Machine Learning in Aero-Engine Fault and Structural Damage Predictions

Within the last decade, engine-related faults have accounted for at least 20% of fatal aviation accidents. Additionally, loss of control in-flight—often tied to structural failures—remains the deadliest mishap in modern aviation. Consequently, predictive maintenance has gained prominence in preventing unplanned failures, cutting direct maintenance costs, and enhancing dispatch reliability. Data-driven approaches, particularly machine learning (ML), are increasingly adopted in predictive maintenance strategies. Leveraging vast operational and sensor data, these methods offer more flexible and efficient fault prediction than traditional model-based techniques, which require extensive modelling and often struggle with evolving operational conditions. In aero-engine fault prediction, prevailing studies rely on synthetic or experimental datasets—where faults are artificially induced and validated using non-aero-engine data—lacking true operational fidelity. This research addresses that gap by leveraging five years of real-world time-series data from ten aircraft, cover over 4,000 engine hours. Advanced pre-processing and feature engineering techniques were employed to handle data-heterogeneity and high-dimensionality, while a novel resampling approach that mitigates computational cost associated with class imbalance is proposed. Model performance is validated using a rigorous, yet less conservative approach offering greater confidence in the model’s prediction. For structural damage prediction, key acoustic emission features were identified to develop high-performance models for identifying and predicting damage sequence in composite structures exposed to seawater degradation and structural damage conditions. Non-destructive ML models were designed to characterise the effects of fabrication temperature-induced porosity on mechanical properties of composite structures. Thus, providing invaluable insights for the autonomous control and data-driven optimisation of aerospace structures. This thesis advances ML application in aerospace engineering, addressing key challenges in aeroengine fault and structural damage prediction. The limitations lay a foundation for further work in transfer learning, reducing false alarms, and adapting models to complex faults and structural damage patterns. Thus, enhancing scalability and real-world applicability.