Research on AI-based Quality Prediction Methods for Electric Vehicle Components

Warranty claim prediction for electric vehicle (EV) traction motors is challenged by extremely rare failure events and limited long-term field history. This study investigates AI-based early-stage quality prediction under severe class imbalance using an industrial-scale dataset of approximately $\mathbf{3 5 0, 0 0 0}$ vehicles with about 1,500 failure cases. We benchmark classical machine-learning models (logistic regression, random forest, AdaBoost, and gradient boosting) and resampling-based pipelines (Borderline-SMOTE and ADASYN) using a multimetric evaluation framework centered on recall, F1-score, PR-AUC, and MCC. Hyperparameters are optimized under minimum operational requirements (e.g., recall $\geq 0.7$ and PR -AUC $\geq 0.1$) to reflect practical screening needs. Among the evaluated models, XGBoost achieves the best overall balance ($\mathrm{PR}-\mathrm{AUC} \approx 0.105$, recall $\approx 0.742$) and enables quantitative estimation of field risk, predicting an approximate claim probability of $1.02 \%$ within the warranty period and $3.94 \%$ over the full durability horizon. The results provide a practical baseline for deploying data-driven quality prediction in low-data, highly imbalanced EV motor applications and motivate future extensions to multimodal and time-dependent modeling. In this study, we benchmark these classical models against state-of-the-art architectures.