVM), Random Forest (RF), CN2 Rule Induction (CN2-RI), and k-Nearest Neighbour (kNN), were evaluated using 10-fold cross-validation. The results indicate substantial differences in model performance across battery-SoH and smartphones, with notable difficulty in correctly identifying Good and Fair SoH and a tendency among several models to classify these conditions as Poor. LR achieved a high number of correct Poor-SoH predictions, while DT and NB showed comparatively more balanced performance across the SoH. The findings demonstrate the potential of ML for smartphone battery-SoH classification but also highlight the effects of SoH imbalance and the need for SoH-sensitive evaluation when selecting reliable predictive models.