Hybrid Ensemble Learning for Real-Time Predictive Maintenance in Vehicular Engine Health Monitoring Systems

Abstract

Vehicular engine failures are a major contributor to unplanned maintenance costs, operational downtime, and reduced road safety. Traditional diagnostic systems often rely on rule-based thresholds or simple machine learning models, which struggle to detect complex, evolving engine faults under real-world conditions. Accurate and interpretable predictive maintenance frameworks are essential to address the increasing demand for intelligent and reliable engine health monitoring. This study proposes a novel hybrid ensemble learning framework for real-time engine fault detection and classification, aimed at enhancing prediction accuracy and diagnostic robustness. The system integrates multiple base classifiers—Random Forest (RF), K-Nearest Neighbors (KNN), and Gradient Boosting (GBM)—within a stacked generalization architecture, using a Logistic Regression meta-learner. Deep learning enhancements, including an attention-augmented LSTM module, are applied to capture temporal patterns in engine sensor data. Real-time data from OBD-II and CAN bus systems are preprocessed using normalization, label encoding, and outlier handling techniques. Experimental evaluations on a multi-parameter engine dataset demonstrate that the proposed ensemble model achieves an accuracy of 90.2%, outperforming baseline models such as RF (86.5%) and SVM (85.4%). The system also reduces the false positive rate to 4.2%, and improves recall for rare but critical faults. This work contributes a scalable, interpretable, and high-performance diagnostic solution with strong applicability to Edge AI, fleet maintenance systems, and connected vehicle ecosystems.