A High-Dimensional Machine Learning Framework for Robust Vehicle Failure Prediction and Predictive Maintenance

The rapid digitalization of modern vehicles has led to the generation of high-dimensional sensor data, enabling advanced predictive maintenance and early failure detection strategies. Traditional diagnostic systems often fail to capture complex nonlinear patterns inherent in such data. This study proposes a robust machine learning benchmarking framework for vehicle failure prediction using a dataset containing 171 sensor-based attributes.
Seven supervised learning algorithms are evaluated under a consistent preprocessing and validation strategy: Logistic Regression, Support Vector Machine, k-Nearest Neighbors, Naïve Bayes, Decision Tree, Random Forest, and Neural Network. Performance is assessed using accuracy, precision, recall, F1-score, and Receiver Operating Characteristic–Area Under Curve (ROC–AUC) metrics to ensure robustness in imbalanced classification settings.
Experimental results reveal that ensemble-based methods, particularly Random Forest, significantly outperform linear, probabilistic, and distance-based models, achieving an AUROC of 0.983 on the test set. The findings provide empirical support for ensemble learning as a reliable and scalable solution for real-world vehicle failure diagnostics and predictive maintenance applications.