Early failure detection in turbofan engines using hybrid ensemble learning

Abstract Early failure detection in turbofan engines is essential for improving aviation safety and reducing maintenance costs. However, reliable fault prediction remains challenging due to sensor noise and severe class imbalance in operational data. This study proposes a hybrid ensemble learning framework that combines noise-robust feature engineering with explainable machine learning for early fault detection. Using the NASA Commercial Modular Aero-Propulsion System Simulation dataset, a multi-scale sliding-window approach is employed to reduce sensor noise, while Random Forest, Extreme Gradient Boosting, and Light Gradient Boosting Machine models are integrated through probability-weighted soft voting. The decision threshold is optimized to improve the detection of imminent failures in safety-critical environments. The proposed framework achieves an Area Under the Curve of 0.991 and an F1-score of 0.708, outperforming existing approaches. Shapley Additive Explanations analysis further shows that the model relies on physically meaningful degradation indicators, including static pressure decreases and temperature increases. The results demonstrate a robust, accurate, and interpretable predictive maintenance solution for aviation applications.