Share
Export Citation
A Comparative Analysis of Isolation Forest, One-Class SVM, and Variational Autoencoder for Anomaly Detection in Diagnostic Trouble Codes (DTC) Data with the Development of an Anomaly Scoring Method
Hasrul
2026 IEEE International Conference on Artificial Intelligence and Mechatronics Systems Aims 2026
Abstract
Detecting subtle, non-linear vehicle sensor anomalies is essential yet challenging for predictive maintenance. We benchmark Isolation Forest (IF), One-Class Support Vector Machine (OC-SVM), and Variational Autoencoder (VAE) on an OBD-II dataset (396 observations) from a Toyota Innova, and propose a Performance-Based Weighting (PBW) ensemble that scales each base model by its validation discriminative power under a strict leakage-free protocol (time-based split, confounder exclusion, and validation-only thresholds and weights). On the held-out test set the PBW ensemble attains AUC-ROC 99.83% and F1 98.98%, matching or slightly exceeding the strongest individual model (OC-SVM) while assigning the weaker VAE the smallest weight (validation AUC ≈ 0.93, weight ≈ 0.28). More importantly, across a 10-fold blocked cross-validation it is the most stable model (AUC-ROC 99.05% ± 1.73%), and a stress test confirms resilience up to 25% perturbation-offering a precise, reliable method for early failure detection in automotive predictive maintenance.